{
 "S1sh3::signage": {
  "fp": "c80056daa4c84bb5",
  "inscriptions": []
 },
 "era_assess::66d6f6580058177c": {
  "subjects": [
   {
    "subject_native": "2000년대 초반 한국 유원지의 타가디스코 (디스코팡팡)",
    "search_terms_native": [
     "월미도 디스코팡팡",
     "타가다디스코",
     "야외 놀이공원 타가디스코",
     "2000년대 디스코팡팡"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 작성되어야 하며 영어나 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 특유의 유원지 타가디스코(디스코팡팡)는 독특한 원형 좌석 배치, 안전벨트 없는 철제 난간 구조를 지니고 있어 일반적인 놀이기구 이미지와는 큰 차이가 있습니다."
   }
  ]
 },
 "era_ref::cd88849d5b634361": {
  "subject": "2000년대 초반 한국 유원지의 타가디스코 (디스코팡팡)",
  "terms": [
   "월미도 디스코팡팡",
   "타가다디스코",
   "야외 놀이공원 타가디스코",
   "2000년대 디스코팡팡"
  ],
  "queries": [
   [
    "2000년대 초반 월미도 디스코팡팡 타가다디스코 야외 놀이공원",
    "옛날 월미도 타가디스코 디스코팡팡 2000년대"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://chillpainai.com/src/wewakeup/scoop/images/e2d7f6f5d57b04247a86224a5634db77cc61173d.jpg",
  "picked_reason_ko": "2번은 한국 유원지의 전형적인 디스코팡팡을 화면 중심에 크게 담아 원형 승강부, 둘레 좌석, 안전 난간, 조종실과 외장 재료를 가장 명확하게 읽을 수 있다.",
  "sha256": "78b2bd5ef938ebef10561905f34533d3d828624bb198d507030c926d2760d2f2",
  "file": "eraref_cd88849d5b634361.png"
 },
 "groupbg::놀이기구": {
  "input_fingerprint": "afd37fc49232de8d",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "놀이기구",
    "tags": [
     "S1sh3"
    ]
   },
   "context_sig": "853f85b81d174f34",
   "era_research_sha": "78b2bd5ef938ebef10561905f34533d3d828624bb198d507030c926d2760d2f2"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: On the passenger deck of a spinning amusement-park ride in the open air.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n놀이공원 타가디스코: 2001년 당시의 유원지 분위기가 나는 야외 놀이공원으로, 크고 둥근 원반형 놀이기구 주변이다. (특징: 빠르게 회전하며 상하로 요동치는 둥근 원반 형태의 대형 놀이기구; 원반 가장자리의 둥근 좌석들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 놀이공원 - 오후\n- 타가디스코를 타고 있는 선영(여, 17세)과 수정을 비롯한 친구들의 모습.\n\nTIME OF DAY (lock): afternoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 초반 한국 유원지의 타가디스코 (디스코팡팡): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: On the passenger deck of a spinning amusement-park ride in the open air.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n놀이공원 타가디스코: 2001년 당시의 유원지 분위기가 나는 야외 놀이공원으로, 크고 둥근 원반형 놀이기구 주변이다. (특징: 빠르게 회전하며 상하로 요동치는 둥근 원반 형태의 대형 놀이기구; 원반 가장자리의 둥근 좌석들)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 놀이공원 - 오후\n- 타가디스코를 타고 있는 선영(여, 17세)과 수정을 비롯한 친구들의 모습.\n\nTIME OF DAY (lock): afternoon.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 초반 한국 유원지의 타가디스코 (디스코팡팡): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_놀이기구_f4a1f9.png",
  "asset_id": "9433a719-eefe-4894-8ffe-e6503bcedbc8",
  "input_asset_ids": [
   "5a818fca-9c16-4015-bbc4-bded64fe2892"
  ],
  "origin_tag": "S1sh3",
  "place_text": "On the passenger deck of a spinning amusement-park ride in the open air.",
  "origin_inputs": {
   "place_text": "On the passenger deck of a spinning amusement-park ride in the open air.",
   "time_of_day_en": "afternoon",
   "conti_asset_id": "5a818fca-9c16-4015-bbc4-bded64fe2892"
  },
  "era_research": {
   "subject": "2000년대 초반 한국 유원지의 타가디스코 (디스코팡팡)",
   "terms": [
    "월미도 디스코팡팡",
    "타가다디스코",
    "야외 놀이공원 타가디스코",
    "2000년대 디스코팡팡"
   ],
   "queries": [
    [
     "2000년대 초반 월미도 디스코팡팡 타가다디스코 야외 놀이공원",
     "옛날 월미도 타가디스코 디스코팡팡 2000년대"
    ]
   ],
   "candidates": 4,
   "picked_index": 2,
   "picked_url": "https://chillpainai.com/src/wewakeup/scoop/images/e2d7f6f5d57b04247a86224a5634db77cc61173d.jpg",
   "picked_reason_ko": "2번은 한국 유원지의 전형적인 디스코팡팡을 화면 중심에 크게 담아 원형 승강부, 둘레 좌석, 안전 난간, 조종실과 외장 재료를 가장 명확하게 읽을 수 있다.",
   "sha256": "78b2bd5ef938ebef10561905f34533d3d828624bb198d507030c926d2760d2f2",
   "file": "eraref_cd88849d5b634361.png",
   "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_cd88849d5b634361.png"
  }
 },
 "era_assess::3e4cbd1652b35abe": {
  "subjects": [
   {
    "subject_native": "2000년대 대한민국 놀이공원의 디스코팡팡",
    "search_terms_native": [
     "디스코팡팡",
     "월미도 디스코팡팡",
     "타가다 디스코"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "한국의 디스코팡팡(타가다)은 특유의 원형 좌석 배치, 안전바가 없는 철제 손잡이 구조, 주변의 DJ 부스 등 고유의 시각적 특징이 있어 일반적인 회전 놀이기구와 크게 다릅니다."
   }
  ]
 },
 "era_fail::00ce9627a304e619": {
  "stage": "research",
  "subject": "2000년대 대한민국 놀이공원의 디스코팡팡",
  "attempts": 5
 },
 "S1sh3::bgfirst_bg": {
  "input_fingerprint": "71e1d71d0790f2b4",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 즐겁게 웃고 있는 김선영의 해맑은 얼굴 클로즈업, 화면 하단에 '2001년'이라는 자막이 오버랩된 상태\n\nLOCATION (lock): On the passenger deck of a spinning amusement-park ride in the open air.\n\nTIME OF DAY (lock): afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At face height beside the moving ride, the camera holds a tight three-quarter close-up as 김선영 tips her head back in unguarded laughter, her face occupying most of the upper frame. The parallel tracking move settles while the lower portion remains clear for the screen-facing ‘2001년’ title.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: amusement ride structure (shaking in motion) — Only an oblique portion beside and behind 김선영 is visible to camera; used as Soft peripheral context behind the close-up, retaining a sense of the moving attraction without competing with the face; ‘2001년’ title overlay (displayed) — The screen-facing graphic reads ‘2001년’ across the lower frame; used as Lower-frame temporal marker kept legible beneath the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light with a restrained neutral-to-cool grade preserves the freshness of her expression without heightened gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 즐겁게 웃고 있는 김선영의 해맑은 얼굴 클로즈업, 화면 하단에 '2001년'이라는 자막이 오버랩된 상태\n\nLOCATION (lock): On the passenger deck of a spinning amusement-park ride in the open air.\n\nTIME OF DAY (lock): afternoon.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At face height beside the moving ride, the camera holds a tight three-quarter close-up as 김선영 tips her head back in unguarded laughter, her face occupying most of the upper frame. The parallel tracking move settles while the lower portion remains clear for the screen-facing ‘2001년’ title.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: amusement ride structure (shaking in motion) — Only an oblique portion beside and behind 김선영 is visible to camera; used as Soft peripheral context behind the close-up, retaining a sense of the moving attraction without competing with the face; ‘2001년’ title overlay (displayed) — The screen-facing graphic reads ‘2001년’ across the lower frame; used as Lower-frame temporal marker kept legible beneath the face.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light with a restrained neutral-to-cool grade preserves the freshness of her expression without heightened gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S1sh3__bgfirst_bg.png",
  "asset_id": "b0dfb259-5517-4d56-9a94-adda76a72d5a",
  "input_asset_ids": [
   "5a818fca-9c16-4015-bbc4-bded64fe2892",
   "9433a719-eefe-4894-8ffe-e6503bcedbc8"
  ]
 },
 "S1sh3": {
  "input_fingerprint": "e6ad7c49d0c20615",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 즐겁게 웃고 있는 김선영의 해맑은 얼굴 클로즈업, 화면 하단에 '2001년'이라는 자막이 오버랩된 상태\n\nLOCATION (lock): On the passenger deck of a spinning amusement-park ride in the open air. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At face height beside the moving ride, the camera holds a tight three-quarter close-up as 김선영 tips her head back in unguarded laughter, her face occupying most of the upper frame. The parallel tracking move settles while the lower portion remains clear for the screen-facing ‘2001년’ title.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: amusement ride structure (shaking in motion) — Only an oblique portion beside and behind 김선영 is visible to camera; used as Soft peripheral context behind the close-up, retaining a sense of the moving attraction without competing with the face; ‘2001년’ title overlay (displayed) — The screen-facing graphic reads ‘2001년’ across the lower frame; used as Lower-frame temporal marker kept legible beneath the face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light with a restrained neutral-to-cool grade preserves the freshness of her expression without heightened gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 즐겁게 웃고 있는 김선영의 해맑은 얼굴 클로즈업, 화면 하단에 '2001년'이라는 자막이 오버랩된 상태\n\nLOCATION (lock): On the passenger deck of a spinning amusement-park ride in the open air. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At face height beside the moving ride, the camera holds a tight three-quarter close-up as 김선영 tips her head back in unguarded laughter, her face occupying most of the upper frame. The parallel tracking move settles while the lower portion remains clear for the screen-facing ‘2001년’ title.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: amusement ride structure (shaking in motion) — Only an oblique portion beside and behind 김선영 is visible to camera; used as Soft peripheral context behind the close-up, retaining a sense of the moving attraction without competing with the face; ‘2001년’ title overlay (displayed) — The screen-facing graphic reads ‘2001년’ across the lower frame; used as Lower-frame temporal marker kept legible beneath the face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light with a restrained neutral-to-cool grade preserves the freshness of her expression without heightened gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): afternoon.\n\nSHOT TEXT (authoritative, Korean): 즐겁게 웃고 있는 김선영의 해맑은 얼굴 클로즈업, 화면 하단에 '2001년'이라는 자막이 오버랩된 상태\n\nLOCATION (lock): On the passenger deck of a spinning amusement-park ride in the open air. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At face height beside the moving ride, the camera holds a tight three-quarter close-up as 김선영 tips her head back in unguarded laughter, her face occupying most of the upper frame. The parallel tracking move settles while the lower portion remains clear for the screen-facing ‘2001년’ title.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: amusement ride structure (shaking in motion) — Only an oblique portion beside and behind 김선영 is visible to camera; used as Soft peripheral context behind the close-up, retaining a sense of the moving attraction without competing with the face; ‘2001년’ title overlay (displayed) — The screen-facing graphic reads ‘2001년’ across the lower frame; used as Lower-frame temporal marker kept legible beneath the face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural afternoon light with a restrained neutral-to-cool grade preserves the freshness of her expression without heightened gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S1sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S1sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870689>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_놀이기구_f4a1f9.png"
    },
    {
     "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870689>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 타이트한 클로즈업 프레이밍을 정확히 구현하였고, 배경의 모션 블러로 놀이기구의 움직임을 훌륭하게 표현하여 가장 높은 점수를 받음."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "프레이밍이 지시사항보다 넓어 얼굴이 화면을 충분히 채우지 못하며, 배경이 선명해 놀이기구의 움직임이 느껴지지 않음."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "고개를 뒤로 젖히고 위를 향해 시선을 두며 환하게 웃고 있음.",
      "built_space": "빨간색 둥근 놀이기구 좌석에 앉아 있으며, 뒤편으로 색색의 원형 안전바가 배치됨.",
      "entities": "김선영(보라색 패딩, 검은색 긴 머리, 비니 누락)이 중심에 있고 화면 하단에 '2001년' 텍스트가 오버랩됨.",
      "hard_violations": [],
      "physics": "좌석에 기대어 앉아 있으며, 배경의 가로 방향 모션 블러가 놀이기구의 회전 상태를 뒷받침함."
     },
     {
      "label": "A",
      "direction": "고개를 들고 위를 바라보며 크게 웃고 있음.",
      "built_space": "원형 놀이기구 내부에 앉아 있고, 뒤로 안전바와 조종석(매표소 형태) 구조물이 선명하게 보임.",
      "entities": "김선영(보라색 패딩, 검은색 비니, 검은색 긴 머리)이 레퍼런스 의상을 모두 갖췄으며, 하단에 '2001년' 자막이 있음.",
      "hard_violations": [],
      "physics": "좌석에 안정적으로 앉아 있으나, 배경이 완전히 정지된 상태로 묘사되어 움직임이 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 타이트한 클로즈업 프레이밍을 정확히 구현하였고, 배경의 모션 블러로 놀이기구의 움직임을 훌륭하게 표현하여 가장 높은 점수를 받음."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "프레이밍이 지시사항보다 넓어 얼굴이 화면을 충분히 채우지 못하며, 배경이 선명해 놀이기구의 움직임이 느껴지지 않음."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "고개를 뒤로 젖히고 위를 향해 시선을 두며 환하게 웃고 있음.",
      "built_space": "빨간색 둥근 놀이기구 좌석에 앉아 있으며, 뒤편으로 색색의 원형 안전바가 배치됨.",
      "entities": "김선영(보라색 패딩, 검은색 긴 머리, 비니 누락)이 중심에 있고 화면 하단에 '2001년' 텍스트가 오버랩됨.",
      "hard_violations": [],
      "physics": "좌석에 기대어 앉아 있으며, 배경의 가로 방향 모션 블러가 놀이기구의 회전 상태를 뒷받침함."
     },
     {
      "label": "A",
      "direction": "고개를 들고 위를 바라보며 크게 웃고 있음.",
      "built_space": "원형 놀이기구 내부에 앉아 있고, 뒤로 안전바와 조종석(매표소 형태) 구조물이 선명하게 보임.",
      "entities": "김선영(보라색 패딩, 검은색 비니, 검은색 긴 머리)이 레퍼런스 의상을 모두 갖췄으며, 하단에 '2001년' 자막이 있음.",
      "hard_violations": [],
      "physics": "좌석에 안정적으로 앉아 있으나, 배경이 완전히 정지된 상태로 묘사되어 움직임이 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 타이트한 클로즈업 프레이밍과 놀이기구의 움직임(모션 블러)을 매우 잘 구현했으나, 레퍼런스의 비니가 누락되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "캐릭터 레퍼런스의 의상(비니 포함)은 정확히 일치하지만, 타이트한 클로즈업 대신 미디엄 샷으로 렌더링되어 프레이밍 지시를 어겼으며 놀이기구의 움직임이 느껴지지 않습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 위쪽 화면 밖을 향하며 크게 웃고 있음.",
      "built_space": "놀이기구의 붉은색 곡선형 등받이 좌석과 금속 난간이 인물 뒤로 배치됨.",
      "entities": "보라색 패딩을 입은 긴 흑발의 소녀(비니 없음), 화면 하단 중앙에 '2001년' 텍스트 존재.",
      "hard_violations": [],
      "physics": "좌석에 기대어 앉아 있으며, 배경의 모션 블러를 통해 회전하는 물리적 움직임을 표현함."
     },
     {
      "label": "B",
      "direction": "시선은 위쪽 화면 밖을 향하며 웃고 있음.",
      "built_space": "놀이기구의 붉은색 좌석과 난간, 뒷배경의 조종 부스까지 넓은 시야로 보임.",
      "entities": "검은색 비니와 보라색 패딩을 착용한 소녀, 화면 하단 중앙에 '2001년' 텍스트 존재.",
      "hard_violations": [],
      "physics": "좌석에 앉아 있으나 배경이 뚜렷하여 물리적인 회전 운동이나 흔들림이 나타나지 않음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 타이트한 클로즈업 프레이밍과 놀이기구의 움직임(모션 블러)을 매우 잘 구현했으나, 레퍼런스의 비니가 누락되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "캐릭터 레퍼런스의 의상(비니 포함)은 정확히 일치하지만, 타이트한 클로즈업 대신 미디엄 샷으로 렌더링되어 프레이밍 지시를 어겼으며 놀이기구의 움직임이 느껴지지 않습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 위쪽 화면 밖을 향하며 크게 웃고 있음.",
      "built_space": "놀이기구의 붉은색 곡선형 등받이 좌석과 금속 난간이 인물 뒤로 배치됨.",
      "entities": "보라색 패딩을 입은 긴 흑발의 소녀(비니 없음), 화면 하단 중앙에 '2001년' 텍스트 존재.",
      "hard_violations": [],
      "physics": "좌석에 기대어 앉아 있으며, 배경의 모션 블러를 통해 회전하는 물리적 움직임을 표현함."
     },
     {
      "label": "A",
      "direction": "시선은 위쪽 화면 밖을 향하며 웃고 있음.",
      "built_space": "놀이기구의 붉은색 좌석과 난간, 뒷배경의 조종 부스까지 넓은 시야로 보임.",
      "entities": "검은색 비니와 보라색 패딩을 착용한 소녀, 화면 하단 중앙에 '2001년' 텍스트 존재.",
      "hard_violations": [],
      "physics": "좌석에 앉아 있으나 배경이 뚜렷하여 물리적인 회전 운동이나 흔들림이 나타나지 않음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "B",
    "direction": "고개를 뒤로 젖히고 위를 향해 시선을 두며 환하게 웃고 있음.",
    "built_space": "빨간색 둥근 놀이기구 좌석에 앉아 있으며, 뒤편으로 색색의 원형 안전바가 배치됨.",
    "entities": "김선영(보라색 패딩, 검은색 긴 머리, 비니 누락)이 중심에 있고 화면 하단에 '2001년' 텍스트가 오버랩됨.",
    "hard_violations": [],
    "physics": "좌석에 기대어 앉아 있으며, 배경의 가로 방향 모션 블러가 놀이기구의 회전 상태를 뒷받침함."
   },
   {
    "label": "A",
    "direction": "고개를 들고 위를 바라보며 크게 웃고 있음.",
    "built_space": "원형 놀이기구 내부에 앉아 있고, 뒤로 안전바와 조종석(매표소 형태) 구조물이 선명하게 보임.",
    "entities": "김선영(보라색 패딩, 검은색 비니, 검은색 긴 머리)이 레퍼런스 의상을 모두 갖췄으며, 하단에 '2001년' 자막이 있음.",
    "hard_violations": [],
    "physics": "좌석에 안정적으로 앉아 있으나, 배경이 완전히 정지된 상태로 묘사되어 움직임이 없음."
   }
  ],
  "totals": {
   "A": 9,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지시된 타이트한 클로즈업 프레이밍을 정확히 구현하였고, 배경의 모션 블러로 놀이기구의 움직임을 훌륭하게 표현하여 가장 높은 점수를 받음."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "프레이밍이 지시사항보다 넓어 얼굴이 화면을 충분히 채우지 못하며, 배경이 선명해 놀이기구의 움직임이 느껴지지 않음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_놀이기구_f4a1f9.png"
   },
   {
    "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870689>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "캐릭터 레퍼런스 이미지에서 착용하고 있는 검은색 비니가 생략되었습니다.",
     "fix_en": "Add the reference's dark grey beanie to her head. Preserve the face, expression, jacket, background, and text overlay.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "화면을 채운 김선영의 얼굴형·이목구비가 레퍼런스 인물과 달라 동일 인물로 보이지 않음",
     "fix_en": "Adjust facial features to match the character reference. Preserve the laughing expression, hair, jacket, background, and text overlay.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "캐릭터 레퍼런스 이미지에서 착용하고 있는 검은색 비니가 생략되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면을 채운 김선영의 얼굴형·이목구비가 레퍼런스 인물과 달라 동일 인물로 보이지 않음",
     "severity": "major"
    },
    {
     "issue_ko": "머리 전체가 보이는데 레퍼런스 의상인 회색 비니가 없음",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "캐릭터 레퍼런스 이미지에서 착용하고 있는 검은색 비니가 생략되었습니다.",
    "fix_en": "Add the reference's dark grey beanie to her head. Preserve the face, expression, jacket, background, and text overlay.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "화면을 채운 김선영의 얼굴형·이목구비가 레퍼런스 인물과 달라 동일 인물로 보이지 않음",
    "fix_en": "Adjust facial features to match the character reference. Preserve the laughing expression, hair, jacket, background, and text overlay.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S1sh3__bgfirst_bg.png",
   "bg_asset_id": "b0dfb259-5517-4d56-9a94-adda76a72d5a",
   "bg_record_key": "S1sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "놀이기구",
   "groupbg_asset_id": "9433a719-eefe-4894-8ffe-e6503bcedbc8"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S1sh3::cine": {
  "applied": true,
  "fingerprint": "fff36ff19ba380174487523ff3c669e24684888c2404a798f2f3ab4a588ad8e0",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S1sh3_sel.png",
  "source_sha256": "f5f748b22efd6be7106be933fee07c5243d92db04c74a722b2c21bbe93e2dbbc",
  "file": "S1sh3_cine.png",
  "latency_ms": 11035
 },
 "S2sh8::signage": {
  "fp": "b6f0557ad0c243a3",
  "inscriptions": [
   {
    "surface_native": "식당 벽면의 안내판",
    "text_native": "물은 셀프",
    "reason_ko": "2000년대 한국 식당 내부의 전형적인 분위기와 사실감을 더하기 위해 주방 근처 벽면에 흔히 붙어 있는 안내 문구를 묘사함."
   }
  ]
 },
 "era_assess::09b4533c903bd401": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 식당 내부와 테이블 집기",
    "search_terms_native": [
     "2000년대 식당 내부",
     "식당 수저테이블",
     "식당 컵 살균기",
     "한국 식당 주방 입구"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 식당 특유의 수저 서랍형 테이블, 자외선 컵 살균기, 스테인리스 식기류 및 주방 입구 가림막 구조는 일반적인 서구식 식당 일러스트와 완전히 다릅니다."
   }
  ]
 },
 "era_ref::0c71c8c760b86593": {
  "subject": "2000년대 한국 식당 내부와 테이블 집기",
  "terms": [
   "2000년대 식당 내부",
   "식당 수저테이블",
   "식당 컵 살균기",
   "한국 식당 주방 입구"
  ],
  "queries": [
   [
    "2000년대 한국 식당 내부 수저 테이블 집기",
    "2000년대 한국 식당 컵 살균기 주방 입구"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDExMjdfMTY3/MDAxNzMyNjg5NDU4NDgx.ly2STHQdRL1KABZSNCH3qNE-xCbJwqj2fVXn8H1s8-og.VblRlgjrgEF5j2uJzrFsrFqnTdUA4qgXnUW10a6Tz48g.JPEG/SE-EAB0A962-15B5-44CC-932D-A70D743E0F70.jpg?type=w800",
  "picked_reason_ko": "식당 홀의 전체 구조와 테이블 배치, 의자, 수저·컵·휴지통 등 일상적인 집기가 함께 선명하게 보여 2000년대 한국 식당의 형태 참고에 가장 적합하다.",
  "sha256": "a937b1ed516f6ca4151f6881e6144e69ae13e1de2225d79d869caf3fcfd9ba45",
  "file": "eraref_0c71c8c760b86593.png"
 },
 "S2sh8::bgfirst_bg": {
  "input_fingerprint": "1eba806617d373eb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 지폐를 손에 쥔 채 혀를 살짝 내밀고 심옥을 향해 미소를 띤 김선영의 표정\n\nLOCATION (lock): Inside the restaurant dining area, beside the tables just outside the kitchen.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just below 김선영’s eye line and over 심옥’s near shoulder, frame a tight three-quarter close-up with 김선영’s playful face in the upper center and the banknotes retained along the lower edge. 김선영 leans toward 심옥, briefly extends her tongue, and keeps her eyes fixed on her mother while the static camera lets the teasing gesture land.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 1,000-won banknotes (held in 김선영's hand) — The visible faces of the notes are angled upward enough to read as 1,000-won bills; used as Small but readable action detail at the lower edge of the close-up; restaurant tables (remaining in the dining area) — Their edges recede obliquely behind 김선영; used as Soft contextual depth behind the exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient light appropriate to the nighttime restaurant is graded in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 식당 내부와 테이블 집기: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 지폐를 손에 쥔 채 혀를 살짝 내밀고 심옥을 향해 미소를 띤 김선영의 표정\n\nLOCATION (lock): Inside the restaurant dining area, beside the tables just outside the kitchen.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just below 김선영’s eye line and over 심옥’s near shoulder, frame a tight three-quarter close-up with 김선영’s playful face in the upper center and the banknotes retained along the lower edge. 김선영 leans toward 심옥, briefly extends her tongue, and keeps her eyes fixed on her mother while the static camera lets the teasing gesture land.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 1,000-won banknotes (held in 김선영's hand) — The visible faces of the notes are angled upward enough to read as 1,000-won bills; used as Small but readable action detail at the lower edge of the close-up; restaurant tables (remaining in the dining area) — Their edges recede obliquely behind 김선영; used as Soft contextual depth behind the exchange.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient light appropriate to the nighttime restaurant is graded in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 식당 내부와 테이블 집기: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S2sh8__bgfirst_bg.png",
  "asset_id": "4d9ed8fe-445d-449f-a769-37fa97f3b65d",
  "input_asset_ids": [
   "4079326f-2973-4ea8-94ff-1c2b281365d7",
   "7e0157b4-fb4e-4756-a7dd-3f15d2dde7f1"
  ],
  "era_research": {
   "subject": "2000년대 한국 식당 내부와 테이블 집기",
   "queries": [
    [
     "2000년대 한국 식당 내부 수저 테이블 집기",
     "2000년대 한국 식당 컵 살균기 주방 입구"
    ]
   ],
   "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDExMjdfMTY3/MDAxNzMyNjg5NDU4NDgx.ly2STHQdRL1KABZSNCH3qNE-xCbJwqj2fVXn8H1s8-og.VblRlgjrgEF5j2uJzrFsrFqnTdUA4qgXnUW10a6Tz48g.JPEG/SE-EAB0A962-15B5-44CC-932D-A70D743E0F70.jpg?type=w800",
   "sha256": "a937b1ed516f6ca4151f6881e6144e69ae13e1de2225d79d869caf3fcfd9ba45",
   "file": "eraref_0c71c8c760b86593.png"
  }
 },
 "S2sh8": {
  "input_fingerprint": "3aa175802b6ca16a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 지폐를 손에 쥔 채 혀를 살짝 내밀고 심옥을 향해 미소를 띤 김선영의 표정\n\nLOCATION (lock): Inside the restaurant dining area, beside the tables just outside the kitchen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just below 김선영’s eye line and over 심옥’s near shoulder, frame a tight three-quarter close-up with 김선영’s playful face in the upper center and the banknotes retained along the lower edge. 김선영 leans toward 심옥, briefly extends her tongue, and keeps her eyes fixed on her mother while the static camera lets the teasing gesture land.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 1,000-won banknotes (held in 김선영's hand) — The visible faces of the notes are angled upward enough to read as 1,000-won bills; used as Small but readable action detail at the lower edge of the close-up; restaurant tables (remaining in the dining area) — Their edges recede obliquely behind 김선영; used as Soft contextual depth behind the exchange.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient light appropriate to the nighttime restaurant is graded in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 김선영 right now — 김선영's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 김선영 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 안내판: \"물은 셀프\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 지폐를 손에 쥔 채 혀를 살짝 내밀고 심옥을 향해 미소를 띤 김선영의 표정\n\nLOCATION (lock): Inside the restaurant dining area, beside the tables just outside the kitchen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just below 김선영’s eye line and over 심옥’s near shoulder, frame a tight three-quarter close-up with 김선영’s playful face in the upper center and the banknotes retained along the lower edge. 김선영 leans toward 심옥, briefly extends her tongue, and keeps her eyes fixed on her mother while the static camera lets the teasing gesture land.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 1,000-won banknotes (held in 김선영's hand) — The visible faces of the notes are angled upward enough to read as 1,000-won bills; used as Small but readable action detail at the lower edge of the close-up; restaurant tables (remaining in the dining area) — Their edges recede obliquely behind 김선영; used as Soft contextual depth behind the exchange.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient light appropriate to the nighttime restaurant is graded in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 김선영 right now — 김선영's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 김선영 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 안내판: \"물은 셀프\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 지폐를 손에 쥔 채 혀를 살짝 내밀고 심옥을 향해 미소를 띤 김선영의 표정\n\nLOCATION (lock): Inside the restaurant dining area, beside the tables just outside the kitchen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just below 김선영’s eye line and over 심옥’s near shoulder, frame a tight three-quarter close-up with 김선영’s playful face in the upper center and the banknotes retained along the lower edge. 김선영 leans toward 심옥, briefly extends her tongue, and keeps her eyes fixed on her mother while the static camera lets the teasing gesture land.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 1,000-won banknotes (held in 김선영's hand) — The visible faces of the notes are angled upward enough to read as 1,000-won bills; used as Small but readable action detail at the lower edge of the close-up; restaurant tables (remaining in the dining area) — Their edges recede obliquely behind 김선영; used as Soft contextual depth behind the exchange.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Low-key ambient light appropriate to the nighttime restaurant is graded in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 김선영 right now — 김선영's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 김선영 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 안내판: \"물은 셀프\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S2sh8__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S2sh8.png"
    },
    {
     "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870689>"
    },
    {
     "label": "PROP REFERENCE — 2001년형 천 원권 지폐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1244193>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L02B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870689>"
    },
    {
     "label": "PROP REFERENCE — 2001년형 천 원권 지폐: the exact object appearing in this shot; match its look, material and wear exactly.",
     "path": "<bytes:1244193>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "지정된 의상(보라색 패딩, 비니)과 지폐 소품을 정확히 구현했으며, 요구된 구도와 표정을 훌륭하게 연출함."
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "배경 구조는 일치하나, 지정된 의상을 완전히 누락했고 지폐에 '1001'이라는 왜곡된 숫자가 표기되어 충실도가 떨어짐."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 1.571
     },
     "adjusted": {
      "A": 1.714,
      "B": 1.571
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "지정된 의상(보라색 패딩, 비니)과 지폐 소품을 정확히 구현했으며, 요구된 구도와 표정을 훌륭하게 연출함."
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "배경 구조는 일치하나, 지정된 의상을 완전히 누락했고 지폐에 '1001'이라는 왜곡된 숫자가 표기되어 충실도가 떨어짐."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "장소 레퍼런스를 무시하고 다른 배경을 생성했으나, 지정된 의상을 정확히 착용하고 주인공이 직접 지폐를 들고 있는 핵심 액션을 올바르게 구현하여 더 높은 점수를 부여함."
     },
     {
      "label": "A",
      "score": 1667,
      "verdict_ko": "배경 구조는 레퍼런스와 유사하나, 인물의 지정 의상이 완전히 누락되었고 지폐에 '1001'이라는 오타가 있으며, 무엇보다 주인공이 아닌 전경 인물이 지폐를 들고 있어 핵심 액션을 위반함."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.5
     },
     "adjusted": {
      "A": 1.667,
      "B": 1.5
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "장소 레퍼런스를 무시하고 다른 배경을 생성했으나, 지정된 의상을 정확히 착용하고 주인공이 직접 지폐를 들고 있는 핵심 액션을 올바르게 구현하여 더 높은 점수를 부여함."
     },
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "배경 구조는 레퍼런스와 유사하나, 인물의 지정 의상이 완전히 누락되었고 지폐에 '1001'이라는 오타가 있으며, 무엇보다 주인공이 아닌 전경 인물이 지폐를 들고 있어 핵심 액션을 위반함."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 3214,
     "B": 3238
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 3214,
   "B": 3238
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "지정된 의상(보라색 패딩, 비니)과 지폐 소품을 정확히 구현했으며, 요구된 구도와 표정을 훌륭하게 연출함."
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "배경 구조는 일치하나, 지정된 의상을 완전히 누락했고 지폐에 '1001'이라는 왜곡된 숫자가 표기되어 충실도가 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L02B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870689>"
   },
   {
    "label": "PROP REFERENCE — 2001년형 천 원권 지폐: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1244193>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "지폐를 쥐고 있는 주된 손이 프롬프트의 지시(김선영의 손)와 달리 카메라 앞쪽 인물(심옥)의 손으로 렌더링되었습니다.",
     "fix_en": "Redraw the hand holding the banknotes at the bottom edge so it belongs to the girl facing the camera; her arm must reach up from her own side with her fingers gripping the money. Preserve the girl's face, expression, hair, clothing, the foreground character's shoulder, the background restaurant set, the signage, and the lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "지폐 중 하나에 '1001'이라는 잘못된 숫자가 인쇄되어 있으며, 레퍼런스의 2001년형 지폐와 디자인이 일치하지 않습니다.",
     "fix_en": "Fan the banknotes tightly and angle them obliquely away from the camera, using shallow focus to blur the printed surfaces so specific text and numbers are illegible, making them read only as generic green currency edges. Preserve the girl's face, expression, hair, clothing, the foreground character, the restaurant interior, and the lighting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "김선영이 레퍼런스 이미지에 명시된 의상(보라색 패딩)이 아닌 베이지색 재킷을 입고 있습니다.",
     "fix_en": "Change the girl's beige jacket to a purple puffer jacket and add a black beanie, matching the character reference.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "화면 중앙 김선영의 얼굴이 레퍼런스 인물과 다르다.",
     "fix_en": "Adjust the girl's facial features to exactly match the provided character reference identity.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "입이 크게 벌어져 혀가 살짝이 아니라 깊게 나와 있다.",
     "fix_en": "Reduce the size of the mouth opening so the tongue protrudes only slightly past the teeth.",
     "severity": "minor",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "지폐를 쥐고 있는 주된 손이 프롬프트의 지시(김선영의 손)와 달리 카메라 앞쪽 인물(심옥)의 손으로 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "지폐 중 하나에 '1001'이라는 잘못된 숫자가 인쇄되어 있으며, 레퍼런스의 2001년형 지폐와 디자인이 일치하지 않습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "김선영이 레퍼런스 이미지에 명시된 의상(보라색 패딩)이 아닌 베이지색 재킷을 입고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙 김선영의 얼굴이 레퍼런스 인물과 다르다.",
     "severity": "major"
    },
    {
     "issue_ko": "김선영이 레퍼런스의 보라 패딩·비니가 아니라 베이지 재킷만 입고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "하단 지폐들에 1001 등 글자가 깨져 있고 2001년 천 원권 디자인이 아니다.",
     "severity": "major"
    },
    {
     "issue_ko": "입이 크게 벌어져 혀가 살짝이 아니라 깊게 나와 있다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "김선영이 레퍼런스 이미지에 명시된 의상(보라색 패딩)이 아닌 베이지색 재킷을 입고 있습니다.",
    "fix_en": "Change the girl's beige jacket to a purple puffer jacket and add a black beanie, matching the character reference.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "화면 중앙 김선영의 얼굴이 레퍼런스 인물과 다르다.",
    "fix_en": "Adjust the girl's facial features to exactly match the provided character reference identity.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "입이 크게 벌어져 혀가 살짝이 아니라 깊게 나와 있다.",
    "fix_en": "Reduce the size of the mouth opening so the tongue protrudes only slightly past the teeth.",
    "severity": "minor",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the hand holding the banknotes at the bottom edge so it belongs to the girl facing the camera; her arm must reach up from her own side with her fingers gripping the money. Preserve the girl's face, expression, hair, clothing, the foreground character's shoulder, the background restaurant set, the signage, and the lighting.\n- Fan the banknotes tightly and angle them obliquely away from the camera, using shallow focus to blur the printed surfaces so specific text and numbers are illegible, making them read only as generic green currency edges. Preserve the girl's face, expression, hair, clothing, the foreground character, the restaurant interior, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "심옥 어깨 너머 타이트 클로즈업·혀를 내밀고 어머니를 보는 미소·하단 지폐는 샷 텍스트에 맞지만, 레퍼런스 얼굴·비니·보라 패딩이 다른 인물·의상으로 바뀌고 지폐 일부가 좌우반전됐다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "신원·보라 패딩·비니·장소 고정을 지켰으나 와이드 정면 스탠딩으로 혀·미소·심옥 방향·오버숄더 클로즈업이 전부 빠졌고, 계약이 묻지 않은 전신·전경을 열어 손실이 더 크다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "김선영이 프레임 왼쪽 전경의 짧은 머리 인물(심옥으로 읽히는 어깨·뒷머리)을 바라보고, 혀를 살짝 내민 채 미소한다. 시선은 카메라가 아니라 그 사람에게 고정되어 있다. 지폐 다발은 아래쪽 가장자리를 향해 펼쳐져 있다.",
      "built_space": "식당 내부 주방 옆이다. 왼쪽에 스테인리스 카운터와 선반, 벽면 흰 안내판, 오른쪽에 나무 테이블과 등받이 의자가 비스듬히 물러나고, 뒤쪽 유리문 너머로 밤 불빛이 보인다. 조향장치 같은 중복 조작면은 없다.",
      "entities": "전면에 짧은 갈색 머리의 성인 여성 뒷모습(심옥으로 추정)이 있고, 주인공은 17세쯤 한국 소녀로 긴 생머리에 베이지색 집업을 입어 레퍼런스의 비니·보라 패딩·동일 얼굴과 불일치한다. 손에 천 원권 여러 장(이황 초상)이 있으나 왼쪽 장들은 1001처럼 좌우반전되어 있다. 벽 안내판에 ‘물은 셀프’가 있다.",
      "hard_violations": [],
      "physics": "김선영의 손이 지폐 다발을 엄지로 쥐고 있어 지폐를 받친다. 상체는 상대 쪽으로 기울인 자세이며 하중은 몸통·팔에 있다. 공중에 뜬 신체나 손 없는 부유물은 없다."
     },
     {
      "label": "B",
      "direction": "김선영이 카메라를 정면으로 응시하고, 오른팔을 렌즈 쪽으로 곧게 뻗어 지폐 묶음을 내민다. 심옥을 향한 시선·몸 기울임·혀 내밀기는 전혀 없다.",
      "built_space": "위치 사진과 거의 같은 식당 전경이다. 왼쪽 싱크·가스대·달력, 정면 유리 미닫이문, 오른쪽 테이블·의자 열과 벽걸이 TV가 보이고, 인물은 주방 카운터 앞 통로에 서 있다. 조작면 중복은 없다.",
      "entities": "레퍼런스와 같은 김선영(회색 비니, 긴 생머리, 보라 패딩, 검정 이너, 청바지, 벨트)이 맞고 2001년형 천 원권 묶음도 초상·‘천원’이 읽힌다. 심옥은 없고 ‘물은 셀프’ 안내판도 보이지 않는다. 표정을 혀 내민 미소가 아닌 무표정이다.",
      "hard_violations": [],
      "physics": "두 발이 바닥에 닿아 선 채 오른손이 지폐 묶음을 쥐고 팔을 뻗는다. 지지가 분명하고 부유는 없다. 다만 그 포즈는 샷이 요구한 기대어 혀를 내미는 동작이 아니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "single_openrouter:x-ai/grok-4.6",
     "models": [
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "심옥 어깨 너머 타이트 클로즈업·혀를 내밀고 어머니를 보는 미소·하단 지폐는 샷 텍스트에 맞지만, 레퍼런스 얼굴·비니·보라 패딩이 다른 인물·의상으로 바뀌고 지폐 일부가 좌우반전됐다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "신원·보라 패딩·비니·장소 고정을 지켰으나 와이드 정면 스탠딩으로 혀·미소·심옥 방향·오버숄더 클로즈업이 전부 빠졌고, 계약이 묻지 않은 전신·전경을 열어 손실이 더 크다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "김선영이 프레임 왼쪽 전경의 짧은 머리 인물(심옥으로 읽히는 어깨·뒷머리)을 바라보고, 혀를 살짝 내민 채 미소한다. 시선은 카메라가 아니라 그 사람에게 고정되어 있다. 지폐 다발은 아래쪽 가장자리를 향해 펼쳐져 있다.",
      "built_space": "식당 내부 주방 옆이다. 왼쪽에 스테인리스 카운터와 선반, 벽면 흰 안내판, 오른쪽에 나무 테이블과 등받이 의자가 비스듬히 물러나고, 뒤쪽 유리문 너머로 밤 불빛이 보인다. 조향장치 같은 중복 조작면은 없다.",
      "entities": "전면에 짧은 갈색 머리의 성인 여성 뒷모습(심옥으로 추정)이 있고, 주인공은 17세쯤 한국 소녀로 긴 생머리에 베이지색 집업을 입어 레퍼런스의 비니·보라 패딩·동일 얼굴과 불일치한다. 손에 천 원권 여러 장(이황 초상)이 있으나 왼쪽 장들은 1001처럼 좌우반전되어 있다. 벽 안내판에 ‘물은 셀프’가 있다.",
      "hard_violations": [],
      "physics": "김선영의 손이 지폐 다발을 엄지로 쥐고 있어 지폐를 받친다. 상체는 상대 쪽으로 기울인 자세이며 하중은 몸통·팔에 있다. 공중에 뜬 신체나 손 없는 부유물은 없다."
     },
     {
      "label": "B",
      "direction": "김선영이 카메라를 정면으로 응시하고, 오른팔을 렌즈 쪽으로 곧게 뻗어 지폐 묶음을 내민다. 심옥을 향한 시선·몸 기울임·혀 내밀기는 전혀 없다.",
      "built_space": "위치 사진과 거의 같은 식당 전경이다. 왼쪽 싱크·가스대·달력, 정면 유리 미닫이문, 오른쪽 테이블·의자 열과 벽걸이 TV가 보이고, 인물은 주방 카운터 앞 통로에 서 있다. 조작면 중복은 없다.",
      "entities": "레퍼런스와 같은 김선영(회색 비니, 긴 생머리, 보라 패딩, 검정 이너, 청바지, 벨트)이 맞고 2001년형 천 원권 묶음도 초상·‘천원’이 읽힌다. 심옥은 없고 ‘물은 셀프’ 안내판도 보이지 않는다. 표정을 혀 내민 미소가 아닌 무표정이다.",
      "hard_violations": [],
      "physics": "두 발이 바닥에 닿아 선 채 오른손이 지폐 묶음을 쥐고 팔을 뻗는다. 지지가 분명하고 부유는 없다. 다만 그 포즈는 샷이 요구한 기대어 혀를 내미는 동작이 아니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "김선영 얼굴·비니·보라 패딩 정체성과 장소는 맞지만 계약이 요구한 심옥 오버숄더 클로즈업·혀 내밀기·미소가 없고 카메라를 향한 와이드 정면이라 순간 자체가 성립하지 않는다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "혀·미소·지폐·오버숄더 연출은 맞추었으나 김선영이 다른 얼굴·생머리·베이지 재킷으로 교체되어 정체성이 상실됐고 심옥 뒷모습까지 추가되어 탈락한다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "김선영이 정면 카메라를 응시하며 지폐 뭉치를 렌즈 쪽으로 내민다. 심옥은 없고 시선·미소·혀가 심옥에 닿지 않으며 표적은 카메라다.",
      "built_space": "식당 홀 전체 와이드로 왼쪽 스테인리스 주방, 달력, 유리 출입문, 오른쪽 식탁·의자·벽걸이 TV가 위치 사진과 거의 같은 자리·비율로 보인다. 인물은 홀 중앙에 서서 카메라를 향한다.",
      "entities": "인물은 레퍼런스와 같은 한국인 소녀, 검은 비니, 보라 패딩, 긴 생머리, 청바지다. 손에 2001년형 천 원권 뭉치가 있다. 심옥은 없고 ‘물은 셀프’ 안내판도 없다.",
      "hard_violations": [],
      "physics": "양발이 바닥에 닿아 서 있고 오른손이 지폐 뭉치를 쥐고 앞으로 뻗는다. 지지가 있다."
     },
     {
      "label": "B",
      "direction": "전면 인물이 왼쪽 전경의 뒷모습(심옥으로 읽힘)을 보고 혀를 내밀며 미소한다. 시선이 그 어깨 너머 인물에게 착지하고 지폐는 둘 사이에 들려 있다.",
      "built_space": "왼쪽 뒤에 스테인리스 주방과 ‘물은 셀프’ 안내판, 오른쪽에 유리문과 식탁 가장자리가 보인다. 왼쪽 전경에 다른 사람의 머리·어깨가 오버숄더로 들어온다.",
      "entities": "전면 인물은 레퍼런스 김선영이 아니다. 비니 없는 다른 얼굴·헤어, 베이지 재킷이다. 짧은 갈색 머리의 심옥 뒷모습이 추가됐다. 천 원권 여러 장(일부 숫자 1001)과 안내판은 있다.",
      "hard_violations": [
       "샷이 보이지 말라고 한 심옥의 머리·어깨가 들어와 인물이 추가됨",
       "김선영이 레퍼런스와 다른 얼굴·헤어·의상의 인물로 교체됨"
      ],
      "physics": "손이 부채꼴 지폐를 쥐고 있고 상체는 상대 쪽으로 숙인다. 혀는 입에서 나와 있으며 공중 부유는 없다."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "single_openrouter:x-ai/grok-4.6",
     "models": [
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "김선영 얼굴·비니·보라 패딩 정체성과 장소는 맞지만 계약이 요구한 심옥 오버숄더 클로즈업·혀 내밀기·미소가 없고 카메라를 향한 와이드 정면이라 순간 자체가 성립하지 않는다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "혀·미소·지폐·오버숄더 연출은 맞추었으나 김선영이 다른 얼굴·생머리·베이지 재킷으로 교체되어 정체성이 상실됐고 심옥 뒷모습까지 추가되어 탈락한다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "김선영이 정면 카메라를 응시하며 지폐 뭉치를 렌즈 쪽으로 내민다. 심옥은 없고 시선·미소·혀가 심옥에 닿지 않으며 표적은 카메라다.",
      "built_space": "식당 홀 전체 와이드로 왼쪽 스테인리스 주방, 달력, 유리 출입문, 오른쪽 식탁·의자·벽걸이 TV가 위치 사진과 거의 같은 자리·비율로 보인다. 인물은 홀 중앙에 서서 카메라를 향한다.",
      "entities": "인물은 레퍼런스와 같은 한국인 소녀, 검은 비니, 보라 패딩, 긴 생머리, 청바지다. 손에 2001년형 천 원권 뭉치가 있다. 심옥은 없고 ‘물은 셀프’ 안내판도 없다.",
      "hard_violations": [],
      "physics": "양발이 바닥에 닿아 서 있고 오른손이 지폐 뭉치를 쥐고 앞으로 뻗는다. 지지가 있다."
     },
     {
      "label": "A",
      "direction": "전면 인물이 왼쪽 전경의 뒷모습(심옥으로 읽힘)을 보고 혀를 내밀며 미소한다. 시선이 그 어깨 너머 인물에게 착지하고 지폐는 둘 사이에 들려 있다.",
      "built_space": "왼쪽 뒤에 스테인리스 주방과 ‘물은 셀프’ 안내판, 오른쪽에 유리문과 식탁 가장자리가 보인다. 왼쪽 전경에 다른 사람의 머리·어깨가 오버숄더로 들어온다.",
      "entities": "전면 인물은 레퍼런스 김선영이 아니다. 비니 없는 다른 얼굴·헤어, 베이지 재킷이다. 짧은 갈색 머리의 심옥 뒷모습이 추가됐다. 천 원권 여러 장(일부 숫자 1001)과 안내판은 있다.",
      "hard_violations": [
       "샷이 보이지 말라고 한 심옥의 머리·어깨가 들어와 인물이 추가됨",
       "김선영이 레퍼런스와 다른 얼굴·헤어·의상의 인물로 교체됨"
      ],
      "physics": "손이 부채꼴 지폐를 쥐고 있고 상체는 상대 쪽으로 숙인다. 혀는 입에서 나와 있으며 공중 부유는 없다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S2sh8__bgfirst_bg.png",
   "bg_asset_id": "4d9ed8fe-445d-449f-a769-37fa97f3b65d",
   "bg_record_key": "S2sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S2sh8::cine": {
  "applied": true,
  "fingerprint": "005ec5b5d2aae9070f390cbf3ba9a27ff2f8ae7aa97963d0962a9ca7ae197243",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S2sh8_sel.png",
  "source_sha256": "989bd65983eabefa3133e37cb1f4070d9dc006a8fdb87ecd29ac5051536e36ee",
  "file": "S2sh8_cine.png",
  "latency_ms": 10554
 },
 "S2sh9::signage": {
  "fp": "dd1d0de1fa969cd4",
  "inscriptions": [
   {
    "surface_native": "출입 유리문",
    "text_native": "당기시오",
    "reason_ko": "한국 식당의 유리 출입문에 흔히 붙어 있는 안내 문구로, 인물이 문을 여닫는 상황의 사실감을 더하기 위해 필요합니다."
   }
  ]
 },
 "S2sh9": {
  "input_fingerprint": "0265c26f45715cd0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한 손으로 출입문을 반쯤 밀어 연 채 서 있는 선영의 아버지와, 그를 향해 허리를 굽혀 고개를 숙인 자세의 김선영\n\nLOCATION (lock): At the restaurant's interior entrance, immediately inside the half-open front door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height inside the dining area, track parallel to 김선영 and open into an oblique medium two-shot of the entrance, with 선영의 아버지 holding the partly opened door on one side and 김선영 bending toward him on the other. His hand and the door remain fully visible as her hurried bow is caught mid-stride, both absorbed in their brief passing greeting rather than the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 김선영 in the middle-right of the frame, midground, moves toward restaurant exit; restaurant exit in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: restaurant entrance door (partly open) — The partly opened door is seen obliquely from the dining-area side, with the father's hand on it; used as Shared threshold that stages the father entering as 김선영 leaves; dining-area tables (present) — Table edges recede diagonally into the restaurant interior; used as Environmental context behind the brief doorway encounter.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient nighttime restaurant illumination remains naturalistic, subdued, and neutrally graded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 선영의 아버지 right now — 선영의 아버지's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 선영의 아버지 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리); 선영의 아버지 (Korean 남성, 중년, 둥근 얼굴, 짧은 검은색 머리) — wearing: 낡은 겨울 점퍼와 편안한 코듀로이 바지를 매치한 수수한 일상복 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 출입 유리문: \"당기시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한 손으로 출입문을 반쯤 밀어 연 채 서 있는 선영의 아버지와, 그를 향해 허리를 굽혀 고개를 숙인 자세의 김선영\n\nLOCATION (lock): At the restaurant's interior entrance, immediately inside the half-open front door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height inside the dining area, track parallel to 김선영 and open into an oblique medium two-shot of the entrance, with 선영의 아버지 holding the partly opened door on one side and 김선영 bending toward him on the other. His hand and the door remain fully visible as her hurried bow is caught mid-stride, both absorbed in their brief passing greeting rather than the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 김선영 in the middle-right of the frame, midground, moves toward restaurant exit; restaurant exit in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: restaurant entrance door (partly open) — The partly opened door is seen obliquely from the dining-area side, with the father's hand on it; used as Shared threshold that stages the father entering as 김선영 leaves; dining-area tables (present) — Table edges recede diagonally into the restaurant interior; used as Environmental context behind the brief doorway encounter.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient nighttime restaurant illumination remains naturalistic, subdued, and neutrally graded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 선영의 아버지 right now — 선영의 아버지's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 선영의 아버지 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리); 선영의 아버지 (Korean 남성, 중년, 둥근 얼굴, 짧은 검은색 머리) — wearing: 낡은 겨울 점퍼와 편안한 코듀로이 바지를 매치한 수수한 일상복 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 출입 유리문: \"당기시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 한 손으로 출입문을 반쯤 밀어 연 채 서 있는 선영의 아버지와, 그를 향해 허리를 굽혀 고개를 숙인 자세의 김선영\n\nLOCATION (lock): At the restaurant's interior entrance, immediately inside the half-open front door. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height inside the dining area, track parallel to 김선영 and open into an oblique medium two-shot of the entrance, with 선영의 아버지 holding the partly opened door on one side and 김선영 bending toward him on the other. His hand and the door remain fully visible as her hurried bow is caught mid-stride, both absorbed in their brief passing greeting rather than the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 김선영 in the middle-right of the frame, midground, moves toward restaurant exit; restaurant exit in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: restaurant entrance door (partly open) — The partly opened door is seen obliquely from the dining-area side, with the father's hand on it; used as Shared threshold that stages the father entering as 김선영 leaves; dining-area tables (present) — Table edges recede diagonally into the restaurant interior; used as Environmental context behind the brief doorway encounter.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient nighttime restaurant illumination remains naturalistic, subdued, and neutrally graded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 선영의 아버지 right now — 선영의 아버지's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 선영의 아버지 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리); 선영의 아버지 (Korean 남성, 중년, 둥근 얼굴, 짧은 검은색 머리) — wearing: 낡은 겨울 점퍼와 편안한 코듀로이 바지를 매치한 수수한 일상복 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 출입 유리문: \"당기시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "아버지는 선영을 응시하고, 선영은 아버지를 향해 고개 숙여 인사함.",
    "built_space": "식당 내부 출입문 앞. 이전 샷의 '물은 셀프' 표지판과 식탁 배치가 일치함.",
    "entities": "두 인물 모두 요구된 외형과 복장을 갖춤. '당기시오' 텍스트 정확함.",
    "hard_violations": [],
    "physics": "아버지의 손이 문손잡이를 단단히 쥐고 있으며, 두 인물 모두 바닥에 지지하여 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "아버지는 선영을 응시하고, 선영은 인사 중임.",
    "built_space": "식당 내부이나, 지정된 벽면 배경 요소가 다른 형태로 변경됨.",
    "entities": "인물 외형과 복장, 출입문 텍스트는 지시사항과 일치함.",
    "hard_violations": [
     "고정된 장소 설정(이전 샷의 배경)을 위반하고 벽면에 임의의 표지판과 문자를 생성함."
    ],
    "physics": "인물들의 자세와 문을 잡은 손의 물리적 지탱이 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "이전 샷의 배경 요소('물은 셀프' 표지판)를 그대로 유지하면서 인물의 동작, 구도, 텍스트 지시를 완벽하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "이전 샷과 동일해야 할 배경에 지시되지 않은 표지판과 왜곡된 문자를 임의로 추가하여 장소의 연속성을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S2sh8_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870689>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "레퍼런스 이미지에서 '물은 셀프' 안내문 아래에 있던 막힌 형태의 스테인리스 작업대가 개방형 다단 선반으로 변경되어 배경 연속성에 어긋납니다.",
     "fix_en": "Replace the open metal shelving unit below the sign with a solid, enclosed stainless steel counter. Keep the sign, characters, and overall room lighting unchanged.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "아버지가 출입문을 반쯤 밀어 연 것이 아니라 거의 닫힌 문을 정면에서 손잡이만 잡고 있다",
     "fix_en": "Redraw the glass door so it is pushed halfway open, with the man's hand actively pushing the frame or glass. Keep the characters' faces, clothing, and the background intact.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "김선영이 출구 쪽으로 걷다 허리 숙여 인사하는 중이 아니라 양손을 주머니에 넣고 서 있다",
     "fix_en": "Redraw the woman's arms so her hands are out of her pockets and held naturally at her sides or front for a mid-stride bow. Keep her face, jacket, the man, and the set exactly as they are.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "선영의 아버지가 낡은 겨울 점퍼가 아닌 갈색 니트 가디건을 입고 있다",
     "fix_en": "Change the man's brown knit jacket into a worn, slightly bulky winter jumper. Keep his face, pose, the woman, and the background elements strictly unchanged.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "식당 출입구가 중좌 배경의 사선이 아니라 프레임 정중앙에 정면으로 놓여 있다",
     "fix_en": "Shift the doorway to the middle-left background and angle it obliquely to the camera.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "식탁이 실내로 후퇴하는 배경이 아니라 화면 하단 전경에 크게 들어와 있다",
     "fix_en": "Remove the foreground tables, leaving only the background tables receding into the interior.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "레퍼런스 이미지에서 '물은 셀프' 안내문 아래에 있던 막힌 형태의 스테인리스 작업대가 개방형 다단 선반으로 변경되어 배경 연속성에 어긋납니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "아버지가 출입문을 반쯤 밀어 연 것이 아니라 거의 닫힌 문을 정면에서 손잡이만 잡고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "김선영이 출구 쪽으로 걷다 허리 숙여 인사하는 중이 아니라 양손을 주머니에 넣고 서 있다",
     "severity": "major"
    },
    {
     "issue_ko": "선영의 아버지가 낡은 겨울 점퍼가 아닌 갈색 니트 가디건을 입고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "식당 출입구가 중좌 배경의 사선이 아니라 프레임 정중앙에 정면으로 놓여 있다",
     "severity": "major"
    },
    {
     "issue_ko": "식탁이 실내로 후퇴하는 배경이 아니라 화면 하단 전경에 크게 들어와 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 6,
  "fix_severity_skipped": [
   {
    "issue_ko": "레퍼런스 이미지에서 '물은 셀프' 안내문 아래에 있던 막힌 형태의 스테인리스 작업대가 개방형 다단 선반으로 변경되어 배경 연속성에 어긋납니다.",
    "fix_en": "Replace the open metal shelving unit below the sign with a solid, enclosed stainless steel counter. Keep the sign, characters, and overall room lighting unchanged.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "아버지가 출입문을 반쯤 밀어 연 것이 아니라 거의 닫힌 문을 정면에서 손잡이만 잡고 있다",
    "fix_en": "Redraw the glass door so it is pushed halfway open, with the man's hand actively pushing the frame or glass. Keep the characters' faces, clothing, and the background intact.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "김선영이 출구 쪽으로 걷다 허리 숙여 인사하는 중이 아니라 양손을 주머니에 넣고 서 있다",
    "fix_en": "Redraw the woman's arms so her hands are out of her pockets and held naturally at her sides or front for a mid-stride bow. Keep her face, jacket, the man, and the set exactly as they are.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "선영의 아버지가 낡은 겨울 점퍼가 아닌 갈색 니트 가디건을 입고 있다",
    "fix_en": "Change the man's brown knit jacket into a worn, slightly bulky winter jumper. Keep his face, pose, the woman, and the background elements strictly unchanged.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "식당 출입구가 중좌 배경의 사선이 아니라 프레임 정중앙에 정면으로 놓여 있다",
    "fix_en": "Shift the doorway to the middle-left background and angle it obliquely to the camera.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "식탁이 실내로 후퇴하는 배경이 아니라 화면 하단 전경에 크게 들어와 있다",
    "fix_en": "Remove the foreground tables, leaving only the background tables receding into the interior.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S2sh8"
  }
 },
 "S2sh9::cine": {
  "applied": true,
  "fingerprint": "4eff9e05d873915ca8b45a710c54b8aecfaee8a22bf5854d9cbe7a942c0a32e2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S2sh9_sel.png",
  "source_sha256": "fb1ff1b72cd39ea7be2b41f8f234d8b1ed791f20746eedbc599492f4a6026c1e",
  "file": "S2sh9_cine.png",
  "latency_ms": 10640
 },
 "S3sh3::signage": {
  "fp": "9b62b34a1af95975",
  "inscriptions": []
 },
 "era_assess::b157f24e1be7aecf": {
  "subjects": [
   {
    "subject_native": "2000년대 한국의 가정용 집전화기",
    "search_terms_native": [
     "2000년대 집전화기",
     "가정용 유선전화기",
     "바텔 전화기",
     "추억의 집전화"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성되어야 하며, 영어나 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 초중반 한국 가정에서 흔히 사용되던 대기업 및 바텔 등의 유무선 집전화기는 특유의 뭉툭한 형태, 버튼 배치, 충전 거치대 형태를 지니고 있어 일반적인 서구식 전화기와 디자인이 다릅니다."
   }
  ]
 },
 "era_ref::b513c4bd2853f799": {
  "subject": "2000년대 한국의 가정용 집전화기",
  "terms": [
   "2000년대 집전화기",
   "가정용 유선전화기",
   "바텔 전화기",
   "추억의 집전화"
  ],
  "queries": [
   [
    "2000년대 한국 가정용 유선전화기 바텔 전화기 추억의 집전화",
    "옛날 바텔 전화기 가정용 집전화기 2000년대"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://jnjcorp.diskn.com/ZZ/M/DK_8098266_M00_01.jpg",
  "picked_reason_ko": "4번은 2000년대 한국 가정에서 흔히 쓰인 유선 집전화기의 본체, 수화기, 숫자 버튼, 발신자 표시창과 색상을 가장 선명하고 온전하게 보여준다.",
  "sha256": "d18d8532fecb8dd09e90b3fc382ecd1c071db484784c3c23013b22af29957236",
  "file": "eraref_b513c4bd2853f799.png"
 },
 "S3sh3::bgfirst_bg": {
  "input_fingerprint": "fe4f427da21b0353",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 민정을 향해 힐끗 시선을 둔 채 수화기 너머로 속삭이듯 고개를 꺾고 있는 김선영\n\nLOCATION (lock): Inside the dark shared bedroom, beside the sleeping sibling and the house telephone.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At mattress height beside 김선영, the settling pan forms a close three-quarter composition with her and the telephone receiver in the foreground while 민정 lies deeper along the same sightline. 김선영 tips her head into the receiver and glances sideways at 민정, whose closed eyes and turned body remain softly legible behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 김선영 in the middle-left of the frame, foreground, looks toward sleeping 민정; 민정 (2001년) in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: telephone receiver (in use) — Its inner side follows the diagonal between 김선영's ear and mouth; used as Held close to 김선영's mouth as the physical reason for her lowered voice; mattress and bedding (occupied by 민정); used as Low horizontal plane linking the foreground caller and sleeping child in depth.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explicitly dark room is rendered with low ambient exposure and restrained neutral-to-cool tonality, without assigning an unmentioned source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 가정용 집전화기: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 민정을 향해 힐끗 시선을 둔 채 수화기 너머로 속삭이듯 고개를 꺾고 있는 김선영\n\nLOCATION (lock): Inside the dark shared bedroom, beside the sleeping sibling and the house telephone.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At mattress height beside 김선영, the settling pan forms a close three-quarter composition with her and the telephone receiver in the foreground while 민정 lies deeper along the same sightline. 김선영 tips her head into the receiver and glances sideways at 민정, whose closed eyes and turned body remain softly legible behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 김선영 in the middle-left of the frame, foreground, looks toward sleeping 민정; 민정 (2001년) in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: telephone receiver (in use) — Its inner side follows the diagonal between 김선영's ear and mouth; used as Held close to 김선영's mouth as the physical reason for her lowered voice; mattress and bedding (occupied by 민정); used as Low horizontal plane linking the foreground caller and sleeping child in depth.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explicitly dark room is rendered with low ambient exposure and restrained neutral-to-cool tonality, without assigning an unmentioned source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 가정용 집전화기: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S3sh3__bgfirst_bg.png",
  "asset_id": "010ba4d8-3ea5-4ca4-b5bd-807bf6f444ad",
  "input_asset_ids": [
   "aff59159-6636-4f38-8282-db3401f4ffbe",
   "cdee95d6-431b-4b8e-bdc4-50b2a81427ce"
  ],
  "era_research": {
   "subject": "2000년대 한국의 가정용 집전화기",
   "queries": [
    [
     "2000년대 한국 가정용 유선전화기 바텔 전화기 추억의 집전화",
     "옛날 바텔 전화기 가정용 집전화기 2000년대"
    ]
   ],
   "picked_url": "https://jnjcorp.diskn.com/ZZ/M/DK_8098266_M00_01.jpg",
   "sha256": "d18d8532fecb8dd09e90b3fc382ecd1c071db484784c3c23013b22af29957236",
   "file": "eraref_b513c4bd2853f799.png"
  }
 },
 "S3sh3": {
  "input_fingerprint": "0be2708e7e17a476",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 민정을 향해 힐끗 시선을 둔 채 수화기 너머로 속삭이듯 고개를 꺾고 있는 김선영\n\nLOCATION (lock): Inside the dark shared bedroom, beside the sleeping sibling and the house telephone. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At mattress height beside 김선영, the settling pan forms a close three-quarter composition with her and the telephone receiver in the foreground while 민정 lies deeper along the same sightline. 김선영 tips her head into the receiver and glances sideways at 민정, whose closed eyes and turned body remain softly legible behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 김선영 in the middle-left of the frame, foreground, looks toward sleeping 민정; 민정 (2001년) in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: telephone receiver (in use) — Its inner side follows the diagonal between 김선영's ear and mouth; used as Held close to 김선영's mouth as the physical reason for her lowered voice; mattress and bedding (occupied by 민정); used as Low horizontal plane linking the foreground caller and sleeping child in depth.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explicitly dark room is rendered with low ambient exposure and restrained neutral-to-cool tonality, without assigning an unmentioned source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 김선영 right now — 김선영's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 김선영 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 민정을 향해 힐끗 시선을 둔 채 수화기 너머로 속삭이듯 고개를 꺾고 있는 김선영\n\nLOCATION (lock): Inside the dark shared bedroom, beside the sleeping sibling and the house telephone. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At mattress height beside 김선영, the settling pan forms a close three-quarter composition with her and the telephone receiver in the foreground while 민정 lies deeper along the same sightline. 김선영 tips her head into the receiver and glances sideways at 민정, whose closed eyes and turned body remain softly legible behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 김선영 in the middle-left of the frame, foreground, looks toward sleeping 민정; 민정 (2001년) in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: telephone receiver (in use) — Its inner side follows the diagonal between 김선영's ear and mouth; used as Held close to 김선영's mouth as the physical reason for her lowered voice; mattress and bedding (occupied by 민정); used as Low horizontal plane linking the foreground caller and sleeping child in depth.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explicitly dark room is rendered with low ambient exposure and restrained neutral-to-cool tonality, without assigning an unmentioned source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 김선영 right now — 김선영's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 김선영 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 민정을 향해 힐끗 시선을 둔 채 수화기 너머로 속삭이듯 고개를 꺾고 있는 김선영\n\nLOCATION (lock): Inside the dark shared bedroom, beside the sleeping sibling and the house telephone. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At mattress height beside 김선영, the settling pan forms a close three-quarter composition with her and the telephone receiver in the foreground while 민정 lies deeper along the same sightline. 김선영 tips her head into the receiver and glances sideways at 민정, whose closed eyes and turned body remain softly legible behind her.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 김선영 in the middle-left of the frame, foreground, looks toward sleeping 민정; 민정 (2001년) in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: telephone receiver (in use) — Its inner side follows the diagonal between 김선영's ear and mouth; used as Held close to 김선영's mouth as the physical reason for her lowered voice; mattress and bedding (occupied by 민정); used as Low horizontal plane linking the foreground caller and sleeping child in depth.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The explicitly dark room is rendered with low ambient exposure and restrained neutral-to-cool tonality, without assigning an unmentioned source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 김선영 right now — 김선영's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 김선영 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S3sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S3sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870689>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L04B04.png"
    },
    {
     "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:870689>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "카메라 앵글, 샷 크기, 그리고 두 인물 간의 시선 및 배치 관계를 요구사항에 맞게 정확히 구현했습니다. 전화기 색상과 의상이 레퍼런스와 다르나 물리적 오류가 없어 우수합니다."
     },
     {
      "label": "B",
      "score": 929,
      "verdict_ko": "요구된 의상을 잘 반영하고 전반적인 분위기를 잡았으나, 전화기 수화기와 패딩 점퍼가 화면 내에 중복으로 존재하는 치명적인 사물 중복 오류가 발생했습니다.  ★위반: [gemini-pro] 배경 협탁의 전화기 본체에 수화기가 정상적으로 거치되어 있음에도 인물이 또 다른 수화기를 들고 있는 사물 중복 오류 / [gemini-pro] 인물이 착용하고 있는 보라색 패딩이 배경 복도 바닥에도 그대로 존재하는 사물 중복 오류"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.429
     },
     "adjusted": {
      "A": 1.571,
      "B": 0.929
     },
     "violations": {
      "B": [
       "[gemini-pro] 배경 협탁의 전화기 본체에 수화기가 정상적으로 거치되어 있음에도 인물이 또 다른 수화기를 들고 있는 사물 중복 오류",
       "[gemini-pro] 인물이 착용하고 있는 보라색 패딩이 배경 복도 바닥에도 그대로 존재하는 사물 중복 오류"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "카메라 앵글, 샷 크기, 그리고 두 인물 간의 시선 및 배치 관계를 요구사항에 맞게 정확히 구현했습니다. 전화기 색상과 의상이 레퍼런스와 다르나 물리적 오류가 없어 우수합니다."
     },
     {
      "label": "B",
      "score": 929,
      "verdict_ko": "요구된 의상을 잘 반영하고 전반적인 분위기를 잡았으나, 전화기 수화기와 패딩 점퍼가 화면 내에 중복으로 존재하는 치명적인 사물 중복 오류가 발생했습니다.  ★위반: [gemini-pro] 배경 협탁의 전화기 본체에 수화기가 정상적으로 거치되어 있음에도 인물이 또 다른 수화기를 들고 있는 사물 중복 오류 / [gemini-pro] 인물이 착용하고 있는 보라색 패딩이 배경 복도 바닥에도 그대로 존재하는 사물 중복 오류"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "캐릭터 복장(보라색 패딩 누락)과 전화기 색상이 레퍼런스와 다르지만, 치명적인 물리적/공간적 오류 없이 샷 텍스트의 구도와 시선을 적절히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "복장과 인물 구도는 훌륭하나, 배경 협탁에 이미 온전한 전화기가 있음에도 주인공이 또 다른 수화기를 들고 있는 복제 오류(Hard Violation)가 발생해 탈락합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "김선영의 시선이 오른쪽 배경 침대에 누워있는 민정을 향함.",
      "built_space": "배경에 침대와 협탁이 있으며, 협탁 위에 전화기 1대가 놓임. 카메라는 전경과 배경을 일직선으로 담음.",
      "entities": "김선영은 보라색 패딩을 입어 레퍼런스와 일치함. 흰색 수화기를 들고 있으며 배경에 민정이 누워있음.",
      "hard_violations": [
       "배경 협탁에 수화기가 놓인 전화기가 온전히 존재하는데, 전경의 주인공이 또 다른 수화기를 들고 있음 (수화기 복제)"
      ],
      "physics": "오른손이 수화기를 쥐고 지탱하나, 전화선이 화면 아래로 끊어져 본체와 연결되지 않음."
     },
     {
      "label": "B",
      "direction": "김선영의 시선이 오른쪽 배경 침대의 민정을 향함.",
      "built_space": "전경에 탁자가 있고 전화기 본체가 놓여 있음. 배경에 침대가 위치함.",
      "entities": "김선영은 남색 스웨터를 입어 복장 레퍼런스와 불일치함. 전화기가 레퍼런스의 흰색 대신 붉은색임. 배경에 민정이 누워있음.",
      "hard_violations": [],
      "physics": "오른손이 붉은색 수화기를 안정적으로 쥐고 있으며, 선이 탁자 위 본체로 올바르게 이어짐."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "캐릭터 복장(보라색 패딩 누락)과 전화기 색상이 레퍼런스와 다르지만, 치명적인 물리적/공간적 오류 없이 샷 텍스트의 구도와 시선을 적절히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "복장과 인물 구도는 훌륭하나, 배경 협탁에 이미 온전한 전화기가 있음에도 주인공이 또 다른 수화기를 들고 있는 복제 오류(Hard Violation)가 발생해 탈락합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "김선영의 시선이 오른쪽 배경 침대에 누워있는 민정을 향함.",
      "built_space": "배경에 침대와 협탁이 있으며, 협탁 위에 전화기 1대가 놓임. 카메라는 전경과 배경을 일직선으로 담음.",
      "entities": "김선영은 보라색 패딩을 입어 레퍼런스와 일치함. 흰색 수화기를 들고 있으며 배경에 민정이 누워있음.",
      "hard_violations": [
       "배경 협탁에 수화기가 놓인 전화기가 온전히 존재하는데, 전경의 주인공이 또 다른 수화기를 들고 있음 (수화기 복제)"
      ],
      "physics": "오른손이 수화기를 쥐고 지탱하나, 전화선이 화면 아래로 끊어져 본체와 연결되지 않음."
     },
     {
      "label": "A",
      "direction": "김선영의 시선이 오른쪽 배경 침대의 민정을 향함.",
      "built_space": "전경에 탁자가 있고 전화기 본체가 놓여 있음. 배경에 침대가 위치함.",
      "entities": "김선영은 남색 스웨터를 입어 복장 레퍼런스와 불일치함. 전화기가 레퍼런스의 흰색 대신 붉은색임. 배경에 민정이 누워있음.",
      "hard_violations": [],
      "physics": "오른손이 붉은색 수화기를 안정적으로 쥐고 있으며, 선이 탁자 위 본체로 올바르게 이어짐."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1577,
     "B": 932
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 1577,
   "B": 932
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "카메라 앵글, 샷 크기, 그리고 두 인물 간의 시선 및 배치 관계를 요구사항에 맞게 정확히 구현했습니다. 전화기 색상과 의상이 레퍼런스와 다르나 물리적 오류가 없어 우수합니다."
   },
   {
    "label": "B",
    "score": 929,
    "verdict_ko": "요구된 의상을 잘 반영하고 전반적인 분위기를 잡았으나, 전화기 수화기와 패딩 점퍼가 화면 내에 중복으로 존재하는 치명적인 사물 중복 오류가 발생했습니다.  ★위반: [gemini-pro] 배경 협탁의 전화기 본체에 수화기가 정상적으로 거치되어 있음에도 인물이 또 다른 수화기를 들고 있는 사물 중복 오류 / [gemini-pro] 인물이 착용하고 있는 보라색 패딩이 배경 복도 바닥에도 그대로 존재하는 사물 중복 오류"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L04B04.png"
   },
   {
    "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870689>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우측 하단 테이블 위 전화기 본체에 수화기가 그대로 놓여 있음에도 김선영이 또 다른 수화기를 손에 들고 있어, 수화기가 중복으로 생성되었습니다.",
     "fix_en": "Remove the extra telephone receiver resting on the red phone base in the bottom right corner, replacing it with the empty cradle structure of the base. Preserve the foreground girl, her current pose and clothing, the receiver in her hand, the coiled cord, the sleeping girl, the lighting, and the background room exactly.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "스케치의 시선 화살표와 프롬프트 지시(민정을 향해 힐끗 시선을 둔 채)와 달리, 김선영의 시선이 배경에 누워있는 민정을 향하지 않고 우측 하단 바닥 쪽을 향하고 있습니다.",
     "fix_en": "Shift the foreground girl's pupils backward to direct her gaze squarely at the sleeping girl in the background. Preserve her head angle, pose, clothing, the telephone, the sleeping girl, and the background exactly.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "전경 김선영의 얼굴이 캐릭터 레퍼런스와 일치하지 않는다",
     "fix_en": "Redraw the foreground girl's face to match the character reference. Preserve her head angle, expression, clothing, the telephone, the sleeping girl, and the background exactly.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "전경 김선영이 레퍼런스 의상(보라 패딩·비니)이 아닌 남색 스웨트셔츠를 입고 있다",
     "fix_en": "Replace the foreground girl's navy top with the purple puffer jacket and grey beanie from the character reference. Preserve her face, hands, the telephone, the sleeping girl, and the background exactly.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우측 하단 테이블 위 전화기 본체에 수화기가 그대로 놓여 있음에도 김선영이 또 다른 수화기를 손에 들고 있어, 수화기가 중복으로 생성되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "스케치의 시선 화살표와 프롬프트 지시(민정을 향해 힐끗 시선을 둔 채)와 달리, 김선영의 시선이 배경에 누워있는 민정을 향하지 않고 우측 하단 바닥 쪽을 향하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전경 김선영의 얼굴이 캐릭터 레퍼런스와 일치하지 않는다",
     "severity": "major"
    },
    {
     "issue_ko": "전경 김선영이 레퍼런스 의상(보라 패딩·비니)이 아닌 남색 스웨트셔츠를 입고 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "스케치의 시선 화살표와 프롬프트 지시(민정을 향해 힐끗 시선을 둔 채)와 달리, 김선영의 시선이 배경에 누워있는 민정을 향하지 않고 우측 하단 바닥 쪽을 향하고 있습니다.",
    "fix_en": "Shift the foreground girl's pupils backward to direct her gaze squarely at the sleeping girl in the background. Preserve her head angle, pose, clothing, the telephone, the sleeping girl, and the background exactly.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "전경 김선영의 얼굴이 캐릭터 레퍼런스와 일치하지 않는다",
    "fix_en": "Redraw the foreground girl's face to match the character reference. Preserve her head angle, expression, clothing, the telephone, the sleeping girl, and the background exactly.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "전경 김선영이 레퍼런스 의상(보라 패딩·비니)이 아닌 남색 스웨트셔츠를 입고 있다",
    "fix_en": "Replace the foreground girl's navy top with the purple puffer jacket and grey beanie from the character reference. Preserve her face, hands, the telephone, the sleeping girl, and the background exactly.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the extra telephone receiver resting on the red phone base in the bottom right corner, replacing it with the empty cradle structure of the base. Preserve the foreground girl, her current pose and clothing, the receiver in her hand, the coiled cord, the sleeping girl, the lighting, and the background room exactly.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "지정된 의상(비니, 패딩)이 누락되었으나, 레이아웃 스케치의 클로즈업 프레이밍과 카메라 각도를 완벽하게 구현하여 우선순위에서 승리함."
     },
     {
      "label": "B",
      "score": 500,
      "verdict_ko": "캐릭터의 의상은 정확히 반영했으나 프레이밍이 지시보다 훨씬 넓으며, 수화기 복제 및 임의의 가구 생성 등 치명적 위반이 발생함.  ★위반: [gemini-pro] 전화기 본체에 수화기가 올려져 있음에도 손에 다른 수화기를 들고 있는 복제 오류(duplicated object) / [gemini-pro] 배경 레퍼런스에 존재하지 않는 둥근 스툴 의자 임의 생성(invented object) / [openrouter:x-ai/grok-4.6] 원본 배경에 없는 나무 스툴을 발명함"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.25
     },
     "adjusted": {
      "A": 1.5,
      "B": 0.5
     },
     "violations": {
      "B": [
       "[gemini-pro] 전화기 본체에 수화기가 올려져 있음에도 손에 다른 수화기를 들고 있는 복제 오류(duplicated object)",
       "[gemini-pro] 배경 레퍼런스에 존재하지 않는 둥근 스툴 의자 임의 생성(invented object)",
       "[openrouter:x-ai/grok-4.6] 원본 배경에 없는 나무 스툴을 발명함"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "지정된 의상(비니, 패딩)이 누락되었으나, 레이아웃 스케치의 클로즈업 프레이밍과 카메라 각도를 완벽하게 구현하여 우선순위에서 승리함."
     },
     {
      "label": "B",
      "score": 500,
      "verdict_ko": "캐릭터의 의상은 정확히 반영했으나 프레이밍이 지시보다 훨씬 넓으며, 수화기 복제 및 임의의 가구 생성 등 치명적 위반이 발생함.  ★위반: [gemini-pro] 전화기 본체에 수화기가 올려져 있음에도 손에 다른 수화기를 들고 있는 복제 오류(duplicated object) / [gemini-pro] 배경 레퍼런스에 존재하지 않는 둥근 스툴 의자 임의 생성(invented object) / [openrouter:x-ai/grok-4.6] 원본 배경에 없는 나무 스툴을 발명함"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "레퍼런스의 복장을 누락한 아쉬움이 있으나, 최우선 조건인 클로즈업 프레이밍과 스케치의 구도, 고개를 숙이고 민정을 바라보는 연출을 완벽하게 구현하여 승리함."
     },
     {
      "label": "A",
      "score": 375,
      "verdict_ko": "레퍼런스의 복장은 정확히 반영했으나, 요구된 클로즈업 프레이밍과 시선 방향을 완전히 무시했으며 원본에 없는 의자를 임의로 추가하여 치명적인 감점을 받음.  ★위반: [gemini-pro] 원본 배경에 존재하지 않는 둥근 나무 의자를 임의로 생성함 (invented objects). / [gemini-pro] 지시된 클로즈업(close-up) 프레이밍을 무시하고 줌아웃된 미디엄 샷으로 렌더링함. / [openrouter:x-ai/grok-4.6] 원본 배경에 없는 나무 스툴을 발명함 / [openrouter:x-ai/grok-4.6] 김선영이 스케치가 두지 않은 방 한가운데 스툴에 앉아 있음"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.375,
      "B": 1.5
     },
     "adjusted": {
      "A": 0.375,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[gemini-pro] 원본 배경에 존재하지 않는 둥근 나무 의자를 임의로 생성함 (invented objects).",
       "[gemini-pro] 지시된 클로즈업(close-up) 프레이밍을 무시하고 줌아웃된 미디엄 샷으로 렌더링함.",
       "[openrouter:x-ai/grok-4.6] 원본 배경에 없는 나무 스툴을 발명함",
       "[openrouter:x-ai/grok-4.6] 김선영이 스케치가 두지 않은 방 한가운데 스툴에 앉아 있음"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "레퍼런스의 복장을 누락한 아쉬움이 있으나, 최우선 조건인 클로즈업 프레이밍과 스케치의 구도, 고개를 숙이고 민정을 바라보는 연출을 완벽하게 구현하여 승리함."
     },
     {
      "label": "B",
      "score": 375,
      "verdict_ko": "레퍼런스의 복장은 정확히 반영했으나, 요구된 클로즈업 프레이밍과 시선 방향을 완전히 무시했으며 원본에 없는 의자를 임의로 추가하여 치명적인 감점을 받음.  ★위반: [gemini-pro] 원본 배경에 존재하지 않는 둥근 나무 의자를 임의로 생성함 (invented objects). / [gemini-pro] 지시된 클로즈업(close-up) 프레이밍을 무시하고 줌아웃된 미디엄 샷으로 렌더링함. / [openrouter:x-ai/grok-4.6] 원본 배경에 없는 나무 스툴을 발명함 / [openrouter:x-ai/grok-4.6] 김선영이 스케치가 두지 않은 방 한가운데 스툴에 앉아 있음"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 3000,
     "B": 875
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S3sh3__bgfirst_bg.png",
   "bg_asset_id": "010ba4d8-3ea5-4ca4-b5bd-807bf6f444ad",
   "bg_record_key": "S3sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S3sh3::cine": {
  "applied": true,
  "fingerprint": "16a4e15c638383807ac6db872c3b7b38d381e796d62074a6118be6bc3285807d",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S3sh3_sel.png",
  "source_sha256": "c3a01535cab685495574eef68dd91b8b3a719006cc5294faf5cd4752d33ec7f5",
  "file": "S3sh3_cine.png",
  "latency_ms": 11673
 },
 "S3sh4::signage": {
  "fp": "65fd0ff79e06ef61",
  "inscriptions": [
   {
    "surface_native": "모니터 화면 팝업창",
    "text_native": "새로운 메시지가 도착했습니다.",
    "reason_ko": "어두운 방 안에서 모니터 알림 불빛이 인물의 얼굴을 비추는 극적인 상황을 연출하기 위해, 화면 속 팝업창에 실제 읽을 수 있는 한국어 알림 문구가 필요합니다."
   }
  ]
 },
 "S3sh4": {
  "input_fingerprint": "70fcad34cbab9605",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 중앙에 알림 팝업창이 떠 있고, 그 화면 불빛이 김선영의 얼굴을 비추는 모습\n\nLOCATION (lock): At the computer desk inside the dark shared bedroom, lit mainly by the monitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From seated head height just behind one edge of the monitor, begin a slow diagonal push with the visible screen in the center foreground and 김선영’s three-quarter face beyond its side in the midground. The notification remains centered on the monitor while she leans toward the chat window, her eyes fixed on its newly opened content rather than the camera.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: monitor with centered chat notification in the middle-center of the frame, foreground; 김선영 in the middle-right of the frame, midground, looks toward chat notification.\n- KEY BACKGROUND ELEMENTS: computer monitor (on with notification open) — The screen face is visible at an oblique angle, with the chat notification centered and the login/chat interface surrounding it; used as Central foreground information plane and partial edge frame around 김선영's face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark room stays underexposed while the monitor’s explicitly visible screen light shapes 김선영’s face in a restrained neutral-to-cool range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 모니터 화면 팝업창: \"새로운 메시지가 도착했습니다.\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 중앙에 알림 팝업창이 떠 있고, 그 화면 불빛이 김선영의 얼굴을 비추는 모습\n\nLOCATION (lock): At the computer desk inside the dark shared bedroom, lit mainly by the monitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From seated head height just behind one edge of the monitor, begin a slow diagonal push with the visible screen in the center foreground and 김선영’s three-quarter face beyond its side in the midground. The notification remains centered on the monitor while she leans toward the chat window, her eyes fixed on its newly opened content rather than the camera.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: monitor with centered chat notification in the middle-center of the frame, foreground; 김선영 in the middle-right of the frame, midground, looks toward chat notification.\n- KEY BACKGROUND ELEMENTS: computer monitor (on with notification open) — The screen face is visible at an oblique angle, with the chat notification centered and the login/chat interface surrounding it; used as Central foreground information plane and partial edge frame around 김선영's face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark room stays underexposed while the monitor’s explicitly visible screen light shapes 김선영’s face in a restrained neutral-to-cool range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 모니터 화면 팝업창: \"새로운 메시지가 도착했습니다.\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 중앙에 알림 팝업창이 떠 있고, 그 화면 불빛이 김선영의 얼굴을 비추는 모습\n\nLOCATION (lock): At the computer desk inside the dark shared bedroom, lit mainly by the monitor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From seated head height just behind one edge of the monitor, begin a slow diagonal push with the visible screen in the center foreground and 김선영’s three-quarter face beyond its side in the midground. The notification remains centered on the monitor while she leans toward the chat window, her eyes fixed on its newly opened content rather than the camera.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: monitor with centered chat notification in the middle-center of the frame, foreground; 김선영 in the middle-right of the frame, midground, looks toward chat notification.\n- KEY BACKGROUND ELEMENTS: computer monitor (on with notification open) — The screen face is visible at an oblique angle, with the chat notification centered and the login/chat interface surrounding it; used as Central foreground information plane and partial edge frame around 김선영's face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark room stays underexposed while the monitor’s explicitly visible screen light shapes 김선영’s face in a restrained neutral-to-cool range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 모니터 화면 팝업창: \"새로운 메시지가 도착했습니다.\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "김선영의 시선은 모니터의 알림창 텍스트를 향하고 있음.",
    "built_space": "모니터가 책상 위에 배치됨. 카메라가 모니터 쪽으로 이동했음에도 배경의 침대와 수면 중인 인물이 이전 샷과 동일한 각도로 남아있음.",
    "entities": "김선영(보라색 패딩과 비니 착용으로 이전 샷 연속성 위반), 모니터 팝업창 텍스트 '새로운 메시지가 도착했습니다.' 일치, 수면 중인 인물.",
    "hard_violations": [
     "physically impossible staging (카메라 이동에 따른 공간 회전 미반영 및 배경 평면 복사)"
    ],
    "physics": "의자에 앉아 키보드에 손을 올린 상태로 신체가 안정적으로 지지됨."
   },
   {
    "label": "B",
    "direction": "김선영의 시선이 화면 중앙의 팝업창을 정확히 향함.",
    "built_space": "카메라가 모니터 뒤에서 인물을 비추는 구도이며, 카메라 시점에 맞게 배경 공간이 자연스럽게 변경됨.",
    "entities": "김선영(이전 샷과 동일한 네이비 상의와 헤어스타일 유지), CRT 모니터 화면의 팝업창 텍스트 '새로운 메시지가 도착했습니다.' 일치.",
    "hard_violations": [],
    "physics": "책상 앞에 정상적으로 앉아 있으며 신체와 팔이 자연스럽게 지지됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "이전 샷의 의상(네이비 티셔츠)과 연속성을 완벽히 유지하며, 지시된 모니터 뷰 카메라 앵글과 팝업창 텍스트를 정확하게 구현함."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "이전 샷의 의상 연속성을 위반하고 캐릭터 레퍼런스의 패딩을 착용했으며, 카메라 시점 이동에도 배경 구도가 그대로 복사되어 물리적 오류가 발생함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S3sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870689>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 장면(Previous Shot Still)에서 배경 침대에 누워있던 잠든 인물이 우측 배경의 침대에서 사라졌습니다.",
     "fix_en": "Restore the sleeping girl in the right background bed, matching her exact pose and clothing from the reference. Preserve Kim Sun-young, the monitor, and the room lighting.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "모니터 화면이 채팅·로그인 창이 아니라 문서 프로그램처럼 보인다",
     "fix_en": "Replace the word processor interface behind the central popup with a chat window. Preserve the central popup text, Kim Sun-young, and lighting.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "모니터 제목줄·툴바·하단에 지정 문구 외의 깨진 영문과 다른 글자가 읽힌다",
     "fix_en": "Remove legible gibberish text on the monitor toolbars and replace with indistinct UI shapes. Preserve the central popup text, Kim Sun-young, and lighting.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "김선영 상의가 이전 컷의 무지 남색이 아니라 앞에 무늬가 있는 남색 스웨트셔츠다",
     "fix_en": "Remove the printed pattern on Kim Sun-young's sweatshirt, making it solid navy. Preserve her face, pose, monitor, and background.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이전 장면(Previous Shot Still)에서 배경 침대에 누워있던 잠든 인물이 우측 배경의 침대에서 사라졌습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "모니터 화면의 팝업창 주변 배경이 지시된 채팅/로그인 인터페이스가 아닌 워드 프로세서(문서 편집) 프로그램으로 렌더링되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "모니터 화면이 채팅·로그인 창이 아니라 문서 프로그램처럼 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "모니터 제목줄·툴바·하단에 지정 문구 외의 깨진 영문과 다른 글자가 읽힌다",
     "severity": "major"
    },
    {
     "issue_ko": "김선영 상의가 이전 컷의 무지 남색이 아니라 앞에 무늬가 있는 남색 스웨트셔츠다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 장면(Previous Shot Still)에서 배경 침대에 누워있던 잠든 인물이 우측 배경의 침대에서 사라졌습니다.",
    "fix_en": "Restore the sleeping girl in the right background bed, matching her exact pose and clothing from the reference. Preserve Kim Sun-young, the monitor, and the room lighting.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "모니터 화면이 채팅·로그인 창이 아니라 문서 프로그램처럼 보인다",
    "fix_en": "Replace the word processor interface behind the central popup with a chat window. Preserve the central popup text, Kim Sun-young, and lighting.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "모니터 제목줄·툴바·하단에 지정 문구 외의 깨진 영문과 다른 글자가 읽힌다",
    "fix_en": "Remove legible gibberish text on the monitor toolbars and replace with indistinct UI shapes. Preserve the central popup text, Kim Sun-young, and lighting.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "김선영 상의가 이전 컷의 무지 남색이 아니라 앞에 무늬가 있는 남색 스웨트셔츠다",
    "fix_en": "Remove the printed pattern on Kim Sun-young's sweatshirt, making it solid navy. Preserve her face, pose, monitor, and background.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S3sh3"
  }
 },
 "S3sh4::cine": {
  "applied": true,
  "fingerprint": "0a9321867f994706b131e282e41c50d4c6c8dbfeb75a42db6ea0e66adc437675",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S3sh4_sel.png",
  "source_sha256": "7c1653f90b3217dd51f48d6c039960fe68be87966d709da6235d8adff5f6b1ae",
  "file": "S3sh4_cine.png",
  "latency_ms": 11243
 },
 "S3sh7::signage": {
  "fp": "41cbe497c6146f13",
  "inscriptions": []
 },
 "S3sh7": {
  "input_fingerprint": "05dc9c4e5f2ff599",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 점퍼를 입고 살금살금 발뒤꿈치를 든 채 방문 밖으로 한 발을 내디딘 상태의 김선영의 실루엣\n\nLOCATION (lock): At the bedroom doorway inside the home, with the dark room behind her and the interior passage beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Low near the floor inside the room, hold the diagonal view across the open doorway as 김선영’s jacketed figure crosses the threshold in restrained silhouette. Her lifted rear heel anchors the lower frame while one foot reaches outside and her torso angles toward the space beyond the room.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 김선영 in the middle-center of the frame, midground, moves toward space outside the bedroom; open bedroom doorway in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bedroom doorway (open) — Seen obliquely from inside the room, with the opening leading away from camera; used as Threshold frame around 김선영's raised heel and first step out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark room and threshold contrast form a subdued silhouette without specifying an unmentioned light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong is wearing the purple padded jacket as she leaves; it must remain on her through her subsequent nighttime walk and meeting with Ji Gukhyeon.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 점퍼를 입고 살금살금 발뒤꿈치를 든 채 방문 밖으로 한 발을 내디딘 상태의 김선영의 실루엣\n\nLOCATION (lock): At the bedroom doorway inside the home, with the dark room behind her and the interior passage beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Low near the floor inside the room, hold the diagonal view across the open doorway as 김선영’s jacketed figure crosses the threshold in restrained silhouette. Her lifted rear heel anchors the lower frame while one foot reaches outside and her torso angles toward the space beyond the room.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 김선영 in the middle-center of the frame, midground, moves toward space outside the bedroom; open bedroom doorway in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bedroom doorway (open) — Seen obliquely from inside the room, with the opening leading away from camera; used as Threshold frame around 김선영's raised heel and first step out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark room and threshold contrast form a subdued silhouette without specifying an unmentioned light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong is wearing the purple padded jacket as she leaves; it must remain on her through her subsequent nighttime walk and meeting with Ji Gukhyeon.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 점퍼를 입고 살금살금 발뒤꿈치를 든 채 방문 밖으로 한 발을 내디딘 상태의 김선영의 실루엣\n\nLOCATION (lock): At the bedroom doorway inside the home, with the dark room behind her and the interior passage beyond. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Low near the floor inside the room, hold the diagonal view across the open doorway as 김선영’s jacketed figure crosses the threshold in restrained silhouette. Her lifted rear heel anchors the lower frame while one foot reaches outside and her torso angles toward the space beyond the room.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 김선영 in the middle-center of the frame, midground, moves toward space outside the bedroom; open bedroom doorway in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: bedroom doorway (open) — Seen obliquely from inside the room, with the opening leading away from camera; used as Threshold frame around 김선영's raised heel and first step out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark room and threshold contrast form a subdued silhouette without specifying an unmentioned light source or color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong is wearing the purple padded jacket as she leaves; it must remain on her through her subsequent nighttime walk and meeting with Ji Gukhyeon.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "인물의 몸통과 시선이 열린 방문 너머 복도를 향하고 있음.",
    "built_space": "왼쪽에 침대가 있는 방 안쪽 바닥 근처에서 방문을 향해 대각선으로 바라보는 로우 앵글 구도.",
    "entities": "보라색 패딩, 청바지, 비니, 흰색 스니커즈를 착용한 김선영의 뒷모습.",
    "hard_violations": [],
    "physics": "방 안쪽에 있는 뒷발(왼발)은 뒤꿈치를 들고 있고, 앞발(오른발)은 복도 바닥을 안정적으로 디디고 있음."
   },
   {
    "label": "B",
    "direction": "인물의 시선과 몸통이 방 밖이 아닌 방 안쪽(카메라 방향)을 향하고 있음.",
    "built_space": "침대가 있는 방 안에서 열린 문을 바라보는 구도이나, 문의 열림 방향과 인물의 위치가 밖에서 안으로 향함.",
    "entities": "보라색 패딩, 청바지, 비니를 착용한 김선영.",
    "hard_violations": [
     "프롬프트는 방 밖으로 한 발을 내딛고 나가는 상황을 요구했으나, 인물이 복도에서 방 안으로 들어오고 있음."
    ],
    "physics": "걷는 자세로 앞발(오른발)이 방 안쪽 바닥을, 뒷발(왼발)이 복도를 딛고 지탱 중임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "뒷발의 뒤꿈치를 들고 방 밖으로 나가는 동선을 로우 앵글과 어두운 조명 아래 정확히 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "인물이 방 밖으로 나가지 않고 안으로 들어오고 있어 제시된 샷의 동선을 정면으로 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 김선영 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S3sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 김선영: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:870689>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 인물을 '실루엣(silhouette)'으로 표현할 것을 지시했으나, 이미지에서는 조명을 받아 보라색 점퍼와 청바지의 색상 및 형태가 선명하게 드러납니다.",
     "fix_en": "Darken the character to cast her in a subdued silhouette, removing the bright lighting on her jacket and jeans. Maintain her pose, the room geometry, the lighting, and the framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 샷(PREVIOUS SHOT)의 인물 외형(옷, 머리 등)을 그대로 유지하라는 지시와 달리, 이전 샷에 없던 비니를 머리에 착용하고 있습니다.",
     "fix_en": "Remove the black beanie, replacing it with the character's long black hair. Maintain her face, pose, purple jacket, the room environment, the light, and the framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "인물이 나가는 방문 바닥에 미닫이문 레일이 그려져 있으나, 이전 샷에 설정된 방문은 둥근 손잡이가 있는 여닫이문입니다.",
     "fix_en": "Replace the sliding door tracks on the floor with a flat wooden threshold. Maintain the character's shoes, pose, the surrounding walls, the light, and the framing.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "카메라가 방 안 바닥 근처 로우가 아니고 중간 높이에서 전신을 거의 수평으로 담고 있다",
     "fix_en": "Crop the bottom of the image to anchor the heel closer to the frame edge, mitigating the high camera angle. Maintain the character's pose, the room layout, the light, and the set.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "왼쪽 침대와 중앙 방문 사이에 이전 장면 그 방에 없던, 창문이 보이는 출입구가 있다",
     "fix_en": "Replace the extra dark doorway on the left wall with a solid interior wall matching the room. Maintain the bed, the character's pose, the main doorway, the light, and the framing.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 인물을 '실루엣(silhouette)'으로 표현할 것을 지시했으나, 이미지에서는 조명을 받아 보라색 점퍼와 청바지의 색상 및 형태가 선명하게 드러납니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷(PREVIOUS SHOT)의 인물 외형(옷, 머리 등)을 그대로 유지하라는 지시와 달리, 이전 샷에 없던 비니를 머리에 착용하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "인물이 나가는 방문 바닥에 미닫이문 레일이 그려져 있으나, 이전 샷에 설정된 방문은 둥근 손잡이가 있는 여닫이문입니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "문턱을 넘는 김선영이 억제된 실루엣이 아니라 화면 중앙에서 보랏빛 점퍼와 몸이 선명히 밝혀져 있다",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 방 안 바닥 근처 로우가 아니고 중간 높이에서 전신을 거의 수평으로 담고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "들린 뒷꿈치가 프레임 하단을 고정하지 않고 발 아래로 바닥과 러그가 넓게 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 침대와 중앙 방문 사이에 이전 장면 그 방에 없던, 창문이 보이는 출입구가 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 인물을 '실루엣(silhouette)'으로 표현할 것을 지시했으나, 이미지에서는 조명을 받아 보라색 점퍼와 청바지의 색상 및 형태가 선명하게 드러납니다.",
    "fix_en": "Darken the character to cast her in a subdued silhouette, removing the bright lighting on her jacket and jeans. Maintain her pose, the room geometry, the lighting, and the framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "이전 샷(PREVIOUS SHOT)의 인물 외형(옷, 머리 등)을 그대로 유지하라는 지시와 달리, 이전 샷에 없던 비니를 머리에 착용하고 있습니다.",
    "fix_en": "Remove the black beanie, replacing it with the character's long black hair. Maintain her face, pose, purple jacket, the room environment, the light, and the framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "인물이 나가는 방문 바닥에 미닫이문 레일이 그려져 있으나, 이전 샷에 설정된 방문은 둥근 손잡이가 있는 여닫이문입니다.",
    "fix_en": "Replace the sliding door tracks on the floor with a flat wooden threshold. Maintain the character's shoes, pose, the surrounding walls, the light, and the framing.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "카메라가 방 안 바닥 근처 로우가 아니고 중간 높이에서 전신을 거의 수평으로 담고 있다",
    "fix_en": "Crop the bottom of the image to anchor the heel closer to the frame edge, mitigating the high camera angle. Maintain the character's pose, the room layout, the light, and the set.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "왼쪽 침대와 중앙 방문 사이에 이전 장면 그 방에 없던, 창문이 보이는 출입구가 있다",
    "fix_en": "Replace the extra dark doorway on the left wall with a solid interior wall matching the room. Maintain the bed, the character's pose, the main doorway, the light, and the framing.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S3sh4"
  }
 },
 "S3sh7::cine": {
  "applied": true,
  "fingerprint": "9e604292d7837ad1100ea861d40d25fe61899c87ff66f55b6dc97f606eb34b81",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S3sh7_sel.png",
  "source_sha256": "09b2fda85afc6a269b9307deff9b67b667a4bb35733bab06f5c3463cccdf48f1",
  "file": "S3sh7_cine.png",
  "latency_ms": 13526
 },
 "S4sh2::signage": {
  "fp": "df227b399d791afe",
  "inscriptions": []
 },
 "era_assess::b1e0cee3423291cd": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 아파트의 주방 및 방문 구조",
    "search_terms_native": [
     "2000년대 아파트 인테리어 체리몰딩",
     "2010년대 아파트 주방 싱크대",
     "한국 아파트 거실 주방 구조",
     "아파트 방문 문틀"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "2000년대에서 2010년대 사이 한국 아파트 특유의 체리색 몰딩, 주방 구조(싱크대 및 아일랜드 배치), 문틀 및 바닥재(장판/강마루) 형태는 일반적인 서구식 주택 이미지와 확연히 다릅니다."
   }
  ]
 },
 "era_ref::e5dcb78d5f8cebc7": {
  "subject": "2000년대~2010년대 한국 아파트의 주방 및 방문 구조",
  "terms": [
   "2000년대 아파트 인테리어 체리몰딩",
   "2010년대 아파트 주방 싱크대",
   "한국 아파트 거실 주방 구조",
   "아파트 방문 문틀"
  ],
  "queries": [
   [
    "2000년대 한국 아파트 인테리어 체리몰딩 거실 주방 방문 문틀",
    "2010년대 한국 아파트 주방 싱크대 거실 주방 구조"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://contents-cdn.viewus.co.kr/image/2023/09/CP-2023-0023/image-93fd9ce3-7bbc-44b2-9a08-07f60f35537b.jpeg",
  "picked_reason_ko": "2000~2010년대 한국 아파트에서 흔한 목재색 ㄱ자형 붙박이 주방과 주방을 구획하는 미닫이문 구조가 한 화면에 선명하게 보여 가장 좋은 참고 사진이다.",
  "sha256": "a5ebc143cc31d59123f6a6791821e6acf95f15e7c5a291e3b616a1b083ce00c3",
  "file": "eraref_e5dcb78d5f8cebc7.png"
 },
 "S4sh2::bgfirst_bg": {
  "input_fingerprint": "c086881f757cec65",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 방문 앞에 서서 부엌 쪽을 향해 목을 빼고 입을 크게 벌린 민정\n\nLOCATION (lock): At the bedroom doorway opening onto the home's sunlit main floor, facing toward the kitchen.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height outside the open bedroom, dolly toward 민정 from her rear three-quarter side and frame slightly upward, leaving meaningful negative space toward the kitchen. She cranes her neck into that space with her mouth open mid-call, attention directed toward her mother in the kitchen rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 민정 (2001년) in the middle-left of the frame, midground, looks toward kitchen; kitchen-side negative space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: open bedroom doorway (open) — The doorway is seen from outside at an oblique angle, with 민정 leaning across its edge; used as Frames 민정's body and establishes the room she has just emerged from; kitchen area (located beyond the doorway sightline); used as Destination of 민정's call and the open negative space ahead of her gaze.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning sunlight described across the floor is kept soft and realistic within a restrained neutral palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 아파트의 주방 및 방문 구조: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 방문 앞에 서서 부엌 쪽을 향해 목을 빼고 입을 크게 벌린 민정\n\nLOCATION (lock): At the bedroom doorway opening onto the home's sunlit main floor, facing toward the kitchen.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height outside the open bedroom, dolly toward 민정 from her rear three-quarter side and frame slightly upward, leaving meaningful negative space toward the kitchen. She cranes her neck into that space with her mouth open mid-call, attention directed toward her mother in the kitchen rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 민정 (2001년) in the middle-left of the frame, midground, looks toward kitchen; kitchen-side negative space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: open bedroom doorway (open) — The doorway is seen from outside at an oblique angle, with 민정 leaning across its edge; used as Frames 민정's body and establishes the room she has just emerged from; kitchen area (located beyond the doorway sightline); used as Destination of 민정's call and the open negative space ahead of her gaze.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning sunlight described across the floor is kept soft and realistic within a restrained neutral palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 아파트의 주방 및 방문 구조: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S4sh2__bgfirst_bg.png",
  "asset_id": "e6923d8f-2834-4f1d-b0c4-e72763e2ffb9",
  "input_asset_ids": [
   "0f2a45e8-6c92-404e-b194-f82da0ab1c65",
   "a33d46e5-2579-41fe-8428-3560a4675072"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국 아파트의 주방 및 방문 구조",
   "queries": [
    [
     "2000년대 한국 아파트 인테리어 체리몰딩 거실 주방 방문 문틀",
     "2010년대 한국 아파트 주방 싱크대 거실 주방 구조"
    ]
   ],
   "picked_url": "https://contents-cdn.viewus.co.kr/image/2023/09/CP-2023-0023/image-93fd9ce3-7bbc-44b2-9a08-07f60f35537b.jpeg",
   "sha256": "a5ebc143cc31d59123f6a6791821e6acf95f15e7c5a291e3b616a1b083ce00c3",
   "file": "eraref_e5dcb78d5f8cebc7.png"
  }
 },
 "S4sh2": {
  "input_fingerprint": "40fa49484cd49f98",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 열린 방문 앞에 서서 부엌 쪽을 향해 목을 빼고 입을 크게 벌린 민정\n\nLOCATION (lock): At the bedroom doorway opening onto the home's sunlit main floor, facing toward the kitchen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height outside the open bedroom, dolly toward 민정 from her rear three-quarter side and frame slightly upward, leaving meaningful negative space toward the kitchen. She cranes her neck into that space with her mouth open mid-call, attention directed toward her mother in the kitchen rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 민정 (2001년) in the middle-left of the frame, midground, looks toward kitchen; kitchen-side negative space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: open bedroom doorway (open) — The doorway is seen from outside at an oblique angle, with 민정 leaning across its edge; used as Frames 민정's body and establishes the room she has just emerged from; kitchen area (located beyond the doorway sightline); used as Destination of 민정's call and the open negative space ahead of her gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning sunlight described across the floor is kept soft and realistic within a restrained neutral palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (2001년) (Korean 소녀, 아동·청소년기, 둥근 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 열린 방문 앞에 서서 부엌 쪽을 향해 목을 빼고 입을 크게 벌린 민정\n\nLOCATION (lock): At the bedroom doorway opening onto the home's sunlit main floor, facing toward the kitchen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height outside the open bedroom, dolly toward 민정 from her rear three-quarter side and frame slightly upward, leaving meaningful negative space toward the kitchen. She cranes her neck into that space with her mouth open mid-call, attention directed toward her mother in the kitchen rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 민정 (2001년) in the middle-left of the frame, midground, looks toward kitchen; kitchen-side negative space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: open bedroom doorway (open) — The doorway is seen from outside at an oblique angle, with 민정 leaning across its edge; used as Frames 민정's body and establishes the room she has just emerged from; kitchen area (located beyond the doorway sightline); used as Destination of 민정's call and the open negative space ahead of her gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning sunlight described across the floor is kept soft and realistic within a restrained neutral palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (2001년) (Korean 소녀, 아동·청소년기, 둥근 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 열린 방문 앞에 서서 부엌 쪽을 향해 목을 빼고 입을 크게 벌린 민정\n\nLOCATION (lock): At the bedroom doorway opening onto the home's sunlit main floor, facing toward the kitchen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height outside the open bedroom, dolly toward 민정 from her rear three-quarter side and frame slightly upward, leaving meaningful negative space toward the kitchen. She cranes her neck into that space with her mouth open mid-call, attention directed toward her mother in the kitchen rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 민정 (2001년) in the middle-left of the frame, midground, looks toward kitchen; kitchen-side negative space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: open bedroom doorway (open) — The doorway is seen from outside at an oblique angle, with 민정 leaning across its edge; used as Frames 민정's body and establishes the room she has just emerged from; kitchen area (located beyond the doorway sightline); used as Destination of 민정's call and the open negative space ahead of her gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning sunlight described across the floor is kept soft and realistic within a restrained neutral palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (2001년) (Korean 소녀, 아동·청소년기, 둥근 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S4sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S4sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 민정 (2001년): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1336104>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L04B03.png"
    },
    {
     "label": "CHARACTER REFERENCE — 민정 (2001년): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1336104>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 750,
      "verdict_ko": "구도와 배경은 우수하나 샷 텍스트에 없는 인물을 추가하여 엄격한 등장인물 제한 규정을 위반함.  ★위반: [gemini-pro] invented people (샷 텍스트에 없는 인물 추가) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 보이지 말라고 한 인물(부엌의 어머니)을 추가함"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "인물 조건은 지켰으나 방문의 얇은 측면 두께에 손잡이가 달린 물리적으로 불가능한 형태의 치명적 오류가 있음.  ★위반: [gemini-pro] physically impossible objects (방문 측면 두께에 문손잡이 부착)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.25,
      "B": 1.75
     },
     "adjusted": {
      "A": 0.75,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[gemini-pro] invented people (샷 텍스트에 없는 인물 추가)",
       "[openrouter:x-ai/grok-4.6] 샷 텍스트가 보이지 말라고 한 인물(부엌의 어머니)을 추가함"
      ],
      "B": [
       "[gemini-pro] physically impossible objects (방문 측면 두께에 문손잡이 부착)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.75,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 750,
      "verdict_ko": "구도와 배경은 우수하나 샷 텍스트에 없는 인물을 추가하여 엄격한 등장인물 제한 규정을 위반함.  ★위반: [gemini-pro] invented people (샷 텍스트에 없는 인물 추가) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 보이지 말라고 한 인물(부엌의 어머니)을 추가함"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "인물 조건은 지켰으나 방문의 얇은 측면 두께에 손잡이가 달린 물리적으로 불가능한 형태의 치명적 오류가 있음.  ★위반: [gemini-pro] physically impossible objects (방문 측면 두께에 문손잡이 부착)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 앵글과 문틀에 기대는 연출, 네거티브 스페이스 활용을 정확히 구현했으며 인물 조건도 엄격히 지켰습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "샷 텍스트에 없는 인물을 배경에 임의로 추가하여 치명적인 규정 위반이 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선과 크게 벌린 입이 부엌 쪽 빈 공간을 정확히 향함.",
      "built_space": "열린 방문, 침실 내부, 부엌의 구조가 참조 사진과 일치하며 문틀에 기대는 위치가 정확함.",
      "entities": "민정의 외모와 의상이 참조와 일치하며 화면에 다른 인물은 없음.",
      "hard_violations": [],
      "physics": "상체를 앞으로 내민 자세가 보이지 않는 하체에 의해 안정적으로 지지됨."
     },
     {
      "label": "B",
      "direction": "시선과 벌린 입이 부엌 쪽을 향함.",
      "built_space": "참조 사진의 공간 구조가 전반적으로 반영됨.",
      "entities": "민정의 외양은 일치하나 부엌에 지시문에 없는 다른 여성이 등장함.",
      "hard_violations": [
       "지시문(Shot Text)에 없는 인물(부엌의 여성) 추가"
      ],
      "physics": "상체를 내민 자세가 바닥에 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 앵글과 문틀에 기대는 연출, 네거티브 스페이스 활용을 정확히 구현했으며 인물 조건도 엄격히 지켰습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "샷 텍스트에 없는 인물을 배경에 임의로 추가하여 치명적인 규정 위반이 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선과 크게 벌린 입이 부엌 쪽 빈 공간을 정확히 향함.",
      "built_space": "열린 방문, 침실 내부, 부엌의 구조가 참조 사진과 일치하며 문틀에 기대는 위치가 정확함.",
      "entities": "민정의 외모와 의상이 참조와 일치하며 화면에 다른 인물은 없음.",
      "hard_violations": [],
      "physics": "상체를 앞으로 내민 자세가 보이지 않는 하체에 의해 안정적으로 지지됨."
     },
     {
      "label": "A",
      "direction": "시선과 벌린 입이 부엌 쪽을 향함.",
      "built_space": "참조 사진의 공간 구조가 전반적으로 반영됨.",
      "entities": "민정의 외양은 일치하나 부엌에 지시문에 없는 다른 여성이 등장함.",
      "hard_violations": [
       "지시문(Shot Text)에 없는 인물(부엌의 여성) 추가"
      ],
      "physics": "상체를 내민 자세가 바닥에 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 753,
     "B": 1507
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 753,
   "B": 1507
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "구도와 배경은 우수하나 샷 텍스트에 없는 인물을 추가하여 엄격한 등장인물 제한 규정을 위반함.  ★위반: [gemini-pro] invented people (샷 텍스트에 없는 인물 추가) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 보이지 말라고 한 인물(부엌의 어머니)을 추가함"
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "인물 조건은 지켰으나 방문의 얇은 측면 두께에 손잡이가 달린 물리적으로 불가능한 형태의 치명적 오류가 있음.  ★위반: [gemini-pro] physically impossible objects (방문 측면 두께에 문손잡이 부착)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L04B03.png"
   },
   {
    "label": "CHARACTER REFERENCE — 민정 (2001년): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1336104>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "방문 손잡이가 문짝의 넓은 면이 아닌 얇은 모서리에 위치하여 물리적으로 불가능한 구조입니다.",
     "fix_en": "Erase the doorknob protruding from the thin edge of the door and draw a standard metallic doorknob properly mounted on the flat, wide surface of the door panel. Preserve the girl, her pose and expression, the door's shape, the kitchen background, and the lighting exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "프롬프트의 지시와 달리 인물의 후면 3/4 측면이 아닌 전면 3/4 측면을 보여주고 있습니다.",
     "fix_en": "Turn the girl's head slightly more towards the kitchen to emphasize her gaze away from the camera, keeping her body pose, the door, the room geometry, and the camera position strictly unchanged.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "방문 손잡이가 문짝의 넓은 면을 관통하지 않고 얇은 모서리 끝에 위치하며, 양쪽 손잡이가 문 표면과 평행하게 튀어나와 있는 물리적으로 불가능한 구조입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "카메라가 인물의 후면 3/4 측면(rear three-quarter side)에서 비추어야 한다는 프롬프트 지시와 달리, 인물의 전면 3/4 측면을 보여주고 있습니다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트의 지시와 달리 인물의 후면 3/4 측면이 아닌 전면 3/4 측면을 보여주고 있습니다.",
    "fix_en": "Turn the girl's head slightly more towards the kitchen to emphasize her gaze away from the camera, keeping her body pose, the door, the room geometry, and the camera position strictly unchanged.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Erase the doorknob protruding from the thin edge of the door and draw a standard metallic doorknob properly mounted on the flat, wide surface of the door panel. Preserve the girl, her pose and expression, the door's shape, the kitchen background, and the lighting exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1875,
      "verdict_ko": "프롬프트가 지시한 인물의 행동, 위치, 시선을 훌륭하게 구현하였으며, A안에서 발생한 방문 손잡이의 물리적 오류를 성공적으로 수정하여 가장 완성도 높은 샷을 제공합니다."
     },
     {
      "label": "A",
      "score": 1417,
      "verdict_ko": "인물의 정체성과 전반적인 배경은 잘 구현되었으나, 방문 손잡이 옆에 물리적으로 불가능한 형태의 구조물이 돌출되는 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 방문 손잡이 옆 문 표면에 물리적으로 불가능한 금속 실린더 구조물이 돌출되어 허공에 떠 있음 (invented objects/physically impossible staging)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.875
     },
     "adjusted": {
      "A": 1.417,
      "B": 1.875
     },
     "violations": {
      "A": [
       "[gemini-pro] 방문 손잡이 옆 문 표면에 물리적으로 불가능한 금속 실린더 구조물이 돌출되어 허공에 떠 있음 (invented objects/physically impossible staging)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1875,
      "verdict_ko": "프롬프트가 지시한 인물의 행동, 위치, 시선을 훌륭하게 구현하였으며, A안에서 발생한 방문 손잡이의 물리적 오류를 성공적으로 수정하여 가장 완성도 높은 샷을 제공합니다."
     },
     {
      "label": "A",
      "score": 1417,
      "verdict_ko": "인물의 정체성과 전반적인 배경은 잘 구현되었으나, 방문 손잡이 옆에 물리적으로 불가능한 형태의 구조물이 돌출되는 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 방문 손잡이 옆 문 표면에 물리적으로 불가능한 금속 실린더 구조물이 돌출되어 허공에 떠 있음 (invented objects/physically impossible staging)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "열린 문의 얇은 모서리 면에 열쇠 구멍이 묘사되는 물리적 공간 오류가 발생했으며, 입을 벌린 모습에서 치아가 뭉개져 보입니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "문의 물리적 구조(래치와 손잡이)를 현실적으로 올바르게 수정하였고, 인물의 치아와 입 안 디테일이 자연스럽게 향상되어 더 완성도가 높습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "민정의 시선과 얼굴 방향이 배경의 부엌을 정확히 향하고 있음.",
      "built_space": "열린 침실 문과 배경의 부엌이 레퍼런스와 동일하게 배치되었으나, 열린 문의 얇은 두께 면(래치가 있는 곳)에 열쇠 구멍이 그려져 있어 물리적으로 불가능한 구조를 띠고 있음.",
      "entities": "민정의 얼굴, 머리 스타일, 체형 등 신원은 레퍼런스와 잘 일치하나 치아가 디테일 없이 뭉개져 있음. 배경의 부엌 가구와 창문 등은 지정된 모습과 일치함.",
      "hard_violations": [
       "열린 문의 얇은 두께 면(모서리)에 열쇠 구멍이 묘사됨 (물리적으로 불가능한 건축 구조)"
      ],
      "physics": "문 밖으로 몸을 내밀고 있는 자세가 바닥에 의해 안정적으로 지탱되고 있으며 옷깃의 형태도 자연스러움."
     },
     {
      "label": "B",
      "direction": "민정의 시선과 고개가 배경에 있는 부엌 쪽을 올바르게 향하고 있음.",
      "built_space": "침실 문의 열린 각도와 부엌의 배치가 레퍼런스와 일치하며, 문의 모서리에 있던 비현실적인 열쇠 구멍이 나사 형태의 래치 플레이트로 올바르게 수정됨.",
      "entities": "민정의 신원이 레퍼런스와 훌륭하게 일치하며, 입을 벌린 모습에서 치아와 입술의 묘사가 A보다 훨씬 사실적임. 부엌의 요소도 정확하게 재현됨.",
      "hard_violations": [],
      "physics": "발과 지면은 보이지 않으나 서서 문 밖을 내다보는 신체의 무게 중심과 기대는 자세가 자연스럽게 지탱되고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "열린 문의 얇은 모서리 면에 열쇠 구멍이 묘사되는 물리적 공간 오류가 발생했으며, 입을 벌린 모습에서 치아가 뭉개져 보입니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "문의 물리적 구조(래치와 손잡이)를 현실적으로 올바르게 수정하였고, 인물의 치아와 입 안 디테일이 자연스럽게 향상되어 더 완성도가 높습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "민정의 시선과 얼굴 방향이 배경의 부엌을 정확히 향하고 있음.",
      "built_space": "열린 침실 문과 배경의 부엌이 레퍼런스와 동일하게 배치되었으나, 열린 문의 얇은 두께 면(래치가 있는 곳)에 열쇠 구멍이 그려져 있어 물리적으로 불가능한 구조를 띠고 있음.",
      "entities": "민정의 얼굴, 머리 스타일, 체형 등 신원은 레퍼런스와 잘 일치하나 치아가 디테일 없이 뭉개져 있음. 배경의 부엌 가구와 창문 등은 지정된 모습과 일치함.",
      "hard_violations": [
       "열린 문의 얇은 두께 면(모서리)에 열쇠 구멍이 묘사됨 (물리적으로 불가능한 건축 구조)"
      ],
      "physics": "문 밖으로 몸을 내밀고 있는 자세가 바닥에 의해 안정적으로 지탱되고 있으며 옷깃의 형태도 자연스러움."
     },
     {
      "label": "A",
      "direction": "민정의 시선과 고개가 배경에 있는 부엌 쪽을 올바르게 향하고 있음.",
      "built_space": "침실 문의 열린 각도와 부엌의 배치가 레퍼런스와 일치하며, 문의 모서리에 있던 비현실적인 열쇠 구멍이 나사 형태의 래치 플레이트로 올바르게 수정됨.",
      "entities": "민정의 신원이 레퍼런스와 훌륭하게 일치하며, 입을 벌린 모습에서 치아와 입술의 묘사가 A보다 훨씬 사실적임. 부엌의 요소도 정확하게 재현됨.",
      "hard_violations": [],
      "physics": "발과 지면은 보이지 않으나 서서 문 밖을 내다보는 신체의 무게 중심과 기대는 자세가 자연스럽게 지탱되고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1425,
     "B": 1880
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S4sh2__bgfirst_bg.png",
   "bg_asset_id": "e6923d8f-2834-4f1d-b0c4-e72763e2ffb9",
   "bg_record_key": "S4sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S4sh2::cine": {
  "applied": true,
  "fingerprint": "5b8eb716b745ab9c8eaad4c667e5c6a549a575b8bbd186c76841e4ed1328969d",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S4sh2_sel.png",
  "source_sha256": "15f336aa28727625a6cea0aefc22dce301f03f150b423f54e626ac8486b0efb8",
  "file": "S4sh2_cine.png",
  "latency_ms": 10622
 },
 "S5sh2::signage": {
  "fp": "2d4ffab247b1ea11",
  "inscriptions": []
 },
 "S5sh2": {
  "input_fingerprint": "7d96d0e1566252c0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강가 수면 위로 엎드린 채 나신으로 떠 있는 김선영의 창백한 몸통\n\nLOCATION (lock): On the open surface of the river close to its bank, in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From above and slightly to one side, the crane settles into an oblique high-angle medium view of 김선영 floating face-down, keeping the composition clinical and non-graphic. Her pale nude torso is fully legible against the surrounding river surface while her face remains turned away and the frame leaves room along her body for the next move toward the ankles.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: river surface (flowing with small waves); used as Environmental field surrounding and carrying the body, with small waves providing restrained motion.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime light renders the pale body and dark river in a subdued neutral-to-cool palette with moderate contrast and no sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's motionless body remains nude and floats horizontally face-down on the river surface beside the bank, with her head and torso directed downward and her limbs extended in the prone orientation. A flesh-colored stocking hangs from the end of one ankle.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강가 수면 위로 엎드린 채 나신으로 떠 있는 김선영의 창백한 몸통\n\nLOCATION (lock): On the open surface of the river close to its bank, in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From above and slightly to one side, the crane settles into an oblique high-angle medium view of 김선영 floating face-down, keeping the composition clinical and non-graphic. Her pale nude torso is fully legible against the surrounding river surface while her face remains turned away and the frame leaves room along her body for the next move toward the ankles.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: river surface (flowing with small waves); used as Environmental field surrounding and carrying the body, with small waves providing restrained motion.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime light renders the pale body and dark river in a subdued neutral-to-cool palette with moderate contrast and no sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's motionless body remains nude and floats horizontally face-down on the river surface beside the bank, with her head and torso directed downward and her limbs extended in the prone orientation. A flesh-colored stocking hangs from the end of one ankle.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강가 수면 위로 엎드린 채 나신으로 떠 있는 김선영의 창백한 몸통\n\nLOCATION (lock): On the open surface of the river close to its bank, in daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From above and slightly to one side, the crane settles into an oblique high-angle medium view of 김선영 floating face-down, keeping the composition clinical and non-graphic. Her pale nude torso is fully legible against the surrounding river surface while her face remains turned away and the frame leaves room along her body for the next move toward the ankles.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: river surface (flowing with small waves); used as Environmental field surrounding and carrying the body, with small waves providing restrained motion.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime light renders the pale body and dark river in a subdued neutral-to-cool palette with moderate contrast and no sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's motionless body remains nude and floats horizontally face-down on the river surface beside the bank, with her head and torso directed downward and her limbs extended in the prone orientation. A flesh-colored stocking hangs from the end of one ankle.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 위에서 약간 옆으로 내려다보며 몸은 강변을 따라 수면에 수평으로 엎드린 채 머리가 둑 쪽, 발이 강 중앙 쪽으로 향한다. 얼굴은 수면 쪽으로 돌려져 보이지 않는다. 시선·총구 등 조준 대상은 없다.",
    "built_space": "열린 강 수면과 왼쪽의 콘크리트 제방·잡초. 차량·조종면·좌석 없음. 카메라가 수면 위 사선 고각에서 몸을 내려다본다.",
    "entities": "검은 긴 머리의 한국인으로 보이는 여성 한 명. 샷이 요구하는 나신·창백한 몸통이 아니라 검은 긴팔·긴바지 의복을 입고 있고, 살색 스타킹이 한쪽이 아니라 양쪽 발에 있다. 다른 인물은 없다.",
    "hard_violations": [],
    "physics": "몸은 강 수면에 떠 받쳐지며 작은 파문이 주변을 감싼다. 팔다리는 엎드린 자세로 뻗어 있고 공중에 떠 있지 않다."
   },
   {
    "label": "B",
    "direction": "사선 고각에서 몸은 수면에 엎드려 머리가 왼쪽 강 쪽, 발이 오른쪽 둑 쪽으로 향한다. 얼굴은 물에 잠겨 카메라를 보지 않는다. 조준 대상은 없다.",
    "built_space": "강 수면과 오른쪽의 돌무더기·잡초 제방, 멀리 다리·건물 실루엣. 실내 구조물은 없다. 카메라가 수면 위 사선에서 전신을 내려다본다.",
    "entities": "검은 긴 머리의 한국인으로 보이는 여성 한 명. 창백한 나신 몸통이 아니라 젖은 베이지 상의와 어두운 바지를 입고 있고, 살색 스타킹이 양쪽 발에 있다. 다른 인물은 없다.",
    "hard_violations": [],
    "physics": "몸이 수면에 떠 받쳐지고 팔다리가 엎드린 채 뻗어 있다. 공중 부양이나 지지 없는 물체는 없다."
   }
  ],
  "gq": {
   "route": "single_openrouter:x-ai/grok-4.6",
   "models": [
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 3,
   "B": 2
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "고각 중경·강변 수면·엎드린 부유는 맞지만 나신 몸통이 검복으로 가려지고 양발 스타킹이 있어 샷 핵심을 놓친다."
   },
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "역시 옷을 입혀 창백한 나신 몸통이 없고 양발 스타킹이며, 바위둑·원경 구조가 위치 록과 더 어긋난다."
   }
  ],
  "refs": [],
  "critique": {
   "issues": [
    {
     "issue_ko": "인물이 나체(나신) 상태여야 한다는 프롬프트의 명시적 지시와 달리, 위아래로 검은색 긴 옷을 완전히 갖춰 입고 있습니다.",
     "fix_en": "Replace all the black clothing with pale, bare human skin across the torso, arms, and legs so the figure is completely nude while floating face-down. Preserve the character's exact pose, her floating black hair, the river water, and the concrete bank.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "한쪽 발목에만 살구색 스타킹이 걸려 있어야 하지만, 양쪽 발 모두에 살구색 천(양말 형태)이 씌워져 있습니다.",
     "fix_en": "Remove the beige fabric covering both feet, rendering the feet and toes as bare skin, and render a single flesh-colored stocking hanging loosely from the end of one ankle only. Preserve the figure's leg positions, the overall pose, and the surrounding river surface.",
     "severity": "critical",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "인물이 나체(나신) 상태여야 한다는 프롬프트의 명시적 지시와 달리, 위아래로 검은색 긴 옷을 완전히 갖춰 입고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "한쪽 발목에만 살구색 스타킹이 걸려 있어야 하지만, 양쪽 발 모두에 살구색 천(양말 형태)이 씌워져 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "강에 엎드린 인물이 나신이 아니라 검은 상의·하의를 입고 있어 창백한 몸통이 전혀 보이지 않는다.",
     "severity": "critical"
    },
    {
     "issue_ko": "살색 스타킹이 한쪽 발목 끝에 걸려 있지 않고 양쪽 발에 양말처럼 신겨져 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 1,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace all the black clothing with pale, bare human skin across the torso, arms, and legs so the figure is completely nude while floating face-down. Preserve the character's exact pose, her floating black hair, the river water, and the concrete bank.\n- Remove the beige fabric covering both feet, rendering the feet and toes as bare skin, and render a single flesh-colored stocking hanging loosely from the end of one ankle only. Preserve the figure's leg positions, the overall pose, and the surrounding river surface.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 900,
      "verdict_ko": "인물이 검은색 옷을 입고 있어 '나신'이라는 핵심 지시를 완전히 어겼으며, 한쪽 발목에 걸쳐져 있어야 할 스타킹이 양발에 신겨져 있으나 치명적 위반은 피함.  ★위반: [openrouter:x-ai/grok-4.6] 나신이어야 할 김선영이 검은 전신 의상을 입고 있어 창백한 몸통이 보이지 않음 / [openrouter:x-ai/grok-4.6] 한쪽 발목에서 늘어져야 할 살색 스타킹이 양발에 신은 양말처럼 보임"
     },
     {
      "label": "B",
      "score": 1167,
      "verdict_ko": "배경 물가에 스노클을 쓴 사람 같은 불필요한 객체가 등장하여 치명적 위반으로 실격되었으며, '나신' 지시도 살구색 전신 타이츠로 잘못 표현함.  ★위반: [gemini-pro] invented people or objects (배경 수면에 스노클/잠수부 마스크 형태의 객체 추가됨, 둑 위의 검은 선) / [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 잠수부(추가 인물)가 프레임에 등장"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.4,
      "B": 1.667
     },
     "adjusted": {
      "A": 0.9,
      "B": 1.167
     },
     "violations": {
      "B": [
       "[gemini-pro] invented people or objects (배경 수면에 스노클/잠수부 마스크 형태의 객체 추가됨, 둑 위의 검은 선)",
       "[openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 잠수부(추가 인물)가 프레임에 등장"
      ],
      "A": [
       "[openrouter:x-ai/grok-4.6] 나신이어야 할 김선영이 검은 전신 의상을 입고 있어 창백한 몸통이 보이지 않음",
       "[openrouter:x-ai/grok-4.6] 한쪽 발목에서 늘어져야 할 살색 스타킹이 양발에 신은 양말처럼 보임"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.6,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 900,
      "verdict_ko": "인물이 검은색 옷을 입고 있어 '나신'이라는 핵심 지시를 완전히 어겼으며, 한쪽 발목에 걸쳐져 있어야 할 스타킹이 양발에 신겨져 있으나 치명적 위반은 피함.  ★위반: [openrouter:x-ai/grok-4.6] 나신이어야 할 김선영이 검은 전신 의상을 입고 있어 창백한 몸통이 보이지 않음 / [openrouter:x-ai/grok-4.6] 한쪽 발목에서 늘어져야 할 살색 스타킹이 양발에 신은 양말처럼 보임"
     },
     {
      "label": "B",
      "score": 1167,
      "verdict_ko": "배경 물가에 스노클을 쓴 사람 같은 불필요한 객체가 등장하여 치명적 위반으로 실격되었으며, '나신' 지시도 살구색 전신 타이츠로 잘못 표현함.  ★위반: [gemini-pro] invented people or objects (배경 수면에 스노클/잠수부 마스크 형태의 객체 추가됨, 둑 위의 검은 선) / [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 잠수부(추가 인물)가 프레임에 등장"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "검열 필터로 인해 '나신' 지시를 어기고 검은 옷을 입어 명도 및 의상 묘사에 실패했으나, A와 같은 치명적 생성 오류가 없어 우위를 차지함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "살구색 타이츠와 한쪽 발에 걸린 스타킹으로 원작 묘사를 시도했으나, 배경에 지시되지 않은 스노클 장비가 생성되는 치명적 오류로 실격됨."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "엎드린 인물의 머리가 화면 왼쪽을 향함.",
      "built_space": "좌측 상단에 경사진 콘크리트 강둑이 배치됨.",
      "entities": "검은색 상하의와 양발에 살구색 양말을 착용한 긴 머리 여성, 강물, 강둑.",
      "hard_violations": [],
      "physics": "몸이 강물 수면 위에 안정적으로 떠 있음."
     },
     {
      "label": "A",
      "direction": "엎드린 인물의 머리가 화면 왼쪽을 향함.",
      "built_space": "좌측 상단에 경사진 콘크리트 강둑이 배치됨.",
      "entities": "살구색 타이츠를 입고 한쪽 발에 천이 늘어진 긴 머리 여성, 강물, 강둑, 수면에 뜬 스노클 마스크.",
      "hard_violations": [
       "배경 수면에 프롬프트에 없는 발명된 물체(스노클 장비)가 생성됨."
      ],
      "physics": "몸이 강물 수면 위에 안정적으로 떠 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "검열 필터로 인해 '나신' 지시를 어기고 검은 옷을 입어 명도 및 의상 묘사에 실패했으나, A와 같은 치명적 생성 오류가 없어 우위를 차지함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "살구색 타이츠와 한쪽 발에 걸린 스타킹으로 원작 묘사를 시도했으나, 배경에 지시되지 않은 스노클 장비가 생성되는 치명적 오류로 실격됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "엎드린 인물의 머리가 화면 왼쪽을 향함.",
      "built_space": "좌측 상단에 경사진 콘크리트 강둑이 배치됨.",
      "entities": "검은색 상하의와 양발에 살구색 양말을 착용한 긴 머리 여성, 강물, 강둑.",
      "hard_violations": [],
      "physics": "몸이 강물 수면 위에 안정적으로 떠 있음."
     },
     {
      "label": "B",
      "direction": "엎드린 인물의 머리가 화면 왼쪽을 향함.",
      "built_space": "좌측 상단에 경사진 콘크리트 강둑이 배치됨.",
      "entities": "살구색 타이츠를 입고 한쪽 발에 천이 늘어진 긴 머리 여성, 강물, 강둑, 수면에 뜬 스노클 마스크.",
      "hard_violations": [
       "배경 수면에 프롬프트에 없는 발명된 물체(스노클 장비)가 생성됨."
      ],
      "physics": "몸이 강물 수면 위에 안정적으로 떠 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 905,
     "B": 1170
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S6sh4::signage": {
  "fp": "2b3c6bd1f13511f1",
  "inscriptions": []
 },
 "era_assess::6e8abcc7ad925b77": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 허름한 국밥집 내부와 테이블 세팅",
    "search_terms_native": [
     "허름한 국밥집 내부",
     "오래된 국밥집 테이블",
     "뚝배기 국밥 상차림",
     "식당 멜라민 그릇"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성되어야 하며, 영어나 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 특유의 국밥집 테이블 세팅(뚝배기 받침, 공깃밥, 멜라민 반찬 그릇, 수저통 등)과 특유의 서민적 인테리어는 일반적인 AI 모델이 정확히 재현하기 어렵고 중국이나 일본 식당처럼 그리기 쉽습니다."
   }
  ]
 },
 "era_ref::fe194f489320edd9": {
  "subject": "2000년대~2010년대 한국의 허름한 국밥집 내부와 테이블 세팅",
  "terms": [
   "허름한 국밥집 내부",
   "오래된 국밥집 테이블",
   "뚝배기 국밥 상차림",
   "식당 멜라민 그릇"
  ],
  "queries": [
   [
    "2000년대 2010년대 한국 허름한 국밥집 내부 오래된 테이블 멜라민 그릇 뚝배기 상차림",
    "한국 오래된 국밥집 테이블 뚝배기 국밥 반찬 멜라민 그릇"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://thumb.mt.co.kr/cdn-cgi/image/w%3D1200%2Ch%3D900%2Cfit%3Dcover%2Cbg%3Dwhilte%2Cf%3Dauto%2Cquality%3Dhigh%2Csharpen%3D2%2Cg%3Dface/21/2022/05/2022052912113877104_5.jpg",
  "picked_reason_ko": "1번은 2000~2010년대 한국의 허름한 국밥집에서 흔히 볼 수 있는 낡은 실내, 원형 목재 식탁, 등받이 없는 의자와 수저통 등의 모습을 가장 명확하게 보여준다.",
  "sha256": "6c177d0b9fb1fe43305abbf9c39bc3758567178326058220d4472022c6689bf5",
  "file": "eraref_fe194f489320edd9.png"
 },
 "groupbg::허름한 국밥집 내부": {
  "input_fingerprint": "1e11e23e7f82c570",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "허름한 국밥집 내부",
    "tags": [
     "S6sh4",
     "S6sh8"
    ]
   },
   "context_sig": "015ff15fdf5e23af",
   "era_research_sha": "6c177d0b9fb1fe43305abbf9c39bc3758567178326058220d4472022c6689bf5"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: At a dining table inside the shabby soup restaurant, facing the wall-mounted television.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n허름한 국밥집 내부: 낡고 소박한 서민식 식당 인테리어와 벽걸이 텔레비전이 있는 공간이다. (특징: 낡은 목재 또는 플라스틱 테이블; 벽에 걸린 얇은 평면 TV; 테이블 위에 놓인 뚝배기 그릇)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 허름한 국밥집 - 저녁\n- 그의 눈에 식당 벽걸이 TV로 '옛날 뉴스' 코너라는 종편 TV 채널이 보인다.\n\nTIME OF DAY (lock): evening.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 허름한 국밥집 내부와 테이블 세팅: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: At a dining table inside the shabby soup restaurant, facing the wall-mounted television.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n허름한 국밥집 내부: 낡고 소박한 서민식 식당 인테리어와 벽걸이 텔레비전이 있는 공간이다. (특징: 낡은 목재 또는 플라스틱 테이블; 벽에 걸린 얇은 평면 TV; 테이블 위에 놓인 뚝배기 그릇)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 허름한 국밥집 - 저녁\n- 그의 눈에 식당 벽걸이 TV로 '옛날 뉴스' 코너라는 종편 TV 채널이 보인다.\n\nTIME OF DAY (lock): evening.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 허름한 국밥집 내부와 테이블 세팅: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_허름한_국밥집_내부_85f3de.png",
  "asset_id": "7b3c1f67-5f6b-4118-84fc-901eef37bb01",
  "input_asset_ids": [
   "26303edb-a0aa-47af-bf12-00d848ea8288"
  ],
  "origin_tag": "S6sh4",
  "place_text": "At a dining table inside the shabby soup restaurant, facing the wall-mounted television.",
  "origin_inputs": {
   "place_text": "At a dining table inside the shabby soup restaurant, facing the wall-mounted television.",
   "time_of_day_en": "evening",
   "conti_asset_id": "26303edb-a0aa-47af-bf12-00d848ea8288"
  },
  "era_research": {
   "subject": "2000년대~2010년대 한국의 허름한 국밥집 내부와 테이블 세팅",
   "terms": [
    "허름한 국밥집 내부",
    "오래된 국밥집 테이블",
    "뚝배기 국밥 상차림",
    "식당 멜라민 그릇"
   ],
   "queries": [
    [
     "2000년대 2010년대 한국 허름한 국밥집 내부 오래된 테이블 멜라민 그릇 뚝배기 상차림",
     "한국 오래된 국밥집 테이블 뚝배기 국밥 반찬 멜라민 그릇"
    ]
   ],
   "candidates": 4,
   "picked_index": 1,
   "picked_url": "https://thumb.mt.co.kr/cdn-cgi/image/w%3D1200%2Ch%3D900%2Cfit%3Dcover%2Cbg%3Dwhilte%2Cf%3Dauto%2Cquality%3Dhigh%2Csharpen%3D2%2Cg%3Dface/21/2022/05/2022052912113877104_5.jpg",
   "picked_reason_ko": "1번은 2000~2010년대 한국의 허름한 국밥집에서 흔히 볼 수 있는 낡은 실내, 원형 목재 식탁, 등받이 없는 의자와 수저통 등의 모습을 가장 명확하게 보여준다.",
   "sha256": "6c177d0b9fb1fe43305abbf9c39bc3758567178326058220d4472022c6689bf5",
   "file": "eraref_fe194f489320edd9.png",
   "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_fe194f489320edd9.png"
  }
 },
 "era_assess::ed20f4647737b943": {
  "subjects": [
   {
    "subject_native": "2000년대 한국의 허름한 국밥집 내부",
    "search_terms_native": [
     "2000년대 국밥집 내부",
     "오래된 식당 내부",
     "기사식당 내부",
     "식당 벽걸이 TV"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성되어야 하며, 영어를 포함한 다른 외국어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 한국의 허름한 국밥집 내부 인테리어와 집기류(뚝배기, 스텐 공기밥그릇, 수저통이 달린 테이블 등) 및 벽걸이 TV의 배치 형태는 한국인에게 매우 친숙하고 독특한 시공간적 특징을 가지므로 일반적인 생성 모델은 이를 재현하지 못하고 서구식 수프 전문점처럼 그릴 우려가 있습니다."
   }
  ]
 },
 "era_ref::3c92d4a5f7ca8eb5": {
  "subject": "2000년대 한국의 허름한 국밥집 내부",
  "terms": [
   "2000년대 국밥집 내부",
   "오래된 식당 내부",
   "기사식당 내부",
   "식당 벽걸이 TV"
  ],
  "queries": [
   [
    "2000년대 한국 허름한 국밥집 오래된 식당 내부 벽걸이 TV",
    "한국 옛날 기사식당 내부 벽걸이 텔레비전"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDEyMjhfMjg3/MDAxNzM1MzE3MzA5MDAx.Y4vd-bqOpFG0kWRg-w57UWMZfA241evP8pFrXXFdzm8g.WshByFXKou5AmEGUifuRA31tmtdHm1Mbt-vmvThzenUg.JPEG/IMG_4715.JPG?type=w800",
  "picked_reason_ko": "2번은 2010년대 한국의 낡고 소박한 국밥집에서 흔히 볼 수 있는 목재 식탁·의자, 얼룩진 벽지, 물병과 양념통, 벽걸이 냉난방기 등이 선명하게 보여 가장 실용적인 내부 참고 사진이다.",
  "sha256": "3d6b17e5ac7bcc52f93c0b900ecca2615506b63106d99dae6ffbf578d32cd7fc",
  "file": "eraref_3c92d4a5f7ca8eb5.png"
 },
 "S6sh4::bgfirst_bg": {
  "input_fingerprint": "1a875e47c5a909b0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): TV 화면을 응시하며 미간에 깊은 주름이 패인 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At a dining table inside the shabby soup restaurant, facing the wall-mounted television.\n\nTIME OF DAY (lock): evening.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the closest point of a slow push, frame 전택수 from slightly above his eye line on his front three-quarter side, placing his face off-center and preserving viewing space toward the television. His brow tightens into a deep crease as he remains absorbed in the news playing just beyond the frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, foreground, looks toward wall-mounted television.\n- KEY BACKGROUND ELEMENTS: dining table edge (supporting the unfinished meal); used as Soft lower-edge context connecting his reaction to the interrupted meal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening interior illumination is complemented by subtle television-screen light, maintaining a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 허름한 국밥집 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): TV 화면을 응시하며 미간에 깊은 주름이 패인 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At a dining table inside the shabby soup restaurant, facing the wall-mounted television.\n\nTIME OF DAY (lock): evening.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the closest point of a slow push, frame 전택수 from slightly above his eye line on his front three-quarter side, placing his face off-center and preserving viewing space toward the television. His brow tightens into a deep crease as he remains absorbed in the news playing just beyond the frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, foreground, looks toward wall-mounted television.\n- KEY BACKGROUND ELEMENTS: dining table edge (supporting the unfinished meal); used as Soft lower-edge context connecting his reaction to the interrupted meal.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening interior illumination is complemented by subtle television-screen light, maintaining a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 허름한 국밥집 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S6sh4__bgfirst_bg.png",
  "asset_id": "62992251-6112-43dc-aa1f-9c81a11350ec",
  "input_asset_ids": [
   "26303edb-a0aa-47af-bf12-00d848ea8288",
   "7b3c1f67-5f6b-4118-84fc-901eef37bb01"
  ],
  "era_research": {
   "subject": "2000년대 한국의 허름한 국밥집 내부",
   "queries": [
    [
     "2000년대 한국 허름한 국밥집 오래된 식당 내부 벽걸이 TV",
     "한국 옛날 기사식당 내부 벽걸이 텔레비전"
    ]
   ],
   "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDEyMjhfMjg3/MDAxNzM1MzE3MzA5MDAx.Y4vd-bqOpFG0kWRg-w57UWMZfA241evP8pFrXXFdzm8g.WshByFXKou5AmEGUifuRA31tmtdHm1Mbt-vmvThzenUg.JPEG/IMG_4715.JPG?type=w800",
   "sha256": "3d6b17e5ac7bcc52f93c0b900ecca2615506b63106d99dae6ffbf578d32cd7fc",
   "file": "eraref_3c92d4a5f7ca8eb5.png"
  }
 },
 "S6sh4": {
  "input_fingerprint": "5a8ca9d5fe45d22b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): TV 화면을 응시하며 미간에 깊은 주름이 패인 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At a dining table inside the shabby soup restaurant, facing the wall-mounted television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the closest point of a slow push, frame 전택수 from slightly above his eye line on his front three-quarter side, placing his face off-center and preserving viewing space toward the television. His brow tightens into a deep crease as he remains absorbed in the news playing just beyond the frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, foreground, looks toward wall-mounted television.\n- KEY BACKGROUND ELEMENTS: dining table edge (supporting the unfinished meal); used as Soft lower-edge context connecting his reaction to the interrupted meal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening interior illumination is complemented by subtle television-screen light, maintaining a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): TV 화면을 응시하며 미간에 깊은 주름이 패인 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At a dining table inside the shabby soup restaurant, facing the wall-mounted television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the closest point of a slow push, frame 전택수 from slightly above his eye line on his front three-quarter side, placing his face off-center and preserving viewing space toward the television. His brow tightens into a deep crease as he remains absorbed in the news playing just beyond the frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, foreground, looks toward wall-mounted television.\n- KEY BACKGROUND ELEMENTS: dining table edge (supporting the unfinished meal); used as Soft lower-edge context connecting his reaction to the interrupted meal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening interior illumination is complemented by subtle television-screen light, maintaining a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): TV 화면을 응시하며 미간에 깊은 주름이 패인 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At a dining table inside the shabby soup restaurant, facing the wall-mounted television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the closest point of a slow push, frame 전택수 from slightly above his eye line on his front three-quarter side, placing his face off-center and preserving viewing space toward the television. His brow tightens into a deep crease as he remains absorbed in the news playing just beyond the frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, foreground, looks toward wall-mounted television.\n- KEY BACKGROUND ELEMENTS: dining table edge (supporting the unfinished meal); used as Soft lower-edge context connecting his reaction to the interrupted meal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained evening interior illumination is complemented by subtle television-screen light, maintaining a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S6sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S6sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_허름한_국밥집_내부_85f3de.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1667,
      "verdict_ko": "클로즈업 샷 크기 지침을 완벽히 충족하진 못했으나, 레퍼런스의 공간 구조와 캐릭터의 지정된 의상을 정확하게 구현하여 더 우수함."
     },
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "인물의 프레이밍은 프롬프트에 약간 더 가깝지만, 필수 지침인 캐릭터의 의상(모자 누락 및 임의의 재킷)을 위반하여 감점됨."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.667
     },
     "adjusted": {
      "A": 1.667,
      "B": 1.667
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.333,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1667,
      "verdict_ko": "클로즈업 샷 크기 지침을 완벽히 충족하진 못했으나, 레퍼런스의 공간 구조와 캐릭터의 지정된 의상을 정확하게 구현하여 더 우수함."
     },
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "인물의 프레이밍은 프롬프트에 약간 더 가깝지만, 필수 지침인 캐릭터의 의상(모자 누락 및 임의의 재킷)을 위반하여 감점됨."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "기준 의상과 공간 구조를 정확히 구현하고 시선 방향도 일치하나, 요구된 얼굴 클로즈업보다 프레임이 넓고 화면 밖이어야 할 TV가 화면 내에 포함되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레임 크기는 클로즈업에 더 가깝지만, 지정된 의상(모자, 재킷)을 완전히 누락했으며 응시 대상인 TV가 시선 뒤쪽에 배치되는 심각한 연출 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "화면 밖 왼쪽을 바라보고 있으나, 시선의 대상이 되어야 할 TV는 인물의 우측 뒤편(화면 좌측 상단)에 위치하여 시선 방향과 어긋납니다.",
      "built_space": "식당 배경과 원목 테이블이 구현되었으나, 인물이 바라보는 방향과 TV의 위치가 기준 사진의 공간 기하학과 모순되게 배치되었습니다.",
      "entities": "얼굴 생김새는 기준과 유사하지만, 지정된 의상(수사 모자, 네이비 경찰 재킷)을 전혀 착용하지 않고 다른 옷을 입고 있습니다.",
      "hard_violations": [
       "인물이 응시해야 할 TV가 시선 방향과 반대인 등 뒤 배경에 배치되어 물리적으로 불가능한 연출(staging)입니다."
      ],
      "physics": "의자에 앉아 식탁에 체중을 싣고 있는 자세가 자연스러우며, 허공에 떠 있는 물체는 없습니다."
     },
     {
      "label": "B",
      "direction": "화면 왼쪽을 향해 시선을 고정하고 있으며, 그 시선이 닿는 동선 상의 왼쪽 배경에 TV가 위치해 있습니다.",
      "built_space": "기준 사진의 공간 구조에 맞게 왼쪽 벽면에 TV가 배치되었고, 인물이 테이블에 앉아 이를 바라보는 공간적 관계가 올바르게 성립합니다.",
      "entities": "인물의 얼굴과 체형뿐만 아니라, 지정된 수사 모자와 로고가 있는 네이비색 경찰 재킷 등 기준 의상을 매우 정확하게 착용하고 있습니다.",
      "hard_violations": [],
      "physics": "식탁에 팔을 얹고 앉아 있는 자세가 지지물에 닿아 있으며, 테이블 위의 그릇과 식기들도 안정적으로 놓여 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "기준 의상과 공간 구조를 정확히 구현하고 시선 방향도 일치하나, 요구된 얼굴 클로즈업보다 프레임이 넓고 화면 밖이어야 할 TV가 화면 내에 포함되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레임 크기는 클로즈업에 더 가깝지만, 지정된 의상(모자, 재킷)을 완전히 누락했으며 응시 대상인 TV가 시선 뒤쪽에 배치되는 심각한 연출 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "화면 밖 왼쪽을 바라보고 있으나, 시선의 대상이 되어야 할 TV는 인물의 우측 뒤편(화면 좌측 상단)에 위치하여 시선 방향과 어긋납니다.",
      "built_space": "식당 배경과 원목 테이블이 구현되었으나, 인물이 바라보는 방향과 TV의 위치가 기준 사진의 공간 기하학과 모순되게 배치되었습니다.",
      "entities": "얼굴 생김새는 기준과 유사하지만, 지정된 의상(수사 모자, 네이비 경찰 재킷)을 전혀 착용하지 않고 다른 옷을 입고 있습니다.",
      "hard_violations": [
       "인물이 응시해야 할 TV가 시선 방향과 반대인 등 뒤 배경에 배치되어 물리적으로 불가능한 연출(staging)입니다."
      ],
      "physics": "의자에 앉아 식탁에 체중을 싣고 있는 자세가 자연스러우며, 허공에 떠 있는 물체는 없습니다."
     },
     {
      "label": "A",
      "direction": "화면 왼쪽을 향해 시선을 고정하고 있으며, 그 시선이 닿는 동선 상의 왼쪽 배경에 TV가 위치해 있습니다.",
      "built_space": "기준 사진의 공간 구조에 맞게 왼쪽 벽면에 TV가 배치되었고, 인물이 테이블에 앉아 이를 바라보는 공간적 관계가 올바르게 성립합니다.",
      "entities": "인물의 얼굴과 체형뿐만 아니라, 지정된 수사 모자와 로고가 있는 네이비색 경찰 재킷 등 기준 의상을 매우 정확하게 착용하고 있습니다.",
      "hard_violations": [],
      "physics": "식탁에 팔을 얹고 앉아 있는 자세가 지지물에 닿아 있으며, 테이블 위의 그릇과 식기들도 안정적으로 놓여 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1673,
     "B": 1670
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 1673,
   "B": 1670
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1667,
    "verdict_ko": "클로즈업 샷 크기 지침을 완벽히 충족하진 못했으나, 레퍼런스의 공간 구조와 캐릭터의 지정된 의상을 정확하게 구현하여 더 우수함."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "인물의 프레이밍은 프롬프트에 약간 더 가깝지만, 필수 지침인 캐릭터의 의상(모자 누락 및 임의의 재킷)을 위반하여 감점됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_허름한_국밥집_내부_85f3de.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트의 '클로즈업' 지시와 인물의 얼굴·어깨가 화면 우측을 크게 채우는 레이아웃 스케치의 비율과 달리, 생성된 이미지는 피사체의 허리 위까지 모두 보이는 미디엄 샷으로 렌더링되어 프레이밍 스케일을 심각하게 위반했습니다.",
     "fix_en": "Darken the character's lower body to draw focus to his face as a mitigation for the incorrect framing scale. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "카메라가 눈선보다 약간 위가 아니라 모자 챙 아래가 드러나는 눈높이·약간 아래에 있다.",
     "fix_en": "Tilt the hat brim down slightly to mimic a higher camera angle. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "재킷 글자가 참조의 SAULIS/POLICE가 아니라 SALIKS/POLICS로 깨져 있다.",
     "fix_en": "Remove the misspelled text on the jacket patches, leaving them as plain dark fabric. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트의 '클로즈업' 지시와 인물의 얼굴·어깨가 화면 우측을 크게 채우는 레이아웃 스케치의 비율과 달리, 생성된 이미지는 피사체의 허리 위까지 모두 보이는 미디엄 샷으로 렌더링되어 프레이밍 스케일을 심각하게 위반했습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신과 식당 전체가 보이는 미디엄 샷이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "카메라가 눈선보다 약간 위가 아니라 모자 챙 아래가 드러나는 눈높이·약간 아래에 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "재킷 글자가 참조의 SAULIS/POLICE가 아니라 SALIKS/POLICS로 깨져 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "카메라가 눈선보다 약간 위가 아니라 모자 챙 아래가 드러나는 눈높이·약간 아래에 있다.",
    "fix_en": "Tilt the hat brim down slightly to mimic a higher camera angle. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "재킷 글자가 참조의 SAULIS/POLICE가 아니라 SALIKS/POLICS로 깨져 있다.",
    "fix_en": "Remove the misspelled text on the jacket patches, leaving them as plain dark fabric. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Darken the character's lower body to draw focus to his face as a mitigation for the incorrect framing scale. Preserve the people present and their positions, their clothing, the set, the light, the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 클로즈업 프레이밍과 캐릭터의 미간 주름, 레이아웃 스케치의 전경 배치를 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 캐릭터를 중경에 작게 배치하여 샷의 기본 의도를 훼손함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선이 왼쪽 벽면의 TV 화면을 정확히 향함.",
      "built_space": "식당 내부, 스케치와 일치하는 전경 우측에 알맞게 위치함.",
      "entities": "전택수의 레퍼런스(얼굴, 의상)와 일치하며 미간의 주름이 명확히 표현됨.",
      "hard_violations": [],
      "physics": "테이블 뒤에 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "시선이 왼쪽 벽면의 TV 화면을 향함.",
      "built_space": "스케치가 요구한 전경이 아닌 중경의 우측 테이블에 잘못 배치됨.",
      "entities": "의상은 일치하나, 샷이 멀어 요구된 미간의 깊은 주름이 드러나지 않음.",
      "hard_violations": [
       "프레이밍 위반 (클로즈업 대신 와이드샷)",
       "레이아웃 스케치 위치 위반 (전경 대신 중경 배치)"
      ],
      "physics": "의자에 자연스럽게 앉아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 클로즈업 프레이밍과 캐릭터의 미간 주름, 레이아웃 스케치의 전경 배치를 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 캐릭터를 중경에 작게 배치하여 샷의 기본 의도를 훼손함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 왼쪽 벽면의 TV 화면을 정확히 향함.",
      "built_space": "식당 내부, 스케치와 일치하는 전경 우측에 알맞게 위치함.",
      "entities": "전택수의 레퍼런스(얼굴, 의상)와 일치하며 미간의 주름이 명확히 표현됨.",
      "hard_violations": [],
      "physics": "테이블 뒤에 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "시선이 왼쪽 벽면의 TV 화면을 향함.",
      "built_space": "스케치가 요구한 전경이 아닌 중경의 우측 테이블에 잘못 배치됨.",
      "entities": "의상은 일치하나, 샷이 멀어 요구된 미간의 깊은 주름이 드러나지 않음.",
      "hard_violations": [
       "프레이밍 위반 (클로즈업 대신 와이드샷)",
       "레이아웃 스케치 위치 위반 (전경 대신 중경 배치)"
      ],
      "physics": "의자에 자연스럽게 앉아 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 클로즈업 구도와 레이아웃을 정확히 따랐으며, 샷 텍스트가 요구한 '미간에 깊은 주름이 패인' 표정을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "요구된 클로즈업 구도를 무시하고 와이드 샷으로 처리했으며, 인물의 미간 주름과 긴장된 표정도 누락되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 왼쪽 벽에 걸린 TV 화면을 향함.",
      "built_space": "식당 내부 배경을 자르지 않고 그대로 사용하여 카메라가 인물로부터 멀리 떨어져 있음.",
      "entities": "경찰 제복을 입은 전택수가 등장하나 얼굴이 다소 젊게 표현되었고 미간 주름이 없음.",
      "hard_violations": [],
      "physics": "테이블 앞 의자에 자연스럽게 앉아 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 프레임 왼쪽 바깥(TV가 있는 방향)을 강렬하게 응시함.",
      "built_space": "레이아웃 스케치에 맞춰 배경과 카메라가 올바르게 전진배치됨.",
      "entities": "제복을 입은 전택수이며, 지시문대로 미간에 깊은 주름이 패인 50대 남성의 얼굴을 정확히 묘사함.",
      "hard_violations": [],
      "physics": "식탁 앞 의자에 앉아 체중을 싣고 있는 자세가 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 클로즈업 구도와 레이아웃을 정확히 따랐으며, 샷 텍스트가 요구한 '미간에 깊은 주름이 패인' 표정을 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 클로즈업 구도를 무시하고 와이드 샷으로 처리했으며, 인물의 미간 주름과 긴장된 표정도 누락되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물의 시선이 왼쪽 벽에 걸린 TV 화면을 향함.",
      "built_space": "식당 내부 배경을 자르지 않고 그대로 사용하여 카메라가 인물로부터 멀리 떨어져 있음.",
      "entities": "경찰 제복을 입은 전택수가 등장하나 얼굴이 다소 젊게 표현되었고 미간 주름이 없음.",
      "hard_violations": [],
      "physics": "테이블 앞 의자에 자연스럽게 앉아 있음."
     },
     {
      "label": "A",
      "direction": "인물의 시선이 프레임 왼쪽 바깥(TV가 있는 방향)을 강렬하게 응시함.",
      "built_space": "레이아웃 스케치에 맞춰 배경과 카메라가 올바르게 전진배치됨.",
      "entities": "제복을 입은 전택수이며, 지시문대로 미간에 깊은 주름이 패인 50대 남성의 얼굴을 정확히 묘사함.",
      "hard_violations": [],
      "physics": "식탁 앞 의자에 앉아 체중을 싣고 있는 자세가 자연스러움."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S6sh4__bgfirst_bg.png",
   "bg_asset_id": "62992251-6112-43dc-aa1f-9c81a11350ec",
   "bg_record_key": "S6sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "허름한 국밥집 내부",
   "groupbg_asset_id": "7b3c1f67-5f6b-4118-84fc-901eef37bb01"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S6sh4::cine": {
  "applied": true,
  "fingerprint": "a59ad77b4b4b9ca9823e9e508b6082476fbb89f8ac6a71d80e5c98b9bbf73cbb",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S6sh4_sel.png",
  "source_sha256": "6a2091a9c5f45a5cd4ce2aa1b536d5584fb11e96a94eb2fb1121fe20ca35b466",
  "file": "S6sh4_cine.png",
  "latency_ms": 11783
 },
 "S5sh2::cine": {
  "applied": true,
  "fingerprint": "604408a8a28a19af719b4e9b4b6ab439ee1e6bba0d09119db40bb7abc15937e4",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S5sh2_sel.png",
  "source_sha256": "631d3291c9e896647bc7204c1e784fbda629e5b6317a79dbd042ad9abc2439dc",
  "file": "S5sh2_cine.png",
  "latency_ms": 12233
 },
 "S6sh8::signage": {
  "fp": "2e14d70c7033c545",
  "inscriptions": [
   {
    "surface_native": "식당 유리문",
    "text_native": "안녕히 가십시오",
    "reason_ko": "식당 출입문 안쪽에 흔히 붙어 있는 고객 배웅용 문구로, 주인공이 밖으로 나가는 상황의 공간적 사실감을 더해줍니다."
   },
   {
    "surface_native": "문손잡이 위의 안내 스티커",
    "text_native": "미시오",
    "reason_ko": "출입문을 열고 나가는 방향을 안내하는 한국의 전형적인 유리문 표기입니다."
   }
  ]
 },
 "S6sh8": {
  "input_fingerprint": "bb34dd074627daa5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 식당 주인이 거스름돈을 든 채 뻗은 팔 너머로, 식당 출입문 밖을 향해 한 발을 내딛는 전택수의 뒷모습\n\nLOCATION (lock): At the restaurant's front doorway, moving from the dining room toward the outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track from just behind 식당 주인’s side at slightly below shoulder height, using the extended arm and change as a strong near-foreground diagonal without allowing the money to overwhelm the frame. Beyond the arm, 전택수 is caught from behind taking one decisive step through the open entrance, while the owner’s attention remains fixed on the departing customer.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, midground, moves toward open restaurant exit; open restaurant exit in the middle-right of the frame, background; 식당 주인 in the middle-left of the frame, foreground, reaches for departing 전택수.\n- KEY BACKGROUND ELEMENTS: change (held out toward 전택수); used as Small action detail held in the owner's extended foreground hand; restaurant entrance door (open) — Seen diagonally from inside the restaurant, with 전택수 crossing away from camera; used as Depth anchor framing 전택수's departure beyond the owner's arm.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic evening restaurant exposure keeps the foreground hand and departing figure readable in restrained neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 식당 주인 right now — 식당 주인's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 식당 주인 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 식당 주인 (Korean 성인, 중년, 세월감 있는 얼굴, 짧은 검은색 머리) — wearing: 편안한 라운드 티셔츠, 고무줄 바지 위로 때가 묻은 면 소재 식당 앞치마를 두른 모습 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 유리문: \"안녕히 가십시오\"\n- 문손잡이 위의 안내 스티커: \"미시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 식당 주인이 거스름돈을 든 채 뻗은 팔 너머로, 식당 출입문 밖을 향해 한 발을 내딛는 전택수의 뒷모습\n\nLOCATION (lock): At the restaurant's front doorway, moving from the dining room toward the outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track from just behind 식당 주인’s side at slightly below shoulder height, using the extended arm and change as a strong near-foreground diagonal without allowing the money to overwhelm the frame. Beyond the arm, 전택수 is caught from behind taking one decisive step through the open entrance, while the owner’s attention remains fixed on the departing customer.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, midground, moves toward open restaurant exit; open restaurant exit in the middle-right of the frame, background; 식당 주인 in the middle-left of the frame, foreground, reaches for departing 전택수.\n- KEY BACKGROUND ELEMENTS: change (held out toward 전택수); used as Small action detail held in the owner's extended foreground hand; restaurant entrance door (open) — Seen diagonally from inside the restaurant, with 전택수 crossing away from camera; used as Depth anchor framing 전택수's departure beyond the owner's arm.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic evening restaurant exposure keeps the foreground hand and departing figure readable in restrained neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 식당 주인 right now — 식당 주인's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 식당 주인 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 식당 주인 (Korean 성인, 중년, 세월감 있는 얼굴, 짧은 검은색 머리) — wearing: 편안한 라운드 티셔츠, 고무줄 바지 위로 때가 묻은 면 소재 식당 앞치마를 두른 모습 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 유리문: \"안녕히 가십시오\"\n- 문손잡이 위의 안내 스티커: \"미시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 식당 주인이 거스름돈을 든 채 뻗은 팔 너머로, 식당 출입문 밖을 향해 한 발을 내딛는 전택수의 뒷모습\n\nLOCATION (lock): At the restaurant's front doorway, moving from the dining room toward the outside. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track from just behind 식당 주인’s side at slightly below shoulder height, using the extended arm and change as a strong near-foreground diagonal without allowing the money to overwhelm the frame. Beyond the arm, 전택수 is caught from behind taking one decisive step through the open entrance, while the owner’s attention remains fixed on the departing customer.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, midground, moves toward open restaurant exit; open restaurant exit in the middle-right of the frame, background; 식당 주인 in the middle-left of the frame, foreground, reaches for departing 전택수.\n- KEY BACKGROUND ELEMENTS: change (held out toward 전택수); used as Small action detail held in the owner's extended foreground hand; restaurant entrance door (open) — Seen diagonally from inside the restaurant, with 전택수 crossing away from camera; used as Depth anchor framing 전택수's departure beyond the owner's arm.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic evening restaurant exposure keeps the foreground hand and departing figure readable in restrained neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 식당 주인 right now — 식당 주인's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 식당 주인 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 식당 주인 (Korean 성인, 중년, 세월감 있는 얼굴, 짧은 검은색 머리) — wearing: 편안한 라운드 티셔츠, 고무줄 바지 위로 때가 묻은 면 소재 식당 앞치마를 두른 모습 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 유리문: \"안녕히 가십시오\"\n- 문손잡이 위의 안내 스티커: \"미시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "식당 주인의 시선과 뻗은 팔이 열린 문 밖으로 나가는 전택수를 명확히 향하고 있습니다.",
    "built_space": "식당 내부는 레퍼런스 이미지와 일치하며, 출입문이 열려 있고 그 너머로 외부 거리가 보입니다. 카메라의 위치와 각도도 지시된 대로 주인의 옆/뒤쪽입니다.",
    "entities": "주인은 지시된 의상(티셔츠, 앞치마)을 입고 있으며, 전택수의 뒷모습은 레퍼런스와 정확히 일치합니다. 유리문과 스티커에 '안녕히 가십시오', '미시오'가 정확히 적혀 있습니다.",
    "hard_violations": [],
    "physics": "주인의 손이 지폐와 동전을 안정적으로 들고 있으며, 전택수는 걷는 동작에 맞게 지면을 딛고 체중을 싣고 있습니다."
   },
   {
    "label": "B",
    "direction": "식당 주인의 시선이 떠나는 손님이 아닌 카메라 렌즈를 정면으로 응시하고 있습니다.",
    "built_space": "식당 출입문이 닫혀 있으며, 배경의 공간이 외부가 아닌 식당 내부 벽면으로 연출되었습니다.",
    "entities": "인물들의 의상과 뒷모습 레퍼런스, 유리문에 적힌 텍스트 등은 프롬프트 지시사항과 잘 일치합니다.",
    "hard_violations": [
     "전택수가 열린 출입문을 나서는 것이 아니라 닫힌 문 옆 실내 안쪽으로 걷고 있음 (동선 및 공간 연출 위반)",
     "식당 주인이 떠나는 손님이 아닌 카메라 렌즈를 쳐다봄 (시선 지시 위반)"
    ],
    "physics": "인물의 걷는 자세와 주인이 손에 든 돈의 형태는 물리적인 지지를 잘 받고 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 8,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "열린 출입문으로 나서는 인물의 동선과 이를 바라보는 주인의 시선, 전경을 가로지르는 팔의 배치 등 프롬프트가 지시한 카메라 구도와 연출을 매우 충실히 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "출입문이 닫혀 있고 인물이 실내 안쪽을 향해 걷고 있으며, 주인이 카메라 렌즈를 정면으로 응시하여 핵심적인 동선 및 시선 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S6sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "식당 주인이 출입문을 향해 나가는 전택수에게 손을 뻗지 않고 카메라 앞쪽 허공을 향해 돈을 내밀고 있으며, 시선 역시 전택수가 아닌 엉뚱한 곳을 향해 있어 프롬프트의 상황과 공간감이 완전히 깨졌습니다.",
     "fix_en": "Redirect the owner's eyes to the man in the background, keeping the head and arm stationary. Preserve all characters, props, set, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "뒷모습으로 걸어 나가는 전택수의 모자 뒷면에 '수사'라는 텍스트가 잘못 배치되어 있습니다 (레퍼런스 이미지에서는 모자 정면에 위치함).",
     "fix_en": "Erase the yellow text from the back of the cap, replacing it with plain dark blue fabric. Preserve the cap's silhouette, the character, and the background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "화면 하단 중앙에 늘어뜨린 식당 주인의 왼손이, 손등이 보이는 상태에서 엄지손가락이 왼쪽을 향하고 있어 해부학적으로 오른손의 형태를 하고 있습니다.",
     "fix_en": "Redraw the hanging left hand to anatomically correct left-hand structure with the thumb pointing toward the thigh. Preserve the arm, apron, and background.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "전경 손에 들린 거스름돈이 한국 원이 아니라 미국 달러 지폐다",
     "fix_en": "Change the foreground banknotes to resemble indistinct 2000s Korean Won in pale yellowish-green tones. Preserve the holding hand, coins, and background.",
     "severity": "critical",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "식당 주인이 출입문을 향해 나가는 전택수에게 손을 뻗지 않고 카메라 앞쪽 허공을 향해 돈을 내밀고 있으며, 시선 역시 전택수가 아닌 엉뚱한 곳을 향해 있어 프롬프트의 상황과 공간감이 완전히 깨졌습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "뒷모습으로 걸어 나가는 전택수의 모자 뒷면에 '수사'라는 텍스트가 잘못 배치되어 있습니다 (레퍼런스 이미지에서는 모자 정면에 위치함).",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 하단 중앙에 늘어뜨린 식당 주인의 왼손이, 손등이 보이는 상태에서 엄지손가락이 왼쪽을 향하고 있어 해부학적으로 오른손의 형태를 하고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "식당 주인이 들고 있는 거스름돈 지폐가 한국 화폐가 아닌 미국 달러 지폐와 유사한 형태를 띠고 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "전경 손에 들린 거스름돈이 한국 원이 아니라 미국 달러 지폐다",
     "severity": "critical"
    },
    {
     "issue_ko": "주인의 뻗은 팔과 돈이 출구의 전택수가 아니라 카메라 전방을 향한다",
     "severity": "critical"
    },
    {
     "issue_ko": "카메라가 주인 옆뒤 어깨 높이가 아니라 주인 얼굴을 거의 정면에서 잡는다",
     "severity": "critical"
    },
    {
     "issue_ko": "식당 주인 시선이 떠나는 전택수가 아니라 카메라·전방 아래를 향한다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redirect the owner's eyes to the man in the background, keeping the head and arm stationary. Preserve all characters, props, set, lighting, and framing.\n- Erase the yellow text from the back of the cap, replacing it with plain dark blue fabric. Preserve the cap's silhouette, the character, and the background.\n- Redraw the hanging left hand to anatomically correct left-hand structure with the thumb pointing toward the thigh. Preserve the arm, apron, and background.\n- Change the foreground banknotes to resemble indistinct 2000s Korean Won in pale yellowish-green tones. Preserve the holding hand, coins, and background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "모자 뒷면을 논리적으로 비워두고, 시대와 배경에 걸맞은 한국 지폐를 정확히 생성하여 전반적인 퀄리티와 프롬프트 충실도가 매우 높습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 텍스트는 잘 구현했으나, 모자 뒷면에 '수사' 글자가 잘못 복제되었고 한국 배경임에도 미국 달러 지폐가 등장하여 감점되었습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "식당 주인의 시선과 뻗은 팔이 문을 나서는 전택수를 향하고 있으며, 전택수의 동선 역시 바깥을 향하고 있음.",
      "built_space": "식당의 내부 구조와 출입문이 레퍼런스와 일치하게 렌더링되었음. 유리문의 '안녕히 가십시오'와 손잡이 위 '미시오' 스티커가 정확한 위치에 올바른 텍스트로 배치됨.",
      "entities": "식당 주인과 전택수의 인상착의, 복장이 프롬프트 및 레퍼런스와 일치함. 주인이 들고 있는 거스름돈이 한국 지폐(인물 초상화 포함)로 바르게 묘사되었으며, 전택수의 모자 뒷면에 불필요한 텍스트가 없음.",
      "hard_violations": [],
      "physics": "주인의 손이 지폐와 동전을 자연스럽게 쥐고 있으며, 전택수의 걷는 자세와 무게 중심이 지면에 안정적으로 지탱되어 있음."
     },
     {
      "label": "A",
      "direction": "식당 주인의 시선과 팔의 방향이 나가는 전택수를 향하고 있으며, 전택수는 문 밖으로 향하고 있음.",
      "built_space": "레퍼런스와 동일한 식당 내부 공간이 구현되었고, 지정된 사이니지('안녕히 가십시오', '미시오')가 유리문과 손잡이 위에 잘 배치됨.",
      "entities": "인물들의 기본 복장은 지시사항을 따르나, 설정상 한국(2001-2017)임에도 식당 주인이 뻗은 손에 명백한 미국 달러 지폐가 들려 있음. 또한 전택수의 모자 뒷면에 앞면에 있어야 할 '수사' 텍스트가 잘못 생성됨.",
      "hard_violations": [
       "모자 앞면의 텍스트가 뒷면에 잘못 복제되어 렌더링됨(leaked text)",
       "공간 배경(한국)에 맞지 않는 미국 달러화가 등장하여 맥락적 오류 발생"
      ],
      "physics": "주인이 손가락으로 지폐를 쥐고 있으며, 전택수가 발을 내딛는 자세가 물리적으로 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "모자 뒷면을 논리적으로 비워두고, 시대와 배경에 걸맞은 한국 지폐를 정확히 생성하여 전반적인 퀄리티와 프롬프트 충실도가 매우 높습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 텍스트는 잘 구현했으나, 모자 뒷면에 '수사' 글자가 잘못 복제되었고 한국 배경임에도 미국 달러 지폐가 등장하여 감점되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "식당 주인의 시선과 뻗은 팔이 문을 나서는 전택수를 향하고 있으며, 전택수의 동선 역시 바깥을 향하고 있음.",
      "built_space": "식당의 내부 구조와 출입문이 레퍼런스와 일치하게 렌더링되었음. 유리문의 '안녕히 가십시오'와 손잡이 위 '미시오' 스티커가 정확한 위치에 올바른 텍스트로 배치됨.",
      "entities": "식당 주인과 전택수의 인상착의, 복장이 프롬프트 및 레퍼런스와 일치함. 주인이 들고 있는 거스름돈이 한국 지폐(인물 초상화 포함)로 바르게 묘사되었으며, 전택수의 모자 뒷면에 불필요한 텍스트가 없음.",
      "hard_violations": [],
      "physics": "주인의 손이 지폐와 동전을 자연스럽게 쥐고 있으며, 전택수의 걷는 자세와 무게 중심이 지면에 안정적으로 지탱되어 있음."
     },
     {
      "label": "A",
      "direction": "식당 주인의 시선과 팔의 방향이 나가는 전택수를 향하고 있으며, 전택수는 문 밖으로 향하고 있음.",
      "built_space": "레퍼런스와 동일한 식당 내부 공간이 구현되었고, 지정된 사이니지('안녕히 가십시오', '미시오')가 유리문과 손잡이 위에 잘 배치됨.",
      "entities": "인물들의 기본 복장은 지시사항을 따르나, 설정상 한국(2001-2017)임에도 식당 주인이 뻗은 손에 명백한 미국 달러 지폐가 들려 있음. 또한 전택수의 모자 뒷면에 앞면에 있어야 할 '수사' 텍스트가 잘못 생성됨.",
      "hard_violations": [
       "모자 앞면의 텍스트가 뒷면에 잘못 복제되어 렌더링됨(leaked text)",
       "공간 배경(한국)에 맞지 않는 미국 달러화가 등장하여 맥락적 오류 발생"
      ],
      "physics": "주인이 손가락으로 지폐를 쥐고 있으며, 전택수가 발을 내딛는 자세가 물리적으로 자연스러움."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "요구된 구도와 인물의 배치, 소품(거스름돈)을 정확히 구현했으며, 출입문 방향으로 나가는 인물의 뒷모습을 자연스럽게 묘사하여 가장 우수한 결과물입니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "전반적인 구도와 요소들은 훌륭하게 구현되었으나, 캐릭터 레퍼런스에서 모자 정면에 위치했던 '수사' 텍스트를 모자 뒷면에 부자연스럽게 추가한 점이 큰 감점 요인입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "식당 주인의 시선은 밖으로 나가는 전택수를 향하고 있으며, 거스름돈을 든 손 역시 그를 향해 뻗어 있습니다. 전택수는 열린 출입문을 향해 걷고 있습니다.",
      "built_space": "식당 내부에서 외부 출입문을 바라보는 구도이며, 요구된 '미시오'와 '안녕히 가십시오' 텍스트가 제 위치에 정확히 묘사되어 있습니다.",
      "entities": "식당 주인(면티, 앞치마)과 전택수(점퍼, 모자 뒷모습)의 인상착의가 레퍼런스와 일치하며, 지폐와 동전 등 거스름돈 소품도 잘 묘사되었습니다.",
      "hard_violations": [],
      "physics": "식당 주인의 서 있는 자세와 팔의 뻗음, 거스름돈을 쥐고 있는 손의 형태가 자연스러우며 전택수의 걷는 자세 또한 바닥에 안정적으로 지탱되어 있습니다."
     },
     {
      "label": "B",
      "direction": "식당 주인의 시선은 밖으로 나가는 전택수를 향하고 있으며, 거스름돈을 든 손 역시 그를 향해 뻗어 있습니다. 전택수는 열린 출입문을 향해 걷고 있습니다.",
      "built_space": "식당 내부에서 외부 출입문을 바라보는 구도이며, 요구된 '미시오'와 '안녕히 가십시오' 텍스트가 출입문에 정확히 적용되어 있습니다.",
      "entities": "식당 주인과 전택수의 인상착의 및 거스름돈 소품이 묘사되었으나, 전택수의 모자 뒷면에 정면에 있어야 할 '수사' 텍스트가 잘못 복제되어 있습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥에 안정적으로 서 있거나 걷고 있으며, 손에 들린 거스름돈 등 물리적 접촉과 지탱이 자연스럽게 표현되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "요구된 구도와 인물의 배치, 소품(거스름돈)을 정확히 구현했으며, 출입문 방향으로 나가는 인물의 뒷모습을 자연스럽게 묘사하여 가장 우수한 결과물입니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "전반적인 구도와 요소들은 훌륭하게 구현되었으나, 캐릭터 레퍼런스에서 모자 정면에 위치했던 '수사' 텍스트를 모자 뒷면에 부자연스럽게 추가한 점이 큰 감점 요인입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "식당 주인의 시선은 밖으로 나가는 전택수를 향하고 있으며, 거스름돈을 든 손 역시 그를 향해 뻗어 있습니다. 전택수는 열린 출입문을 향해 걷고 있습니다.",
      "built_space": "식당 내부에서 외부 출입문을 바라보는 구도이며, 요구된 '미시오'와 '안녕히 가십시오' 텍스트가 제 위치에 정확히 묘사되어 있습니다.",
      "entities": "식당 주인(면티, 앞치마)과 전택수(점퍼, 모자 뒷모습)의 인상착의가 레퍼런스와 일치하며, 지폐와 동전 등 거스름돈 소품도 잘 묘사되었습니다.",
      "hard_violations": [],
      "physics": "식당 주인의 서 있는 자세와 팔의 뻗음, 거스름돈을 쥐고 있는 손의 형태가 자연스러우며 전택수의 걷는 자세 또한 바닥에 안정적으로 지탱되어 있습니다."
     },
     {
      "label": "A",
      "direction": "식당 주인의 시선은 밖으로 나가는 전택수를 향하고 있으며, 거스름돈을 든 손 역시 그를 향해 뻗어 있습니다. 전택수는 열린 출입문을 향해 걷고 있습니다.",
      "built_space": "식당 내부에서 외부 출입문을 바라보는 구도이며, 요구된 '미시오'와 '안녕히 가십시오' 텍스트가 출입문에 정확히 적용되어 있습니다.",
      "entities": "식당 주인과 전택수의 인상착의 및 거스름돈 소품이 묘사되었으나, 전택수의 모자 뒷면에 정면에 있어야 할 '수사' 텍스트가 잘못 복제되어 있습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥에 안정적으로 서 있거나 걷고 있으며, 손에 들린 거스름돈 등 물리적 접촉과 지탱이 자연스럽게 표현되었습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 18
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S6sh4"
  }
 },
 "S6sh8::cine": {
  "applied": true,
  "fingerprint": "38510ae239c4dcd35b7a0d89057b3cf889d5013c03bcce8e51841267a96ad932",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S6sh8_sel.png",
  "source_sha256": "5324d5c73456ee1bc992fad031fad9ea9e3f598d5722ff49ab9f26c36ea13a0c",
  "file": "S6sh8_cine.png",
  "latency_ms": 11265
 },
 "S7sh4::confined_fp_apt": {
  "applies": false,
  "reason_ko": "샷이 자동차 내부에 위치하기는 하지만 지갑 안의 흑백사진을 보여주는 매우 타이트한 클로즈업 샷이므로 좌석 배치나 차량 조종 장치와의 상대적인 위치가 중요하게 작용하지 않습니다.",
  "input_fingerprint": "e68c235e48ecf527"
 },
 "S7sh4::signage": {
  "fp": "823d41b183509b28",
  "inscriptions": [
   {
    "surface_native": "학생증 카드",
    "text_native": "학생증\n한빛고등학교",
    "reason_ko": "지갑의 투명 창에 꽂힌 사진이 고등학교 학생증임을 보여주어 캐릭터의 신분과 시대적 배경을 사실적으로 묘사하기 위함."
   }
  ]
 },
 "S7sh4": {
  "input_fingerprint": "6e8a7f6d3ad9a72c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 활짝 펼쳐진 지갑 안쪽, 교복 입은 소녀의 흑백사진이 꽂혀 있는 투명 창 클로즈업\n\nLOCATION (lock): Inside the parked car at the driver's seat, surrounded closely by the seat, dashboard, windshield, and controls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a steep high angle just above the fully opened wallet, hold a close, slightly offset view as the dolly-in settles, placing the inner window and black-and-white schoolgirl photograph at the center. The photograph occupies less than two-fifths of the frame, with the surrounding wallet edges preserving its physical scale and context.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: open worn wallet (Fully open) — Its fully spread inner side faces upward toward the camera; used as Provides the close framing boundary around the photograph; wallet's transparent inner window (The photograph is inserted behind it) — The camera looks through the window directly at the photograph beneath it; used as Keeps the inserted photograph visible while forming a thin foreground layer over it; black-and-white photograph (Inserted in the wallet) — The image-bearing face is visible to camera and shows a girl in a school uniform; used as Primary focal content centered inside the wallet window.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued evening ambient light with a restrained neutral-to-cool palette and natural tonal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old wallet is open, with the black-and-white photograph of a school-uniformed girl fixed inside its transparent sleeve.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 학생증 카드: \"학생증\n한빛고등학교\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 활짝 펼쳐진 지갑 안쪽, 교복 입은 소녀의 흑백사진이 꽂혀 있는 투명 창 클로즈업\n\nLOCATION (lock): Inside the parked car at the driver's seat, surrounded closely by the seat, dashboard, windshield, and controls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a steep high angle just above the fully opened wallet, hold a close, slightly offset view as the dolly-in settles, placing the inner window and black-and-white schoolgirl photograph at the center. The photograph occupies less than two-fifths of the frame, with the surrounding wallet edges preserving its physical scale and context.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: open worn wallet (Fully open) — Its fully spread inner side faces upward toward the camera; used as Provides the close framing boundary around the photograph; wallet's transparent inner window (The photograph is inserted behind it) — The camera looks through the window directly at the photograph beneath it; used as Keeps the inserted photograph visible while forming a thin foreground layer over it; black-and-white photograph (Inserted in the wallet) — The image-bearing face is visible to camera and shows a girl in a school uniform; used as Primary focal content centered inside the wallet window.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued evening ambient light with a restrained neutral-to-cool palette and natural tonal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old wallet is open, with the black-and-white photograph of a school-uniformed girl fixed inside its transparent sleeve.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 학생증 카드: \"학생증\n한빛고등학교\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 활짝 펼쳐진 지갑 안쪽, 교복 입은 소녀의 흑백사진이 꽂혀 있는 투명 창 클로즈업\n\nLOCATION (lock): Inside the parked car at the driver's seat, surrounded closely by the seat, dashboard, windshield, and controls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a steep high angle just above the fully opened wallet, hold a close, slightly offset view as the dolly-in settles, placing the inner window and black-and-white schoolgirl photograph at the center. The photograph occupies less than two-fifths of the frame, with the surrounding wallet edges preserving its physical scale and context.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: open worn wallet (Fully open) — Its fully spread inner side faces upward toward the camera; used as Provides the close framing boundary around the photograph; wallet's transparent inner window (The photograph is inserted behind it) — The camera looks through the window directly at the photograph beneath it; used as Keeps the inserted photograph visible while forming a thin foreground layer over it; black-and-white photograph (Inserted in the wallet) — The image-bearing face is visible to camera and shows a girl in a school uniform; used as Primary focal content centered inside the wallet window.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued evening ambient light with a restrained neutral-to-cool palette and natural tonal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old wallet is open, with the black-and-white photograph of a school-uniformed girl fixed inside its transparent sleeve.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 학생증 카드: \"학생증\n한빛고등학교\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "카메라는 차량 내부 표면에 놓인 지갑 안쪽을 위에서 아래로 내려다봄.",
    "built_space": "자동차 실내 콘솔 또는 좌석의 질감이 배경으로 보임.",
    "entities": "완전히 펼쳐진 낡은 지갑, 투명 창, 교복을 입은 소녀의 흑백 사진, 지정된 '학생증 / 한빛고등학교' 텍스트 모두 일치함.",
    "hard_violations": [],
    "physics": "지갑이 차량 내부 표면에 안정적으로 지지되어 있음."
   },
   {
    "label": "A",
    "direction": "카메라는 차량 내부 표면에 놓인 지갑 안쪽을 위에서 아래로 내려다봄.",
    "built_space": "자동차 실내 콘솔 또는 좌석의 표면이 배경에 위치함.",
    "entities": "펼쳐진 지갑, 투명 창, 정확한 텍스트가 확인되나, 사진이 프롬프트의 흑백(black-and-white) 요구와 달리 컬러로 렌더링됨.",
    "hard_violations": [],
    "physics": "지갑이 차량 내부 표면에 안정적으로 놓여 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 8,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "프롬프트가 요구한 '흑백사진' 지시사항과 텍스트를 정확하게 구현하여 완성도가 높습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "텍스트는 정확하게 구현되었으나, 명시된 '흑백사진' 대신 컬러 사진이 생성되어 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L10B01.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "사진 속 소녀의 교복 명찰에 의미를 알 수 없는 뭉개진 문자가 생성됨.",
     "fix_en": "Repair would remove the meaningless squiggly text from the girl's nametag in the photograph.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "지갑이 놓인 배경이 레퍼런스 사진의 차량 내부 재질(회색 직물 시트 및 플라스틱 콘솔)과 일치하지 않는 스티치가 있는 검은색 가죽으로 묘사됨.",
     "fix_en": "Repair would change the background beneath the wallet to grey fabric and dark plastic to match the reference car interior.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "지갑의 가죽 표면이 프롬프트에서 지시한 '낡은(worn)' 상태가 아니라 새것처럼 깨끗하고 매끄럽게 표현됨.",
     "fix_en": "Repair would add scratches, scuffs, and worn edges to the black leather wallet.",
     "severity": "minor",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "사진 속 소녀의 교복 명찰에 의미를 알 수 없는 뭉개진 문자가 생성됨.",
     "severity": "major"
    },
    {
     "issue_ko": "지갑이 놓인 배경이 레퍼런스 사진의 차량 내부 재질(회색 직물 시트 및 플라스틱 콘솔)과 일치하지 않는 스티치가 있는 검은색 가죽으로 묘사됨.",
     "severity": "major"
    },
    {
     "issue_ko": "지갑의 가죽 표면이 프롬프트에서 지시한 '낡은(worn)' 상태가 아니라 새것처럼 깨끗하고 매끄럽게 표현됨.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "사진 속 소녀의 교복 명찰에 의미를 알 수 없는 뭉개진 문자가 생성됨.",
    "fix_en": "Repair would remove the meaningless squiggly text from the girl's nametag in the photograph.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "지갑이 놓인 배경이 레퍼런스 사진의 차량 내부 재질(회색 직물 시트 및 플라스틱 콘솔)과 일치하지 않는 스티치가 있는 검은색 가죽으로 묘사됨.",
    "fix_en": "Repair would change the background beneath the wallet to grey fabric and dark plastic to match the reference car interior.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "지갑의 가죽 표면이 프롬프트에서 지시한 '낡은(worn)' 상태가 아니라 새것처럼 깨끗하고 매끄럽게 표현됨.",
    "fix_en": "Repair would add scratches, scuffs, and worn edges to the black leather wallet.",
    "severity": "minor",
    "observation_index": 2
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S7sh4::cine": {
  "applied": true,
  "fingerprint": "d86e5b89cc88c38b74eb8fed5d9af57b83570b83eefb915cb33201f80aa1a148",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S7sh4_sel.png",
  "source_sha256": "a480c903b6744ed574e0da5e7c704ec20135f2c54c8aebae66742d0d9a1b6f77",
  "file": "S7sh4_cine.png",
  "latency_ms": 11419
 },
 "S7sh6::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 숏은 자동차 운전석 내부를 배경으로 하며, 인물이 앞유리를 바라보며 시동 키를 돌리는 조작을 하고 있으므로 운전석 및 제어 장치의 정확한 공간 배치가 매우 중요합니다.",
  "input_fingerprint": "c5bf8258fc7b3b2d"
 },
 "S7sh6::signage": {
  "fp": "9f1630545629ea50",
  "inscriptions": [
   {
    "surface_native": "대시보드 위의 주차 연락처 안내판",
    "text_native": "연락처 010-1234-5678",
    "reason_ko": "2000년대 한국 자동차 내부 대시보드에 흔히 비치되는 주차 알림 연락처판을 재현하여 사실성을 더하기 위함."
   }
  ]
 },
 "confinedfp::88cf3285b010": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/confinedfp_base_88cf3285b010.png",
  "place_text": "Inside the car's driver compartment, facing the windshield with the ignition and dashboard close at hand.",
  "input_fingerprint": "a851dccdd08fae58"
 },
 "S7sh6::confined_fp": {
  "reads": {
   "controls": "Ignition attached to the right side of the Driver station.",
   "mirrors": "None depicted.",
   "camera": "Positioned in the Front Passenger area on the left, facing right towards the Driver station.",
   "occupants": "전택수 occupies the Driver seat."
  },
  "mismatches": [
   "The text describes the ignition key as a 'near-foreground detail at lower right' and notes the camera looks 'across 전택수's hand at the ignition toward his face'. However, the diagram places the camera in the left-hand passenger seat and the ignition on the far right side of the right-hand driver's seat, meaning the ignition would physically be in the background, furthest from the camera, not in the near-foreground."
  ],
  "scene_description_en": "The camera is situated on the left side of the cabin in the empty front passenger area, looking diagonally rightward. In the right-center of the screen, in the midground, Taeksu sits in the driver's seat facing forward, presenting his left side to the camera. On the far right of the frame, in the background beyond Taeksu, the ignition control is visible. The left foreground and midground remain open, showing the empty passenger space. The empty rear seat stretches across the background from left to right behind the front seats.",
  "fixed": true,
  "input_fingerprint": "8dcac3aa4134ff94"
 },
 "era_assess::d47906bd64fe1708": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 대한민국 자동차 운전석 및 대시보드",
    "search_terms_native": [
     "2000년대 국산차 실내",
     "2010년대 자동차 대시보드",
     "국산 세단 운전석 내부"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대에서 2010년대 사이 한국에서 널리 타던 국산 세단(현대 소나타, 아반떼 등) 고유의 대시보드, 센터페시아 레이아웃 및 아날로그 계기판 형태를 정확하게 묘사하기 위함입니다."
   }
  ]
 },
 "era_ref::9e5928b133ecedd0": {
  "subject": "2000년대~2010년대 대한민국 자동차 운전석 및 대시보드",
  "terms": [
   "2000년대 국산차 실내",
   "2010년대 자동차 대시보드",
   "국산 세단 운전석 내부"
  ],
  "queries": [
   [
    "2000년대 국산 세단 실내 운전석 대시보드",
    "2010년대 국산 자동차 세단 운전석 내부 대시보드"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://cnp1753-prod.s3.amazonaws.com/cars/5BfIQApsvUxXxT/tbyLhgWtQh5T9EjHPPvH.jpg",
  "picked_reason_ko": "대한민국 내수용 르노삼성 승용차의 운전석과 대시보드가 정면에서 선명하게 보여, 2010년대의 일상적인 구성·재료·비례·조작계를 가장 충실히 참고할 수 있다.",
  "sha256": "61d1f31e90cb37ce69c6956a80afddcfbab18353630fe852766df874c4e03729",
  "file": "eraref_9e5928b133ecedd0.png"
 },
 "S7sh6": {
  "input_fingerprint": "a38ff87371ad53d5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 굳은 결의가 담긴 눈빛으로 정면 앞유리를 바라보며 시동 키를 쥐고 돌리는 자세의 전택수\n\nLOCATION (lock): Inside the car's driver compartment, facing the windshield with the ignition and dashboard close at hand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At dashboard height on the passenger side, settle from the track into a medium three-quarter view diagonally across 전택수's hand at the ignition toward his face. The turned key remains a modest near-foreground detail at lower right while his upper body occupies the midground, his eyes fixed through the windshield rather than toward the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: ignition key (Turned to start the car) — The inserted key projects from the ignition toward 전택수's turning hand; used as Near-foreground action detail connecting his hand to the decision to depart; dashboard (No additional state specified) — Its passenger-side span recedes diagonally toward the driver's position; used as Supplies spatial context between the foreground hand and 전택수's face; windshield (No additional state specified) — The windshield lies beyond his face and shows the exterior ahead; used as Provides the distant direction of 전택수's resolved gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued evening ambient light preserves readable eyes and hand detail without introducing a specified light source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old wallet containing the girl's black-and-white photograph has been closed and remains with Taeksu as he starts the car.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 대시보드 위의 주차 연락처 안내판: \"연락처 010-1234-5678\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is situated on the left side of the cabin in the empty front passenger area, looking diagonally rightward. In the right-center of the screen, in the midground, Taeksu sits in the driver's seat facing forward, presenting his left side to the camera. On the far right of the frame, in the background beyond Taeksu, the ignition control is visible. The left foreground and midground remain open, showing the empty passenger space. The empty rear seat stretches across the background from left to right behind the front seats.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 굳은 결의가 담긴 눈빛으로 정면 앞유리를 바라보며 시동 키를 쥐고 돌리는 자세의 전택수\n\nLOCATION (lock): Inside the car's driver compartment, facing the windshield with the ignition and dashboard close at hand. The shot takes place here.\n",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is situated on the left side of the cabin in the empty front passenger area, looking diagonally rightward. In the right-center of the screen, in the midground, Taeksu sits in the driver's seat facing forward, presenting his left side to the camera. On the far right of the frame, in the background beyond Taeksu, the ignition control is visible. The left foreground and midground remain open, showing the empty passenger space. The empty rear seat stretches across the background from left to right behind the front seats.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): evening.\n\nSHOT TEXT (authoritative, Korean): 굳은 결의가 담긴 눈빛으로 정면 앞유리를 바라보며 시동 키를 쥐고 돌리는 자세의 전택수\n\nLOCATION (lock): Inside the car's driver compartment, facing the windshield with the ignition and dashboard close at hand. The shot takes place here.\n\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S7sh6_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대~2010년대 대한민국 자동차 운전석 및 대시보드: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_9e5928b133ecedd0.png"
    },
    {
     "label": "전택수",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S7sh6_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대~2010년대 대한민국 자동차 운전석 및 대시보드: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_9e5928b133ecedd0.png"
    },
    {
     "label": "전택수",
     "path": "<bytes:917702>"
    }
   ]
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.667,
    "B": 1.667
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1667,
   "A": 1667
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "시선이 정면 앞유리를 향해 샷 텍스트의 액션을 따랐으나, 카메라 위치가 지정된 조수석 대시보드가 아닌 뒷좌석으로 물러나 근경 구도 요구를 놓쳤습니다."
   },
   {
    "label": "A",
    "score": 1667,
    "verdict_ko": "카메라 위치가 뒷좌석으로 밀려나 구도를 어긴 데다, 운전자의 시선마저 정면이 아닌 우측 창문 밖을 향해 핵심 지시에서 크게 벗어났습니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S7sh6_confinedfp.png"
   },
   {
    "label": "PERIOD REFERENCE — 2000년대~2010년대 대한민국 자동차 운전석 및 대시보드: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_9e5928b133ecedd0.png"
   },
   {
    "label": "전택수",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라와 가장 가까운 인물의 왼쪽 어깨에서 이어진 팔이 시동 키를 잡고 있으나, 그 끝에 있는 손은 오른손 형태(엄지손가락이 왼쪽을 향함)로 그려져 해부학적으로 불가능합니다.",
     "fix_en": "Redraw the hand holding the ignition key into an anatomically correct left hand with the thumb correctly positioned. Preserve the character, his position, police uniform, the set, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "지시문에서 왼쪽 전경과 중경은 빈 조수석 공간을 보여주며 '열려 있어야(remain open)' 한다고 명시했으나, 조수석 시트가 전경을 크게 가로막고 있습니다.",
     "fix_en": "Remove the tall seat backrest in the left foreground to reveal an open passenger space. Preserve the character, his position, clothing, the set, lighting, and framing.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "지시문에서 빈 뒷좌석이 앞좌석 뒤로 '배경을 가로질러(stretches across the background)' 보여야 한다고 했으나, 배경이 대부분 우측 차창 밖 풍경으로 채워져 뒷좌석의 가로 배치 구조가 나타나지 않습니다.",
     "fix_en": "Add a rear passenger seat stretching horizontally across the background behind the driver. Preserve the character, his position, clothing, the set, lighting, and framing.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 중앙의 전택수가 앞유리가 아니라 카메라 쪽을 바라보고 있다",
     "fix_en": "Adjust the character's face and eyes to gaze straight ahead toward the windshield. Preserve the character, his position, clothing, the set, lighting, and framing.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라와 가장 가까운 인물의 왼쪽 어깨에서 이어진 팔이 시동 키를 잡고 있으나, 그 끝에 있는 손은 오른손 형태(엄지손가락이 왼쪽을 향함)로 그려져 해부학적으로 불가능합니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "지시문에서 왼쪽 전경과 중경은 빈 조수석 공간을 보여주며 '열려 있어야(remain open)' 한다고 명시했으나, 조수석 시트가 전경을 크게 가로막고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "지시문에서 빈 뒷좌석이 앞좌석 뒤로 '배경을 가로질러(stretches across the background)' 보여야 한다고 했으나, 배경이 대부분 우측 차창 밖 풍경으로 채워져 뒷좌석의 가로 배치 구조가 나타나지 않습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙의 전택수가 앞유리가 아니라 카메라 쪽을 바라보고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "앞좌석 뒤 배경에 왼쪽에서 오른쪽으로 이어지는 빈 뒷좌석이 보이지 않는다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "지시문에서 왼쪽 전경과 중경은 빈 조수석 공간을 보여주며 '열려 있어야(remain open)' 한다고 명시했으나, 조수석 시트가 전경을 크게 가로막고 있습니다.",
    "fix_en": "Remove the tall seat backrest in the left foreground to reveal an open passenger space. Preserve the character, his position, clothing, the set, lighting, and framing.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "지시문에서 빈 뒷좌석이 앞좌석 뒤로 '배경을 가로질러(stretches across the background)' 보여야 한다고 했으나, 배경이 대부분 우측 차창 밖 풍경으로 채워져 뒷좌석의 가로 배치 구조가 나타나지 않습니다.",
    "fix_en": "Add a rear passenger seat stretching horizontally across the background behind the driver. Preserve the character, his position, clothing, the set, lighting, and framing.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "화면 중앙의 전택수가 앞유리가 아니라 카메라 쪽을 바라보고 있다",
    "fix_en": "Adjust the character's face and eyes to gaze straight ahead toward the windshield. Preserve the character, his position, clothing, the set, lighting, and framing.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the hand holding the ignition key into an anatomically correct left hand with the thumb correctly positioned. Preserve the character, his position, police uniform, the set, lighting, and framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "평면도와 씬 레이아웃이 명시한 '왼쪽 조수석에서 오른쪽 운전석을 바라보는 구도(인물의 왼쪽 측면 노출)'를 완전히 반대인 좌핸들 구조로 렌더링하여 하드 위반이 발생했으나, 저녁 시간대 설정과 의상 디테일은 정확히 반영하여 B보다 낫습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "A 후보와 마찬가지로 지정된 좌우 구도를 반대로 렌더링하는 하드 위반을 범했으며, 저녁 시간대 지시마저 무시하고 창밖을 대낮으로 표현하여 더욱 부적합합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 정면 앞유리를 향하고 있습니다. 오른손은 스티어링 휠 오른쪽의 시동 장치(타겟)를 향해 뻗어 이를 쥐고 있습니다.",
      "built_space": "차량 내부 공간입니다. 1개의 스티어링 휠과 레퍼런스를 잘 따른 대시보드가 있으나, 운전석이 왼쪽, 조수석이 오른쪽에 있는 구조(LHD)입니다. 카메라는 오른쪽 조수석을 차지하고 왼쪽 운전석을 바라보고 있습니다.",
      "entities": "전택수(한국인 남성)는 레퍼런스와 일치하는 남색 재킷, 체크 셔츠, '수사' 캡 모자를 착용하고 있습니다. 재킷의 패치 디테일도 비교적 정확합니다.",
      "hard_violations": [
       "평면도 및 씬 레이아웃 위반: 프롬프트는 카메라가 왼쪽 조수석에 있고 인물의 왼쪽 측면이 화면의 우중간에 보여야 한다고 명시했으나, 차량의 좌우 구조를 반대로(좌핸들) 렌더링하여 오른쪽 조수석에서 인물의 오른쪽 측면을 촬영함."
      ],
      "physics": "인물은 운전석 시트에 앉아 체중을 지지하고 있으며, 오른손으로 시동 키를 쥐고 돌리는 자연스러운 자세를 취하고 있습니다. 떠 있거나 물리적으로 불가능하게 지탱된 물체는 없습니다."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 정면 앞유리를 향하고 있습니다. 오른손은 스티어링 휠 오른쪽의 시동 키(타겟)를 쥐고 있습니다.",
      "built_space": "차량 내부 공간입니다. 1개의 스티어링 휠이 대시보드 안쪽 축과 다소 어색하게 연결(융합)되어 있습니다. 운전석이 왼쪽, 조수석이 오른쪽에 있는 구조이며, 카메라는 오른쪽 조수석에 위치합니다.",
      "entities": "전택수(한국인 남성)는 남색 재킷, 체크 셔츠, '수사' 캡 모자를 착용하고 있습니다.",
      "hard_violations": [
       "평면도 및 씬 레이아웃 위반: 프롬프트는 카메라가 왼쪽 조수석에 있고 인물의 왼쪽 측면이 화면의 우중간에 보여야 한다고 명시했으나, 차량의 좌우 구조를 반대로(좌핸들) 렌더링하여 오른쪽 조수석에서 인물의 오른쪽 측면을 촬영함."
      ],
      "physics": "인물은 운전석 시트에 앉아 안정적으로 지탱되고 있으며, 오른손으로 시동 키를 잡고 있습니다. 지지되지 않고 떠 있는 물체는 없습니다."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "평면도와 씬 레이아웃이 명시한 '왼쪽 조수석에서 오른쪽 운전석을 바라보는 구도(인물의 왼쪽 측면 노출)'를 완전히 반대인 좌핸들 구조로 렌더링하여 하드 위반이 발생했으나, 저녁 시간대 설정과 의상 디테일은 정확히 반영하여 B보다 낫습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "A 후보와 마찬가지로 지정된 좌우 구도를 반대로 렌더링하는 하드 위반을 범했으며, 저녁 시간대 지시마저 무시하고 창밖을 대낮으로 표현하여 더욱 부적합합니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 정면 앞유리를 향하고 있습니다. 오른손은 스티어링 휠 오른쪽의 시동 장치(타겟)를 향해 뻗어 이를 쥐고 있습니다.",
      "built_space": "차량 내부 공간입니다. 1개의 스티어링 휠과 레퍼런스를 잘 따른 대시보드가 있으나, 운전석이 왼쪽, 조수석이 오른쪽에 있는 구조(LHD)입니다. 카메라는 오른쪽 조수석을 차지하고 왼쪽 운전석을 바라보고 있습니다.",
      "entities": "전택수(한국인 남성)는 레퍼런스와 일치하는 남색 재킷, 체크 셔츠, '수사' 캡 모자를 착용하고 있습니다. 재킷의 패치 디테일도 비교적 정확합니다.",
      "hard_violations": [
       "평면도 및 씬 레이아웃 위반: 프롬프트는 카메라가 왼쪽 조수석에 있고 인물의 왼쪽 측면이 화면의 우중간에 보여야 한다고 명시했으나, 차량의 좌우 구조를 반대로(좌핸들) 렌더링하여 오른쪽 조수석에서 인물의 오른쪽 측면을 촬영함."
      ],
      "physics": "인물은 운전석 시트에 앉아 체중을 지지하고 있으며, 오른손으로 시동 키를 쥐고 돌리는 자연스러운 자세를 취하고 있습니다. 떠 있거나 물리적으로 불가능하게 지탱된 물체는 없습니다."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 정면 앞유리를 향하고 있습니다. 오른손은 스티어링 휠 오른쪽의 시동 키(타겟)를 쥐고 있습니다.",
      "built_space": "차량 내부 공간입니다. 1개의 스티어링 휠이 대시보드 안쪽 축과 다소 어색하게 연결(융합)되어 있습니다. 운전석이 왼쪽, 조수석이 오른쪽에 있는 구조이며, 카메라는 오른쪽 조수석에 위치합니다.",
      "entities": "전택수(한국인 남성)는 남색 재킷, 체크 셔츠, '수사' 캡 모자를 착용하고 있습니다.",
      "hard_violations": [
       "평면도 및 씬 레이아웃 위반: 프롬프트는 카메라가 왼쪽 조수석에 있고 인물의 왼쪽 측면이 화면의 우중간에 보여야 한다고 명시했으나, 차량의 좌우 구조를 반대로(좌핸들) 렌더링하여 오른쪽 조수석에서 인물의 오른쪽 측면을 촬영함."
      ],
      "physics": "인물은 운전석 시트에 앉아 안정적으로 지탱되고 있으며, 오른손으로 시동 키를 잡고 있습니다. 지지되지 않고 떠 있는 물체는 없습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "평면도가 지시한 카메라 및 좌석 위치(우핸들 레이아웃)를 완벽하게 구현하면서 레퍼런스의 대시보드 디테일을 살렸고, 저녁 시간대와 인물 착장까지 모두 정확하게 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "차량 내부 구조가 물리적으로 불가능하게 렌더링되었으며(왼쪽 조수석, 중앙 운전석, 우측 센터 콘솔), 저녁 시간대 설정 및 배경의 뒷좌석 지시를 모두 누락했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수는 프레임 우측(정면)을 향해 앞유리 쪽을 바라보고 있으며, 왼손을 뻗어 시동 키를 쥐고 돌리고 있습니다.",
      "built_space": "차량 내부 구조가 물리적으로 불가능하게 렌더링되었습니다. 왼쪽 전경에 조수석이 있고 중앙에 운전석이 있는데, 센터 콘솔과 내비게이션 화면이 운전석 우측에 위치해 있어 차량 구조가 성립하지 않습니다. 또한 배경에 뒷좌석이 보이지 않고 창밖으로 밝은 빈 벽만 보입니다.",
      "entities": "전택수의 얼굴과 의상(수사 모자, 남색 재킷, 체크 셔츠)은 캐릭터 레퍼런스와 잘 일치합니다. 하지만 시간대가 저녁이 아닌 주간이나 스튜디오 조명처럼 묘사되어 'evening' 고정 조건을 위반했습니다.",
      "hard_violations": [
       "물리적으로 불가능한 공간 구조 (왼쪽 전경에 조수석이 존재함에도 중앙 운전석의 우측에 센터 콘솔이 배치되는 모순된 차량 내부)"
      ],
      "physics": "전택수는 시트에 안정적으로 앉아 있으며, 왼손으로 키를 확실히 쥐고 있습니다. 지지되지 않고 떠 있는 물체나 신체는 없습니다."
     },
     {
      "label": "B",
      "direction": "전택수는 굳은 결의가 담긴 눈빛으로 프레임 우측(정면 앞유리)을 바라보고 있으며, 왼손을 뻗어 스티어링 칼럼 우측의 시동 키를 쥐고 돌리고 있습니다.",
      "built_space": "평면도가 지시한 카메라 위치(왼쪽 조수석)와 운전석 위치(오른쪽)에 맞춰 우핸들(RHD) 구조를 정확하게 구현했습니다. 왼쪽 전경의 조수석과 오른쪽 운전석 사이에 센터 콘솔이 올바르게 위치하며, 배경에는 지시된 빈 뒷좌석 공간과 창밖 풍경이 모두 보입니다.",
      "entities": "전택수의 인상착의(얼굴, 모자, 재킷 패치, 셔츠, 바지, 벨트)가 레퍼런스와 완벽히 일치합니다. 2000년대~2010년대 차량 대시보드의 형태와 질감도 훌륭하게 재현되었으며, 창밖으로 야경이 보여 'evening' 시간대 지정 조건도 훌륭하게 충족했습니다.",
      "hard_violations": [],
      "physics": "전택수는 시트에 안정적으로 앉아 체중이 지지되고 있으며, 키를 잡은 손의 형태와 그립도 자연스럽게 밀착되어 있습니다. 허공에 뜬 물체는 없습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "평면도가 지시한 카메라 및 좌석 위치(우핸들 레이아웃)를 완벽하게 구현하면서 레퍼런스의 대시보드 디테일을 살렸고, 저녁 시간대와 인물 착장까지 모두 정확하게 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "차량 내부 구조가 물리적으로 불가능하게 렌더링되었으며(왼쪽 조수석, 중앙 운전석, 우측 센터 콘솔), 저녁 시간대 설정 및 배경의 뒷좌석 지시를 모두 누락했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수는 프레임 우측(정면)을 향해 앞유리 쪽을 바라보고 있으며, 왼손을 뻗어 시동 키를 쥐고 돌리고 있습니다.",
      "built_space": "차량 내부 구조가 물리적으로 불가능하게 렌더링되었습니다. 왼쪽 전경에 조수석이 있고 중앙에 운전석이 있는데, 센터 콘솔과 내비게이션 화면이 운전석 우측에 위치해 있어 차량 구조가 성립하지 않습니다. 또한 배경에 뒷좌석이 보이지 않고 창밖으로 밝은 빈 벽만 보입니다.",
      "entities": "전택수의 얼굴과 의상(수사 모자, 남색 재킷, 체크 셔츠)은 캐릭터 레퍼런스와 잘 일치합니다. 하지만 시간대가 저녁이 아닌 주간이나 스튜디오 조명처럼 묘사되어 'evening' 고정 조건을 위반했습니다.",
      "hard_violations": [
       "물리적으로 불가능한 공간 구조 (왼쪽 전경에 조수석이 존재함에도 중앙 운전석의 우측에 센터 콘솔이 배치되는 모순된 차량 내부)"
      ],
      "physics": "전택수는 시트에 안정적으로 앉아 있으며, 왼손으로 키를 확실히 쥐고 있습니다. 지지되지 않고 떠 있는 물체나 신체는 없습니다."
     },
     {
      "label": "A",
      "direction": "전택수는 굳은 결의가 담긴 눈빛으로 프레임 우측(정면 앞유리)을 바라보고 있으며, 왼손을 뻗어 스티어링 칼럼 우측의 시동 키를 쥐고 돌리고 있습니다.",
      "built_space": "평면도가 지시한 카메라 위치(왼쪽 조수석)와 운전석 위치(오른쪽)에 맞춰 우핸들(RHD) 구조를 정확하게 구현했습니다. 왼쪽 전경의 조수석과 오른쪽 운전석 사이에 센터 콘솔이 올바르게 위치하며, 배경에는 지시된 빈 뒷좌석 공간과 창밖 풍경이 모두 보입니다.",
      "entities": "전택수의 인상착의(얼굴, 모자, 재킷 패치, 셔츠, 바지, 벨트)가 레퍼런스와 완벽히 일치합니다. 2000년대~2010년대 차량 대시보드의 형태와 질감도 훌륭하게 재현되었으며, 창밖으로 야경이 보여 'evening' 시간대 지정 조건도 훌륭하게 충족했습니다.",
      "hard_violations": [],
      "physics": "전택수는 시트에 안정적으로 앉아 체중이 지지되고 있으며, 키를 잡은 손의 형태와 그립도 자연스럽게 밀착되어 있습니다. 허공에 뜬 물체는 없습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "confined_fp": {
   "base_key": "confinedfp::88cf3285b010",
   "apt_reason": "이 숏은 자동차 운전석 내부를 배경으로 하며, 인물이 앞유리를 바라보며 시동 키를 돌리는 조작을 하고 있으므로 운전석 및 제어 장치의 정확한 공간 배치가 매우 중요합니다.",
   "fixed": true,
   "mismatches": [
    "The text describes the ignition key as a 'near-foreground detail at lower right' and notes the camera looks 'across 전택수's hand at the ignition toward his face'. However, the diagram places the camera in the left-hand passenger seat and the ignition on the far right side of the right-hand driver's seat, meaning the ignition would physically be in the background, furthest from the camera, not in the near-foreground."
   ],
   "era_research": {
    "subject": "2000년대~2010년대 대한민국 자동차 운전석 및 대시보드",
    "queries": [
     [
      "2000년대 국산 세단 실내 운전석 대시보드",
      "2010년대 국산 자동차 세단 운전석 내부 대시보드"
     ]
    ],
    "picked_url": "https://cnp1753-prod.s3.amazonaws.com/cars/5BfIQApsvUxXxT/tbyLhgWtQh5T9EjHPPvH.jpg",
    "sha256": "61d1f31e90cb37ce69c6956a80afddcfbab18353630fe852766df874c4e03729",
    "file": "eraref_9e5928b133ecedd0.png"
   }
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S7sh4"
  }
 },
 "S7sh6::cine": {
  "applied": true,
  "fingerprint": "48c4b5388a73e5858eb8cef07495ace7709a28f47e23a218e677f987f7b3c817",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S7sh6_sel.png",
  "source_sha256": "40ea5ffded88f40da8ff8e69db28736b800fe0c9218b342002bc954309c4591b",
  "file": "S7sh6_cine.png",
  "latency_ms": 10472
 },
 "S8sh1::signage": {
  "fp": "7661d7ebc2106c6b",
  "inscriptions": [
   {
    "surface_native": "정문 기둥 현판",
    "text_native": "남부경찰서",
    "reason_ko": "경찰서 정문 진입로라는 공간적 배경을 명확히 전달하기 위해 입구 기둥에 경찰서 명칭을 표시합니다."
   }
  ]
 },
 "S8sh1": {
  "input_fingerprint": "5eff50f91908419e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 높이 솟은 메타세콰이어 나무들이 빽빽하게 늘어선 길 한가운데를 가로지르는 검은색 SUV의 앞모습\n\nLOCATION (lock): On the tree-lined exterior approach road leading through the police station's main gate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track backward near hood height on the avenue axis, maintaining a wide frontal view as the black SUV advances slowly at center. The vehicle remains a realistic portion of the lower-middle frame while the tall rows of metasequoia trees rise on both sides and establish the road's depth.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: black SUV in the middle-center of the frame, midground, moves toward camera position along the avenue; metasequoia-lined avenue in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: black SUV (Moving slowly into the grounds) — Its front faces the camera as it enters slowly; used as Primary moving subject held on the central road axis; tall metasequoia trees (Closely spaced in long rows); used as Forms two dense vertical rows that frame the vehicle and carry perspective into depth; entrance avenue (Extends beneath the front gate); used as Establishes the central approach line beneath the vehicle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 정문 기둥 현판: \"남부경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 높이 솟은 메타세콰이어 나무들이 빽빽하게 늘어선 길 한가운데를 가로지르는 검은색 SUV의 앞모습\n\nLOCATION (lock): On the tree-lined exterior approach road leading through the police station's main gate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track backward near hood height on the avenue axis, maintaining a wide frontal view as the black SUV advances slowly at center. The vehicle remains a realistic portion of the lower-middle frame while the tall rows of metasequoia trees rise on both sides and establish the road's depth.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: black SUV in the middle-center of the frame, midground, moves toward camera position along the avenue; metasequoia-lined avenue in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: black SUV (Moving slowly into the grounds) — Its front faces the camera as it enters slowly; used as Primary moving subject held on the central road axis; tall metasequoia trees (Closely spaced in long rows); used as Forms two dense vertical rows that frame the vehicle and carry perspective into depth; entrance avenue (Extends beneath the front gate); used as Establishes the central approach line beneath the vehicle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 정문 기둥 현판: \"남부경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 높이 솟은 메타세콰이어 나무들이 빽빽하게 늘어선 길 한가운데를 가로지르는 검은색 SUV의 앞모습\n\nLOCATION (lock): On the tree-lined exterior approach road leading through the police station's main gate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track backward near hood height on the avenue axis, maintaining a wide frontal view as the black SUV advances slowly at center. The vehicle remains a realistic portion of the lower-middle frame while the tall rows of metasequoia trees rise on both sides and establish the road's depth.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: black SUV in the middle-center of the frame, midground, moves toward camera position along the avenue; metasequoia-lined avenue in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: black SUV (Moving slowly into the grounds) — Its front faces the camera as it enters slowly; used as Primary moving subject held on the central road axis; tall metasequoia trees (Closely spaced in long rows); used as Forms two dense vertical rows that frame the vehicle and carry perspective into depth; entrance avenue (Extends beneath the front gate); used as Establishes the central approach line beneath the vehicle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 정문 기둥 현판: \"남부경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "차량의 전면이 카메라를 향해 정면으로 위치하고 있으며 접근 중임.",
    "built_space": "경찰서 건물이 배경에 있고, 정문 기둥과 열린 철문이 양옆에 배치됨. 차량은 도로 중앙에 있음.",
    "entities": "검은색 SUV, 메타세쿼이아 나무, 경찰서 건물, '남부경찰서' 텍스트 모두 확인됨.",
    "hard_violations": [],
    "physics": "차량의 바퀴가 도로 표면에 안정적으로 접지되어 있으며 그림자가 자연스러움."
   },
   {
    "label": "B",
    "direction": "차량의 전면이 카메라를 향해 정면으로 위치하고 있으며 접근 중임.",
    "built_space": "배경의 경찰서 건물과 정문 기둥, 철문이 원본 위치와 동일하게 배치됨.",
    "entities": "검은색 SUV, 메타세쿼이아 나무, 경찰서 건물, '남부경찰서' 텍스트 모두 확인됨.",
    "hard_violations": [],
    "physics": "차량의 바퀴가 도로에 닿아 있으며 그림자 방향이 일관됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지정된 텍스트 현판과 차량, 배경을 완벽히 구현했으며, 시네마틱한 대기감과 심도가 돋보입니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "지정된 요소들을 잘 구현하였으나, 차량 헤드라이트의 디테일과 대기감 면에서 A에 비해 약간 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L12B01.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "왼쪽 정문 기둥에 있는 '남부경찰서' 텍스트가 표면의 질감이나 조명에 자연스럽게 융화되지 않고, 평면적인 디지털 글꼴이 스티커처럼 합성된 것처럼 보입니다.",
     "fix_en": "Render the text on the left pillar as a physical engraved or mounted plaque matching the stone's lighting and texture. Preserve the SUV, trees, and background building.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "카메라가 후드 높이가 아니라 눈높이 부근 정면 와이드로, 검은 SUV·정문·청사가 위치 참조와 비슷한 높은 시점에서 잡혀 있다.",
     "fix_en": "Lower the camera to near hood height. Preserve the vehicle, trees, and architecture.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경 경찰서 청사 입면의 창 개수와 배치가 위치 참조 사진의 고정 건축과 다르다.",
     "fix_en": "Match the building's facade window count and arrangement to the reference. Preserve the SUV and trees.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "샷이 요구한 빽빽한 메타세콰이어 가로수 대신 양쪽 나무 간격이 넓어 중·원경 건물 벽이 크게 드러난다.",
     "fix_en": "Increase the density of the tree rows to closely match the reference. Preserve the SUV and background building.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "왼쪽 정문 기둥에 있는 '남부경찰서' 텍스트가 표면의 질감이나 조명에 자연스럽게 융화되지 않고, 평면적인 디지털 글꼴이 스티커처럼 합성된 것처럼 보입니다.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 후드 높이가 아니라 눈높이 부근 정면 와이드로, 검은 SUV·정문·청사가 위치 참조와 비슷한 높은 시점에서 잡혀 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 경찰서 청사 입면의 창 개수와 배치가 위치 참조 사진의 고정 건축과 다르다.",
     "severity": "major"
    },
    {
     "issue_ko": "샷이 요구한 빽빽한 메타세콰이어 가로수 대신 양쪽 나무 간격이 넓어 중·원경 건물 벽이 크게 드러난다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 정문 기둥에 있는 '남부경찰서' 텍스트가 표면의 질감이나 조명에 자연스럽게 융화되지 않고, 평면적인 디지털 글꼴이 스티커처럼 합성된 것처럼 보입니다.",
    "fix_en": "Render the text on the left pillar as a physical engraved or mounted plaque matching the stone's lighting and texture. Preserve the SUV, trees, and background building.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "카메라가 후드 높이가 아니라 눈높이 부근 정면 와이드로, 검은 SUV·정문·청사가 위치 참조와 비슷한 높은 시점에서 잡혀 있다.",
    "fix_en": "Lower the camera to near hood height. Preserve the vehicle, trees, and architecture.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경 경찰서 청사 입면의 창 개수와 배치가 위치 참조 사진의 고정 건축과 다르다.",
    "fix_en": "Match the building's facade window count and arrangement to the reference. Preserve the SUV and trees.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "샷이 요구한 빽빽한 메타세콰이어 가로수 대신 양쪽 나무 간격이 넓어 중·원경 건물 벽이 크게 드러난다.",
    "fix_en": "Increase the density of the tree rows to closely match the reference. Preserve the SUV and background building.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S8sh1::cine": {
  "applied": true,
  "fingerprint": "2c1c774404b80f63c6825e308ee0ff679b919a92c27faa257374e9937caf8106",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S8sh1_sel.png",
  "source_sha256": "849195f573e0e33251f99772511ea65b08947b5a7337b60f50adacba052b07d5",
  "file": "S8sh1_cine.png",
  "latency_ms": 10875
 },
 "S9sh2::signage": {
  "fp": "a825838f13543dac",
  "inscriptions": [
   {
    "surface_native": "경찰서 건물 외벽 간판",
    "text_native": "나주경찰서",
    "reason_ko": "설정상 나주 경찰서 건물 외관임을 명확히 나타내기 위해 건물 전면에 경찰서 명칭 표기가 필요합니다."
   }
  ]
 },
 "era_assess::69fa84e42bf110ba": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 대한민국 경찰차",
    "search_terms_native": [
     "2010년대 경찰차",
     "순찰차 소나타",
     "경찰 순찰차 2000년대"
    ],
    "language_lock_native": "검색어는 오직 한국어로만 작성되어야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "해당 시기(2001~2017년) 한국 경찰차의 고유한 청색·백색·황색 도색 패턴과 경광등 스타일은 일반적인 서구식 순찰차나 현대식 도색과 확연히 다릅니다."
   },
   {
    "subject_native": "2000년대~2010년대 한국 경찰서 및 지구대 외관",
    "search_terms_native": [
     "지구대 외관 2010년대",
     "파출소 건물 2000년대",
     "경찰관서 표지판"
    ],
    "language_lock_native": "검색어는 오직 한국어로만 작성되어야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국의 지구대 및 파출소 외관, 고유의 참수리 문양 경찰 CI 마크, 특유의 전면 간판 디자인은 이 시기만의 한국적 디테일이 필요합니다."
   }
  ]
 },
 "era_ref::45ce26a729b17a09": {
  "subject": "2000년대~2010년대 대한민국 경찰차",
  "terms": [
   "2010년대 경찰차",
   "순찰차 소나타",
   "경찰 순찰차 2000년대"
  ],
  "queries": [
   [
    "2010년대 대한민국 경찰차 순찰차 쏘나타",
    "2000년대 대한민국 경찰 순찰차 쏘나타"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://images-wixmp-ed30a86b8c4ca887773594c2.wixmp.com/f/8c634ae3-bb00-46f1-bc8a-7120a3f89f8c/d9r21vk-d647f003-19c8-475a-906a-e090653b0a3c.jpg/v1/fill/w_1024%2Ch_683%2Cq_75%2Cstrp/sonata_police_vehicle_by_toyonda_d9r21vk-fullview.jpg?token=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJzdWIiOiJ1cm46YXBwOjdlMGQxODg5ODIyNjQzNzNhNWYwZDQxNWVhMGQyNmUwIiwiaXNzIjoidXJuOmFwcDo3ZTBkMTg4OTgyMjY0MzczYTVmMGQ0MTVlYTBkMjZlMCIsIm9iaiI6W1t7ImhlaWdodCI6Ijw9NjgzIiwicGF0aCI6Ii9mLzhjNjM0YWUzLWJiMDAtNDZmMS1iYzhhLTcxMjBhM2Y4OWY4Yy9kOXIyMXZrLWQ2NDdmMDAzLTE5YzgtNDc1YS05MDZhLWUwOTA2NTNiMGEzYy5qcGciLCJ3aWR0aCI6Ijw9MTAyNCJ9XV0sImF1ZCI6WyJ1cm46c2VydmljZTppbWFnZS5vcGVyYXRpb25zIl19.NY3KEFUBI-RFT42SUwgg5Ghx7vOj6jetRqMID85CbtM",
  "picked_reason_ko": "2번은 2010년대 대한민국의 일상적인 현대 쏘나타 경찰차를 선명한 사선 구도로 보여 주어 차체 비례, 도색, 표장, 경광등과 번호판을 가장 잘 확인할 수 있다.",
  "sha256": "c7537031087baffec9b555a061eae310ef8ee0ba755750edf4c159f2668e978a",
  "file": "eraref_45ce26a729b17a09.png"
 },
 "S9sh2::bgfirst_bg": {
  "input_fingerprint": "4e70f3f49e4589ea",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개를 들어 올린 채 차 밖에 서 있는 전택수의 어깨 너머로 보이는, 오래된 외관의 나주 경찰서 건물 전경\n\nLOCATION (lock): In the forecourt directly outside the police station, beside the parked vehicle and facing the building facade.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Crane to shoulder height behind and slightly beside 전택수, ending in a wide over-the-shoulder composition that places his raised shoulder and back along one lower edge. The modest police station exterior spreads across the background as he lifts his chin and studies the building, which remains the clear visual focus beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police station building; police station building in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Naju Police Station building (Shown in full exterior view) — Its modest exterior frontage faces the camera across the forecourt; used as Principal reveal beyond 전택수's shoulder, establishing his destination.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light renders the exterior in a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.\n\nPERIOD REFERENCE — 2000년대~2010년대 대한민국 경찰차: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 고개를 들어 올린 채 차 밖에 서 있는 전택수의 어깨 너머로 보이는, 오래된 외관의 나주 경찰서 건물 전경\n\nLOCATION (lock): In the forecourt directly outside the police station, beside the parked vehicle and facing the building facade.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Crane to shoulder height behind and slightly beside 전택수, ending in a wide over-the-shoulder composition that places his raised shoulder and back along one lower edge. The modest police station exterior spreads across the background as he lifts his chin and studies the building, which remains the clear visual focus beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police station building; police station building in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Naju Police Station building (Shown in full exterior view) — Its modest exterior frontage faces the camera across the forecourt; used as Principal reveal beyond 전택수's shoulder, establishing his destination.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light renders the exterior in a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.\n\nPERIOD REFERENCE — 2000년대~2010년대 대한민국 경찰차: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S9sh2__bgfirst_bg.png",
  "asset_id": "43b468ac-1b7b-4655-bec1-a157785bb40f",
  "input_asset_ids": [
   "6a5e6ca4-7dfe-428c-8bc9-dafd6e5e7054",
   "58d53119-8071-41f9-8ad5-f436ff73af3d",
   "9a057280-ea34-4e04-894f-2ae68b9a0e17"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 대한민국 경찰차",
   "queries": [
    [
     "2010년대 대한민국 경찰차 순찰차 쏘나타",
     "2000년대 대한민국 경찰 순찰차 쏘나타"
    ]
   ],
   "picked_url": "https://images-wixmp-ed30a86b8c4ca887773594c2.wixmp.com/f/8c634ae3-bb00-46f1-bc8a-7120a3f89f8c/d9r21vk-d647f003-19c8-475a-906a-e090653b0a3c.jpg/v1/fill/w_1024%2Ch_683%2Cq_75%2Cstrp/sonata_police_vehicle_by_toyonda_d9r21vk-fullview.jpg?token=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJzdWIiOiJ1cm46YXBwOjdlMGQxODg5ODIyNjQzNzNhNWYwZDQxNWVhMGQyNmUwIiwiaXNzIjoidXJuOmFwcDo3ZTBkMTg4OTgyMjY0MzczYTVmMGQ0MTVlYTBkMjZlMCIsIm9iaiI6W1t7ImhlaWdodCI6Ijw9NjgzIiwicGF0aCI6Ii9mLzhjNjM0YWUzLWJiMDAtNDZmMS1iYzhhLTcxMjBhM2Y4OWY4Yy9kOXIyMXZrLWQ2NDdmMDAzLTE5YzgtNDc1YS05MDZhLWUwOTA2NTNiMGEzYy5qcGciLCJ3aWR0aCI6Ijw9MTAyNCJ9XV0sImF1ZCI6WyJ1cm46c2VydmljZTppbWFnZS5vcGVyYXRpb25zIl19.NY3KEFUBI-RFT42SUwgg5Ghx7vOj6jetRqMID85CbtM",
   "sha256": "c7537031087baffec9b555a061eae310ef8ee0ba755750edf4c159f2668e978a",
   "file": "eraref_45ce26a729b17a09.png"
  }
 },
 "S9sh2": {
  "input_fingerprint": "2bfe69d0b701afce",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 고개를 들어 올린 채 차 밖에 서 있는 전택수의 어깨 너머로 보이는, 오래된 외관의 나주 경찰서 건물 전경\n\nLOCATION (lock): In the forecourt directly outside the police station, beside the parked vehicle and facing the building facade. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Crane to shoulder height behind and slightly beside 전택수, ending in a wide over-the-shoulder composition that places his raised shoulder and back along one lower edge. The modest police station exterior spreads across the background as he lifts his chin and studies the building, which remains the clear visual focus beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police station building; police station building in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Naju Police Station building (Shown in full exterior view) — Its modest exterior frontage faces the camera across the forecourt; used as Principal reveal beyond 전택수's shoulder, establishing his destination.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light renders the exterior in a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 경찰서 건물 외벽 간판: \"나주경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 고개를 들어 올린 채 차 밖에 서 있는 전택수의 어깨 너머로 보이는, 오래된 외관의 나주 경찰서 건물 전경\n\nLOCATION (lock): In the forecourt directly outside the police station, beside the parked vehicle and facing the building facade. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Crane to shoulder height behind and slightly beside 전택수, ending in a wide over-the-shoulder composition that places his raised shoulder and back along one lower edge. The modest police station exterior spreads across the background as he lifts his chin and studies the building, which remains the clear visual focus beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police station building; police station building in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Naju Police Station building (Shown in full exterior view) — Its modest exterior frontage faces the camera across the forecourt; used as Principal reveal beyond 전택수's shoulder, establishing his destination.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light renders the exterior in a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 경찰서 건물 외벽 간판: \"나주경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 고개를 들어 올린 채 차 밖에 서 있는 전택수의 어깨 너머로 보이는, 오래된 외관의 나주 경찰서 건물 전경\n\nLOCATION (lock): In the forecourt directly outside the police station, beside the parked vehicle and facing the building facade. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Crane to shoulder height behind and slightly beside 전택수, ending in a wide over-the-shoulder composition that places his raised shoulder and back along one lower edge. The modest police station exterior spreads across the background as he lifts his chin and studies the building, which remains the clear visual focus beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police station building; police station building in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: Naju Police Station building (Shown in full exterior view) — Its modest exterior frontage faces the camera across the forecourt; used as Principal reveal beyond 전택수's shoulder, establishing his destination.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning ambient light renders the exterior in a subdued neutral-to-cool procedural tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 경찰서 건물 외벽 간판: \"나주경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S9sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S9sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L12B01.png"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_district_police_station_sel.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 건물 구조와 장소 배경을 정확히 합성하였고 캐릭터의 복장 레퍼런스를 충실히 반영하여 프롬프트 지시를 훌륭하게 따랐습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "캐릭터의 복장이 레퍼런스와 완전히 다르며, 건물 외관 역시 지정된 구조 레퍼런스가 아닌 장소 사진의 건물을 잘못 차용했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물은 고개를 들어 배경의 경찰서 건물을 바라보고 있습니다.",
      "built_space": "건물이 프레임 중앙에 배치되었으나, 지정된 구조 레퍼런스(오래된 외관)가 아닌 장소 레퍼런스의 현대식 건물을 잘못 렌더링했습니다.",
      "entities": "인물이 갈색 가죽 재킷을 입고 있어 캐릭터 레퍼런스(경찰 점퍼와 모자)와 불일치하며, 차량에 프롬프트에 없는 추가 텍스트가 포함되었습니다.",
      "hard_violations": [],
      "physics": "인물은 지면에 안정적으로 서 있고 차량도 지면에 정상적으로 주차되어 있습니다."
     },
     {
      "label": "B",
      "direction": "인물은 정면의 건물을 향해 시선을 두고 있으나, 고개를 뚜렷하게 들어 올린 묘사는 부족합니다.",
      "built_space": "장소 레퍼런스의 가로수길 배경에 구조 레퍼런스의 오래된 건물 외관을 성공적으로 결합하여 중앙에 배치했습니다.",
      "entities": "인물은 캐릭터 레퍼런스와 일치하는 점퍼와 수사 모자를 착용했으며, 간판 텍스트('나주경찰서')도 정확히 반영되었습니다.",
      "hard_violations": [],
      "physics": "인물과 주차된 차량 모두 지면에 자연스럽게 닿아 있으며 무리 없는 물리 법칙을 보여줍니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 건물 구조와 장소 배경을 정확히 합성하였고 캐릭터의 복장 레퍼런스를 충실히 반영하여 프롬프트 지시를 훌륭하게 따랐습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "캐릭터의 복장이 레퍼런스와 완전히 다르며, 건물 외관 역시 지정된 구조 레퍼런스가 아닌 장소 사진의 건물을 잘못 차용했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물은 고개를 들어 배경의 경찰서 건물을 바라보고 있습니다.",
      "built_space": "건물이 프레임 중앙에 배치되었으나, 지정된 구조 레퍼런스(오래된 외관)가 아닌 장소 레퍼런스의 현대식 건물을 잘못 렌더링했습니다.",
      "entities": "인물이 갈색 가죽 재킷을 입고 있어 캐릭터 레퍼런스(경찰 점퍼와 모자)와 불일치하며, 차량에 프롬프트에 없는 추가 텍스트가 포함되었습니다.",
      "hard_violations": [],
      "physics": "인물은 지면에 안정적으로 서 있고 차량도 지면에 정상적으로 주차되어 있습니다."
     },
     {
      "label": "B",
      "direction": "인물은 정면의 건물을 향해 시선을 두고 있으나, 고개를 뚜렷하게 들어 올린 묘사는 부족합니다.",
      "built_space": "장소 레퍼런스의 가로수길 배경에 구조 레퍼런스의 오래된 건물 외관을 성공적으로 결합하여 중앙에 배치했습니다.",
      "entities": "인물은 캐릭터 레퍼런스와 일치하는 점퍼와 수사 모자를 착용했으며, 간판 텍스트('나주경찰서')도 정확히 반영되었습니다.",
      "hard_violations": [],
      "physics": "인물과 주차된 차량 모두 지면에 자연스럽게 닿아 있으며 무리 없는 물리 법칙을 보여줍니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "요청된 숄더뷰 구도와 '고개를 들어 올린' 액션을 정확히 구현하여 높은 우선순위를 충족했으나, 인물의 의상(갈색 재킷)이 레퍼런스와 전혀 일치하지 않아 감점되었습니다.  ★위반: [openrouter:x-ai/grok-4.6] 위치·구조 레퍼런스에 없는 순찰차와 차량 로고·영문 POLICE·번호판 등 가독 문자 창작"
     },
     {
      "label": "A",
      "score": 1667,
      "verdict_ko": "인물의 의상은 레퍼런스와 일치하지만, 카메라가 인물 바로 뒤에 위치해 '고개를 든' 핵심 액션이 보이지 않으며, 모자 뒷면에 앞면의 텍스트가 잘못 생성되는 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.429
     },
     "adjusted": {
      "A": 1.667,
      "B": 1.179
     },
     "violations": {
      "B": [
       "[openrouter:x-ai/grok-4.6] 위치·구조 레퍼런스에 없는 순찰차와 차량 로고·영문 POLICE·번호판 등 가독 문자 창작"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.571,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "요청된 숄더뷰 구도와 '고개를 들어 올린' 액션을 정확히 구현하여 높은 우선순위를 충족했으나, 인물의 의상(갈색 재킷)이 레퍼런스와 전혀 일치하지 않아 감점되었습니다.  ★위반: [openrouter:x-ai/grok-4.6] 위치·구조 레퍼런스에 없는 순찰차와 차량 로고·영문 POLICE·번호판 등 가독 문자 창작"
     },
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "인물의 의상은 레퍼런스와 일치하지만, 카메라가 인물 바로 뒤에 위치해 '고개를 든' 핵심 액션이 보이지 않으며, 모자 뒷면에 앞면의 텍스트가 잘못 생성되는 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1182,
     "B": 1674
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "인물은 고개를 들어 배경의 경찰서 건물을 바라보고 있습니다.",
    "built_space": "건물이 프레임 중앙에 배치되었으나, 지정된 구조 레퍼런스(오래된 외관)가 아닌 장소 레퍼런스의 현대식 건물을 잘못 렌더링했습니다.",
    "entities": "인물이 갈색 가죽 재킷을 입고 있어 캐릭터 레퍼런스(경찰 점퍼와 모자)와 불일치하며, 차량에 프롬프트에 없는 추가 텍스트가 포함되었습니다.",
    "hard_violations": [],
    "physics": "인물은 지면에 안정적으로 서 있고 차량도 지면에 정상적으로 주차되어 있습니다."
   },
   {
    "label": "B",
    "direction": "인물은 정면의 건물을 향해 시선을 두고 있으나, 고개를 뚜렷하게 들어 올린 묘사는 부족합니다.",
    "built_space": "장소 레퍼런스의 가로수길 배경에 구조 레퍼런스의 오래된 건물 외관을 성공적으로 결합하여 중앙에 배치했습니다.",
    "entities": "인물은 캐릭터 레퍼런스와 일치하는 점퍼와 수사 모자를 착용했으며, 간판 텍스트('나주경찰서')도 정확히 반영되었습니다.",
    "hard_violations": [],
    "physics": "인물과 주차된 차량 모두 지면에 자연스럽게 닿아 있으며 무리 없는 물리 법칙을 보여줍니다."
   }
  ],
  "totals": {
   "A": 1182,
   "B": 1674
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 건물 구조와 장소 배경을 정확히 합성하였고 캐릭터의 복장 레퍼런스를 충실히 반영하여 프롬프트 지시를 훌륭하게 따랐습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "캐릭터의 복장이 레퍼런스와 완전히 다르며, 건물 외관 역시 지정된 구조 레퍼런스가 아닌 장소 사진의 건물을 잘못 차용했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L12B01.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_district_police_station_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "경찰서 건물 외관이 지정된 STRUCTURE LOOK 레퍼런스(오래된 평면적 구조)를 따르지 않고 LOCATION 사진의 현대적인 건물 형태를 그대로 모방함 (배경 중앙).",
     "fix_en": "Replace the entire police station building in the background with the grey concrete structure from the STRUCTURE LOOK reference, matching its exact windows, proportions, and central tower. Maintain the foreground character, the parked car, the driveway, the rows of trees, and the morning lighting.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "레퍼런스 상 모자 정면에 위치해야 할 '수사' 텍스트가 인물이 쓰고 있는 모자 뒷면에 잘못 렌더링됨 (좌측 하단).",
     "fix_en": "Remove the yellow text '수사' from the back of the dark blue cap, leaving only plain dark blue fabric and the adjustment strap. Preserve the character's pose, his jacket, the parked car, and the background environment.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "프롬프트에서 '고개를 들어 올린 채(lifts his chin)'를 지시했으나 인물의 고개가 들리지 않고 평행을 유지함 (좌측 하단).",
     "fix_en": "Tilt the character's head slightly upward to show him lifting his chin, angling the back of the cap downwards toward his neck. Preserve the character's jacket, his position in the frame, the parked car, the trees, and the background building.",
     "severity": "minor",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "경찰서 건물 외관이 지정된 STRUCTURE LOOK 레퍼런스(오래된 평면적 구조)를 따르지 않고 LOCATION 사진의 현대적인 건물 형태를 그대로 모방함 (배경 중앙).",
     "severity": "critical"
    },
    {
     "issue_ko": "레퍼런스 상 모자 정면에 위치해야 할 '수사' 텍스트가 인물이 쓰고 있는 모자 뒷면에 잘못 렌더링됨 (좌측 하단).",
     "severity": "major"
    },
    {
     "issue_ko": "프롬프트에서 '고개를 들어 올린 채(lifts his chin)'를 지시했으나 인물의 고개가 들리지 않고 평행을 유지함 (좌측 하단).",
     "severity": "minor"
    },
    {
     "issue_ko": "배경 중앙의 경찰서 건물이 STRUCTURE LOOK에 잠긴 회색 콘크리트 형태·비율·창호가 아닌 흰 타일 건물이다",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "레퍼런스 상 모자 정면에 위치해야 할 '수사' 텍스트가 인물이 쓰고 있는 모자 뒷면에 잘못 렌더링됨 (좌측 하단).",
    "fix_en": "Remove the yellow text '수사' from the back of the dark blue cap, leaving only plain dark blue fabric and the adjustment strap. Preserve the character's pose, his jacket, the parked car, and the background environment.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "프롬프트에서 '고개를 들어 올린 채(lifts his chin)'를 지시했으나 인물의 고개가 들리지 않고 평행을 유지함 (좌측 하단).",
    "fix_en": "Tilt the character's head slightly upward to show him lifting his chin, angling the back of the cap downwards toward his neck. Preserve the character's jacket, his position in the frame, the parked car, the trees, and the background building.",
    "severity": "minor",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the entire police station building in the background with the grey concrete structure from the STRUCTURE LOOK reference, matching its exact windows, proportions, and central tower. Maintain the foreground character, the parked car, the driveway, the rows of trees, and the morning lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 오버더숄더 앵글, 좌측 하단 배치, 건물을 향한 시선과 주차된 차량 등 프레이밍 지시를 충실히 따랐으며 간판 텍스트도 명확히 구현됨. (단, 모자 앞면의 글자가 뒷면에 렌더링된 점은 사소한 오류임)"
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "캐릭터의 정면 외형과 복장은 참조와 일치하나, 필수적인 오버더숄더 구도와 건물 쪽 시선 지시를 완전히 무시하고 정면을 응시하게 렌더링하여 프레이밍에서 크게 실패함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수는 화면 안쪽의 경찰서 건물을 향해 시선을 두고 있으며, 턱을 살짝 든 상태임.",
      "built_space": "경찰서 진입로와 가로수, 주차된 차량이 위치하며, 카메라 시점에 맞게 배경 건물이 원경에 자리 잡음.",
      "entities": "화면 좌측 하단의 전택수(뒷모습, 남색 재킷), 좌측에 주차된 검은색 차량, '나주경찰서' 간판이 있는 배경 건물.",
      "hard_violations": [],
      "physics": "전택수는 지면에 안정적으로 서 있으며, 차량들 역시 도로 위에 정상적으로 안착해 있음."
     },
     {
      "label": "B",
      "direction": "전택수는 정면의 카메라를 똑바로 응시하고 있음.",
      "built_space": "가로수가 늘어선 진입로와 배경의 건물이 배치되어 있으나 주변 주차 차량이 없음.",
      "entities": "화면 중앙의 전택수(얼굴 및 복장은 참조와 일치), 텍스트가 누락된 배경 경찰서 건물.",
      "hard_violations": [],
      "physics": "전택수는 진입로 지면 위에 두 발로 안정적으로 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 오버더숄더 앵글, 좌측 하단 배치, 건물을 향한 시선과 주차된 차량 등 프레이밍 지시를 충실히 따랐으며 간판 텍스트도 명확히 구현됨. (단, 모자 앞면의 글자가 뒷면에 렌더링된 점은 사소한 오류임)"
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "캐릭터의 정면 외형과 복장은 참조와 일치하나, 필수적인 오버더숄더 구도와 건물 쪽 시선 지시를 완전히 무시하고 정면을 응시하게 렌더링하여 프레이밍에서 크게 실패함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수는 화면 안쪽의 경찰서 건물을 향해 시선을 두고 있으며, 턱을 살짝 든 상태임.",
      "built_space": "경찰서 진입로와 가로수, 주차된 차량이 위치하며, 카메라 시점에 맞게 배경 건물이 원경에 자리 잡음.",
      "entities": "화면 좌측 하단의 전택수(뒷모습, 남색 재킷), 좌측에 주차된 검은색 차량, '나주경찰서' 간판이 있는 배경 건물.",
      "hard_violations": [],
      "physics": "전택수는 지면에 안정적으로 서 있으며, 차량들 역시 도로 위에 정상적으로 안착해 있음."
     },
     {
      "label": "B",
      "direction": "전택수는 정면의 카메라를 똑바로 응시하고 있음.",
      "built_space": "가로수가 늘어선 진입로와 배경의 건물이 배치되어 있으나 주변 주차 차량이 없음.",
      "entities": "화면 중앙의 전택수(얼굴 및 복장은 참조와 일치), 텍스트가 누락된 배경 경찰서 건물.",
      "hard_violations": [],
      "physics": "전택수는 진입로 지면 위에 두 발로 안정적으로 서 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 오버더숄더 카메라 구도, 인물의 배치 및 시선, 지정된 간판 텍스트까지 프롬프트의 지시를 충실히 구현하여 훌륭한 결과물을 만들었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물의 뒷모습과 어깨 너머로 건물을 바라보는 오버더숄더 구도 지시를 완전히 무시하고 정면을 묘사하여 핵심 연출을 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "인물이 카메라를 등지고 경찰서 건물 정면을 향해 시선을 두고 있습니다.",
      "built_space": "경찰서 앞마당과 나무, 진입로 및 인물 옆과 건물 앞 주차된 차량이 기준 이미지와 일치하게 배치되어 있습니다.",
      "entities": "전택수의 뒷모습(지정된 재킷 착용)이 보이며, 건물 외벽에 '나주경찰서'라는 텍스트가 정확히 렌더링되었습니다.",
      "hard_violations": [],
      "physics": "인물이 바닥을 안정적으로 딛고 서 있습니다."
     },
     {
      "label": "A",
      "direction": "인물이 목표물(건물)을 보지 않고 카메라 렌즈를 정면으로 응시하고 있습니다.",
      "built_space": "경찰서 앞마당과 건물 외관이 묘사되었으나 인물 옆에 있어야 할 주차된 차량이 생략되었습니다.",
      "entities": "전택수의 외형과 복장은 기준과 일치하나, 건물에 요구된 '나주경찰서' 간판이 없습니다.",
      "hard_violations": [
       "지정된 카메라 구도 및 인물 연출(오버더숄더 샷, 등을 보이며 건물을 향한 시선, 좌측 하단 배치)을 완전히 위반하고 정면 전신 샷을 렌더링함."
      ],
      "physics": "인물이 바닥을 안정적으로 딛고 서 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 오버더숄더 카메라 구도, 인물의 배치 및 시선, 지정된 간판 텍스트까지 프롬프트의 지시를 충실히 구현하여 훌륭한 결과물을 만들었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 뒷모습과 어깨 너머로 건물을 바라보는 오버더숄더 구도 지시를 완전히 무시하고 정면을 묘사하여 핵심 연출을 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물이 카메라를 등지고 경찰서 건물 정면을 향해 시선을 두고 있습니다.",
      "built_space": "경찰서 앞마당과 나무, 진입로 및 인물 옆과 건물 앞 주차된 차량이 기준 이미지와 일치하게 배치되어 있습니다.",
      "entities": "전택수의 뒷모습(지정된 재킷 착용)이 보이며, 건물 외벽에 '나주경찰서'라는 텍스트가 정확히 렌더링되었습니다.",
      "hard_violations": [],
      "physics": "인물이 바닥을 안정적으로 딛고 서 있습니다."
     },
     {
      "label": "B",
      "direction": "인물이 목표물(건물)을 보지 않고 카메라 렌즈를 정면으로 응시하고 있습니다.",
      "built_space": "경찰서 앞마당과 건물 외관이 묘사되었으나 인물 옆에 있어야 할 주차된 차량이 생략되었습니다.",
      "entities": "전택수의 외형과 복장은 기준과 일치하나, 건물에 요구된 '나주경찰서' 간판이 없습니다.",
      "hard_violations": [
       "지정된 카메라 구도 및 인물 연출(오버더숄더 샷, 등을 보이며 건물을 향한 시선, 좌측 하단 배치)을 완전히 위반하고 정면 전신 샷을 렌더링함."
      ],
      "physics": "인물이 바닥을 안정적으로 딛고 서 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S9sh2__bgfirst_bg.png",
   "bg_asset_id": "43b468ac-1b7b-4655-bec1-a157785bb40f",
   "bg_record_key": "S9sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S9sh2::cine": {
  "applied": true,
  "fingerprint": "a5444ecdd05abccd802125c609601dee25514c1f0a0fd9ff90430dceac4e669e",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S9sh2_sel.png",
  "source_sha256": "eb70dfa6dfd97e1ef5834be4f50535b8cf4b0355f95f2baff2f3e2dba4c6c338",
  "file": "S9sh2_cine.png",
  "latency_ms": 11126
 },
 "S10sh4::signage": {
  "fp": "fa27ec8caf1ab41f",
  "inscriptions": [
   {
    "surface_native": "사무실 문 위 부서 안내판",
    "text_native": "강력1팀",
    "reason_ko": "사건이 일어나는 공간이 경찰서 강력계 사무실 내부임을 사실적으로 묘사하기 위해 문에 부서 안내판 표기가 필요합니다."
   }
  ]
 },
 "era_assess::53e1b89fde57b475": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 경찰서 강력반 복도 및 출입문",
    "search_terms_native": [
     "경찰서 강력반 복도",
     "경찰서 강력계 문",
     "2000년대 경찰서 내부",
     "경찰서 강력계 표지판"
    ],
    "language_lock_native": "이 검색은 오직 한국어로만 수행되어야 하며, 다른 언어로 번역하거나 다른 언어의 검색어를 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 경찰서 강력반 특유의 투박한 철제 문, 푸른색 계열의 부서 안내 아크릴 표지판, 벽면의 홍보 포스터 등 고유의 시각적 요소를 반영하지 못하면 서구식 경찰서나 일반 사무실처럼 그려질 위험이 큽니다."
   }
  ]
 },
 "era_ref::f168a728fc17769b": {
  "subject": "2000년대~2010년대 한국 경찰서 강력반 복도 및 출입문",
  "terms": [
   "경찰서 강력반 복도",
   "경찰서 강력계 문",
   "2000년대 경찰서 내부",
   "경찰서 강력계 표지판"
  ],
  "queries": [
   [
    "2000년대 한국 경찰서 강력반 복도 강력계 출입문 표지판",
    "2010년대 한국 경찰서 내부 강력계 문 복도"
   ],
   [
    "\"강력계\" \"경찰서\" \"복도\"",
    "\"강력반\" \"경찰서 내부\"",
    "\"강력계 사무실\" 경찰서 문",
    "\"강력팀\" 경찰서 복도"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://ojsfile.ohmynews.com/STD_IMG_FILE/2024/0305/IE003268468_STD.jpg",
  "picked_reason_ko": "강력팀 표지, 유리 출입문, 내부 복도와 사무공간의 재료·비례·설비가 선명하게 보여 2001~2017년 한국 경찰서 강력반의 가장 직접적이고 읽기 좋은 참고 사진이다.",
  "sha256": "cfd4da335b0552622f3083e449dcf5c1b2805192d1829a65a88da050144e2436",
  "file": "eraref_f168a728fc17769b.png"
 },
 "S10sh4::bgfirst_bg": {
  "input_fingerprint": "eb9587afc56f1ddb",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 사무실 안에서 다른 형사가 문을 반쯤 열고 선 배경 앞, 중년 여성이 서의용의 가슴팍을 두 손으로 세게 밀치는 mid-impact 순간\n\nLOCATION (lock): In the police-station corridor directly outside the violent-crimes office doorway.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At chest height beside the office doorway, a brief handheld tightening catches the exact impact from an oblique three-quarter axis rather than between the opponents. The 50대 중년 여성 drives both palms into 서의용's chest at center-left as his weight breaks backward to the right, while 다른 형사 braces the half-open door in the background and watches the struggle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: violent-crimes office door (Half open) — The door is seen obliquely from the hallway, with its opening exposing part of the office beyond; used as Frames the background detective and preserves the woman's route back into the office; police station hallway (No additional state specified); used as Provides the lateral space into which 서의용 begins to lose balance.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior, held in a restrained neutral-to-cool palette with documentary immediacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 경찰서 강력반 복도 및 출입문: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 사무실 안에서 다른 형사가 문을 반쯤 열고 선 배경 앞, 중년 여성이 서의용의 가슴팍을 두 손으로 세게 밀치는 mid-impact 순간\n\nLOCATION (lock): In the police-station corridor directly outside the violent-crimes office doorway.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At chest height beside the office doorway, a brief handheld tightening catches the exact impact from an oblique three-quarter axis rather than between the opponents. The 50대 중년 여성 drives both palms into 서의용's chest at center-left as his weight breaks backward to the right, while 다른 형사 braces the half-open door in the background and watches the struggle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: violent-crimes office door (Half open) — The door is seen obliquely from the hallway, with its opening exposing part of the office beyond; used as Frames the background detective and preserves the woman's route back into the office; police station hallway (No additional state specified); used as Provides the lateral space into which 서의용 begins to lose balance.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior, held in a restrained neutral-to-cool palette with documentary immediacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 경찰서 강력반 복도 및 출입문: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S10sh4__bgfirst_bg.png",
  "asset_id": "7a70f7ca-9234-4df3-bb88-991603c7d506",
  "input_asset_ids": [
   "3d3f6d63-5314-4070-b731-fbc39951971f",
   "1d7f70b8-abe2-48d7-ab2b-c99a90a28e26"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국 경찰서 강력반 복도 및 출입문",
   "queries": [
    [
     "2000년대 한국 경찰서 강력반 복도 강력계 출입문 표지판",
     "2010년대 한국 경찰서 내부 강력계 문 복도"
    ],
    [
     "\"강력계\" \"경찰서\" \"복도\"",
     "\"강력반\" \"경찰서 내부\"",
     "\"강력계 사무실\" 경찰서 문",
     "\"강력팀\" 경찰서 복도"
    ]
   ],
   "picked_url": "https://ojsfile.ohmynews.com/STD_IMG_FILE/2024/0305/IE003268468_STD.jpg",
   "sha256": "cfd4da335b0552622f3083e449dcf5c1b2805192d1829a65a88da050144e2436",
   "file": "eraref_f168a728fc17769b.png"
  }
 },
 "S10sh4": {
  "input_fingerprint": "509130351dfe1b3a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사무실 안에서 다른 형사가 문을 반쯤 열고 선 배경 앞, 중년 여성이 서의용의 가슴팍을 두 손으로 세게 밀치는 mid-impact 순간\n\nLOCATION (lock): In the police-station corridor directly outside the violent-crimes office doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside the office doorway, a brief handheld tightening catches the exact impact from an oblique three-quarter axis rather than between the opponents. The 50대 중년 여성 drives both palms into 서의용's chest at center-left as his weight breaks backward to the right, while 다른 형사 braces the half-open door in the background and watches the struggle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: violent-crimes office door (Half open) — The door is seen obliquely from the hallway, with its opening exposing part of the office beyond; used as Frames the background detective and preserves the woman's route back into the office; police station hallway (No additional state specified); used as Provides the lateral space into which 서의용 begins to lose balance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior, held in a restrained neutral-to-cool palette with documentary immediacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 50대 중년 여성 (Korean 여성, 50대, 넓은 얼굴, 검은색 짧은 머리) — wearing: 눈에 띄는 형광색 아웃도어 등산복 상의와 활동적인 검은색 등산바지; 50대 남성 경찰 (Korean 남성, 50대, 각진 얼굴, 짧게 정돈된 검은색 머리) — wearing: 일선 경찰관이 착용하는 연한 청회색 셔츠와 짙은 남색 바지로 구성된 기본 근무복 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 문 위 부서 안내판: \"강력1팀\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사무실 안에서 다른 형사가 문을 반쯤 열고 선 배경 앞, 중년 여성이 서의용의 가슴팍을 두 손으로 세게 밀치는 mid-impact 순간\n\nLOCATION (lock): In the police-station corridor directly outside the violent-crimes office doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside the office doorway, a brief handheld tightening catches the exact impact from an oblique three-quarter axis rather than between the opponents. The 50대 중년 여성 drives both palms into 서의용's chest at center-left as his weight breaks backward to the right, while 다른 형사 braces the half-open door in the background and watches the struggle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: violent-crimes office door (Half open) — The door is seen obliquely from the hallway, with its opening exposing part of the office beyond; used as Frames the background detective and preserves the woman's route back into the office; police station hallway (No additional state specified); used as Provides the lateral space into which 서의용 begins to lose balance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior, held in a restrained neutral-to-cool palette with documentary immediacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 50대 중년 여성 (Korean 여성, 50대, 넓은 얼굴, 검은색 짧은 머리) — wearing: 눈에 띄는 형광색 아웃도어 등산복 상의와 활동적인 검은색 등산바지; 50대 남성 경찰 (Korean 남성, 50대, 각진 얼굴, 짧게 정돈된 검은색 머리) — wearing: 일선 경찰관이 착용하는 연한 청회색 셔츠와 짙은 남색 바지로 구성된 기본 근무복 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 문 위 부서 안내판: \"강력1팀\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사무실 안에서 다른 형사가 문을 반쯤 열고 선 배경 앞, 중년 여성이 서의용의 가슴팍을 두 손으로 세게 밀치는 mid-impact 순간\n\nLOCATION (lock): In the police-station corridor directly outside the violent-crimes office doorway. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside the office doorway, a brief handheld tightening catches the exact impact from an oblique three-quarter axis rather than between the opponents. The 50대 중년 여성 drives both palms into 서의용's chest at center-left as his weight breaks backward to the right, while 다른 형사 braces the half-open door in the background and watches the struggle.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: violent-crimes office door (Half open) — The door is seen obliquely from the hallway, with its opening exposing part of the office beyond; used as Frames the background detective and preserves the woman's route back into the office; police station hallway (No additional state specified); used as Provides the lateral space into which 서의용 begins to lose balance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior, held in a restrained neutral-to-cool palette with documentary immediacy.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 50대 중년 여성 (Korean 여성, 50대, 넓은 얼굴, 검은색 짧은 머리) — wearing: 눈에 띄는 형광색 아웃도어 등산복 상의와 활동적인 검은색 등산바지; 50대 남성 경찰 (Korean 남성, 50대, 각진 얼굴, 짧게 정돈된 검은색 머리) — wearing: 일선 경찰관이 착용하는 연한 청회색 셔츠와 짙은 남색 바지로 구성된 기본 근무복 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 문 위 부서 안내판: \"강력1팀\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S10sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S10sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L13B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "레퍼런스 장소의 세부 요소(좌측 문의 텍스트와 디자인)를 완벽하게 재현하고 인물의 복장과 구도를 잘 살렸으나, 여성이 손바닥으로 미는 대신 옷깃을 움켜쥐고 있는 점이 감점 요인입니다."
     },
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "인물 배치와 복장은 지시사항을 따랐으나, 장소 레퍼런스의 디테일이 부족하고 여성의 손 형태가 다소 뭉개져 묘사되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.857,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.857,
      "B": 1.714
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.143,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "레퍼런스 장소의 세부 요소(좌측 문의 텍스트와 디자인)를 완벽하게 재현하고 인물의 복장과 구도를 잘 살렸으나, 여성이 손바닥으로 미는 대신 옷깃을 움켜쥐고 있는 점이 감점 요인입니다."
     },
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "인물 배치와 복장은 지시사항을 따랐으나, 장소 레퍼런스의 디테일이 부족하고 여성의 손 형태가 다소 뭉개져 묘사되었습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "손바닥으로 가슴을 미는 대신 옷깃을 움켜쥐는 자세로 연출된 점은 아쉬우나, 기준 사진의 복도 공간과 특정 간판들을 완벽에 가깝게 재현하여 우선순위(공간 일치)에서 크게 앞섭니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "프롬프트의 손바닥으로 미는 동작을 반영하지 못하고 옷을 움켜쥐었으며, 기준 사진의 배경(좌측 문 로고 및 복도 끝 계단 구조)을 정확히 구현하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "중년 여성은 서의용을 바라보며 옷깃을 움켜쥐고 있고, 서의용도 여성을 마주 본다. 배경의 형사는 두 사람의 실랑이를 주시한다.",
      "built_space": "경찰서 복도. 우측 사무실 문이 복도 바깥쪽으로 열려 있으며, 좌측 벽면과 배경의 복도 구조(계단, 창문 등)가 기준 사진의 공간과 다르게 렌더링되었다.",
      "entities": "중년 여성(형광색 상의), 서의용(기준 사진 복장 일치), 다른 형사(근무복) 모두 지정된 인물 특성에 부합한다. 문 위 간판에 '강력1팀'이 적혀 있다.",
      "hard_violations": [],
      "physics": "여성이 두 발을 바닥에 지지하고 있으나 손바닥으로 미는 대신 재킷을 움켜쥐어 당기고 있으며, 서의용은 뒤로 밀려 체중이 쏠려 있다."
     },
     {
      "label": "B",
      "direction": "중년 여성은 서의용의 옷을 움켜쥔 채 그를 응시하고, 서의용은 놀란 표정으로 여성을 본다. 배경의 형사는 열린 문틈으로 상황을 지켜본다.",
      "built_space": "기준 사진의 공간을 훌륭하게 재현함. 좌측의 '수사 책임자실' 문과 로고, 안쪽으로 열린 우측 사무실 문, 그리고 배경의 계단과 창문 구조가 정확하다.",
      "entities": "중년 여성(형광색 상의), 서의용(기준 사진 일치), 다른 형사(근무복 착용)가 적절히 묘사되었다. 문 위 간판 텍스트가 정확하게 렌더링되었다.",
      "hard_violations": [],
      "physics": "여성이 서의용을 손바닥으로 밀지 않고 재킷을 꽉 움켜쥐고 있으며, 서의용은 균형을 잃고 뒤로 물러나는 자세다. 모든 인물이 바닥과 구조물에 안정적으로 지지되어 있다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "손바닥으로 가슴을 미는 대신 옷깃을 움켜쥐는 자세로 연출된 점은 아쉬우나, 기준 사진의 복도 공간과 특정 간판들을 완벽에 가깝게 재현하여 우선순위(공간 일치)에서 크게 앞섭니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "프롬프트의 손바닥으로 미는 동작을 반영하지 못하고 옷을 움켜쥐었으며, 기준 사진의 배경(좌측 문 로고 및 복도 끝 계단 구조)을 정확히 구현하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "중년 여성은 서의용을 바라보며 옷깃을 움켜쥐고 있고, 서의용도 여성을 마주 본다. 배경의 형사는 두 사람의 실랑이를 주시한다.",
      "built_space": "경찰서 복도. 우측 사무실 문이 복도 바깥쪽으로 열려 있으며, 좌측 벽면과 배경의 복도 구조(계단, 창문 등)가 기준 사진의 공간과 다르게 렌더링되었다.",
      "entities": "중년 여성(형광색 상의), 서의용(기준 사진 복장 일치), 다른 형사(근무복) 모두 지정된 인물 특성에 부합한다. 문 위 간판에 '강력1팀'이 적혀 있다.",
      "hard_violations": [],
      "physics": "여성이 두 발을 바닥에 지지하고 있으나 손바닥으로 미는 대신 재킷을 움켜쥐어 당기고 있으며, 서의용은 뒤로 밀려 체중이 쏠려 있다."
     },
     {
      "label": "A",
      "direction": "중년 여성은 서의용의 옷을 움켜쥔 채 그를 응시하고, 서의용은 놀란 표정으로 여성을 본다. 배경의 형사는 열린 문틈으로 상황을 지켜본다.",
      "built_space": "기준 사진의 공간을 훌륭하게 재현함. 좌측의 '수사 책임자실' 문과 로고, 안쪽으로 열린 우측 사무실 문, 그리고 배경의 계단과 창문 구조가 정확하다.",
      "entities": "중년 여성(형광색 상의), 서의용(기준 사진 일치), 다른 형사(근무복 착용)가 적절히 묘사되었다. 문 위 간판 텍스트가 정확하게 렌더링되었다.",
      "hard_violations": [],
      "physics": "여성이 서의용을 손바닥으로 밀지 않고 재킷을 꽉 움켜쥐고 있으며, 서의용은 균형을 잃고 뒤로 물러나는 자세다. 모든 인물이 바닥과 구조물에 안정적으로 지지되어 있다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1864,
     "B": 1719
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 1864,
   "B": 1719
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "레퍼런스 장소의 세부 요소(좌측 문의 텍스트와 디자인)를 완벽하게 재현하고 인물의 복장과 구도를 잘 살렸으나, 여성이 손바닥으로 미는 대신 옷깃을 움켜쥐고 있는 점이 감점 요인입니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "인물 배치와 복장은 지시사항을 따랐으나, 장소 레퍼런스의 디테일이 부족하고 여성의 손 형태가 다소 뭉개져 묘사되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L13B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우측 문을 잡고 있는 경찰관의 손 아래에 여분의 손이 하나 더 문 가장자리를 잡고 있는 해부학적 오류가 있음.",
     "fix_en": "Remove the extra hand gripping the door edge immediately below the police officer's left hand, replacing it with the plain gray surface of the door edge. Preserve the officer's real hand on the handle, all three characters, their poses and clothing, the lighting, and the corridor setting exactly.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "여성이 손바닥으로 남성의 가슴을 밀쳐야(drives both palms) 한다는 지시와 달리, 양손을 주먹 쥐고 남성의 재킷을 움켜쥐고 있음.",
     "fix_en": "Change the woman's hands from closed fists gripping the jacket to open hands with both palms pressed flat against the man's chest to push him. Preserve the characters' identities, their overall poses, their clothing, the lighting, and the background exactly.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "배경을 원본과 정확히 동일하게 유지하라는 지시에도 불구하고, 우측 사무실 문이 원본의 유리문에서 작은 창이 달린 불투명한 철문으로 변형됨.",
     "fix_en": "Change the solid gray door on the right to a full-length framed glass door matching the reference background, showing the office interior through the glass. Preserve the police officer's position and pose holding the door, all other characters, their clothing, the lighting, and the rest of the background exactly.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "허공을 향해 있는 서의용의 왼손 손가락들이 서로 뭉개져 있어 형태가 부자연스러움.",
     "fix_en": "Redraw the man's extended left hand so the fingers are anatomically distinct and properly shaped, removing the melted appearance. Preserve the hand's placement, all characters, their poses, their clothing, the lighting, and the background exactly.",
     "severity": "minor",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우측 문을 잡고 있는 경찰관의 손 아래에 여분의 손이 하나 더 문 가장자리를 잡고 있는 해부학적 오류가 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "여성이 손바닥으로 남성의 가슴을 밀쳐야(drives both palms) 한다는 지시와 달리, 양손을 주먹 쥐고 남성의 재킷을 움켜쥐고 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "배경을 원본과 정확히 동일하게 유지하라는 지시에도 불구하고, 우측 사무실 문이 원본의 유리문에서 작은 창이 달린 불투명한 철문으로 변형됨.",
     "severity": "major"
    },
    {
     "issue_ko": "허공을 향해 있는 서의용의 왼손 손가락들이 서로 뭉개져 있어 형태가 부자연스러움.",
     "severity": "minor"
    },
    {
     "issue_ko": "중년 여성이 서의용 가슴에 두 손바닥을 밀어붙이는 것이 아니라 주먹을 쥐고 재킷 앞섶을 움켜쥐고 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "여성이 손바닥으로 남성의 가슴을 밀쳐야(drives both palms) 한다는 지시와 달리, 양손을 주먹 쥐고 남성의 재킷을 움켜쥐고 있음.",
    "fix_en": "Change the woman's hands from closed fists gripping the jacket to open hands with both palms pressed flat against the man's chest to push him. Preserve the characters' identities, their overall poses, their clothing, the lighting, and the background exactly.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "배경을 원본과 정확히 동일하게 유지하라는 지시에도 불구하고, 우측 사무실 문이 원본의 유리문에서 작은 창이 달린 불투명한 철문으로 변형됨.",
    "fix_en": "Change the solid gray door on the right to a full-length framed glass door matching the reference background, showing the office interior through the glass. Preserve the police officer's position and pose holding the door, all other characters, their clothing, the lighting, and the rest of the background exactly.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "허공을 향해 있는 서의용의 왼손 손가락들이 서로 뭉개져 있어 형태가 부자연스러움.",
    "fix_en": "Redraw the man's extended left hand so the fingers are anatomically distinct and properly shaped, removing the melted appearance. Preserve the hand's placement, all characters, their poses, their clothing, the lighting, and the background exactly.",
    "severity": "minor",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the extra hand gripping the door edge immediately below the police officer's left hand, replacing it with the plain gray surface of the door edge. Preserve the officer's real hand on the handle, all three characters, their poses and clothing, the lighting, and the corridor setting exactly.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 샷 크기, 인물들의 역할과 의상, 배경과 표지판을 비교적 잘 구현했으나, 여성이 손바닥으로 가슴을 미는 대신 옷깃을 쥐고 있으며 서의용의 얼굴이 레퍼런스와 다소 차이가 납니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "서의용이 밀쳐지는 대신 문간에 서 있고 지시문에 없는 새로운 남성이 밀쳐지고 있으며, 여성의 의상과 프레이밍(풀샷) 모두 지시사항을 심각하게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "여성은 밀쳐지는 남성을 향해 힘을 가하며 바라보고, 배경의 경찰관은 두 사람의 실랑이를 주시하고 있습니다.",
      "built_space": "경찰서 복도와 반쯤 열린 사무실 문의 구조가 스케치와 일치하며, 문 상단에 '강력1팀' 표지판이 정확히 배치되었습니다.",
      "entities": "여성은 형광색 등산복과 검은 바지를 입었고, 남성은 레퍼런스의 가죽 재킷과 비니를 착용했으나 얼굴이 정확히 일치하지 않습니다. 경찰관은 지정된 근무복을 입고 있습니다.",
      "hard_violations": [],
      "physics": "여성의 체중이 앞으로 실려 있고 남성은 뒤로 중심을 잃으며 무너지는 자세가 중력에 맞게 자연스럽게 표현되었습니다."
     },
     {
      "label": "B",
      "direction": "여성은 벽에 밀어붙인 남성을 향하고 있으며, 문간에 선 남성(서의용)이 이를 바라보고 있습니다.",
      "built_space": "복도와 문의 위치는 존재하나, 인물들이 스케치의 배치와 달리 벽면 쪽에 치우쳐 있습니다.",
      "entities": "여성이 정장을 입고 있으며, 서의용 레퍼런스 인물이 경찰관 대신 문간에 서 있습니다. 지시문에 없는 파란 셔츠의 남성이 등장합니다.",
      "hard_violations": [
       "지시문에 없는 새로운 인물 등장",
       "프롬프트에 명시된 경찰관 누락",
       "지정된 인물의 역할 및 위치가 완전히 뒤바뀜"
      ],
      "physics": "인물들은 바닥에 발을 딛고 서 있으며 자세 자체는 물리적으로 가능하나 프롬프트의 상황과는 다릅니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 샷 크기, 인물들의 역할과 의상, 배경과 표지판을 비교적 잘 구현했으나, 여성이 손바닥으로 가슴을 미는 대신 옷깃을 쥐고 있으며 서의용의 얼굴이 레퍼런스와 다소 차이가 납니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "서의용이 밀쳐지는 대신 문간에 서 있고 지시문에 없는 새로운 남성이 밀쳐지고 있으며, 여성의 의상과 프레이밍(풀샷) 모두 지시사항을 심각하게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "여성은 밀쳐지는 남성을 향해 힘을 가하며 바라보고, 배경의 경찰관은 두 사람의 실랑이를 주시하고 있습니다.",
      "built_space": "경찰서 복도와 반쯤 열린 사무실 문의 구조가 스케치와 일치하며, 문 상단에 '강력1팀' 표지판이 정확히 배치되었습니다.",
      "entities": "여성은 형광색 등산복과 검은 바지를 입었고, 남성은 레퍼런스의 가죽 재킷과 비니를 착용했으나 얼굴이 정확히 일치하지 않습니다. 경찰관은 지정된 근무복을 입고 있습니다.",
      "hard_violations": [],
      "physics": "여성의 체중이 앞으로 실려 있고 남성은 뒤로 중심을 잃으며 무너지는 자세가 중력에 맞게 자연스럽게 표현되었습니다."
     },
     {
      "label": "B",
      "direction": "여성은 벽에 밀어붙인 남성을 향하고 있으며, 문간에 선 남성(서의용)이 이를 바라보고 있습니다.",
      "built_space": "복도와 문의 위치는 존재하나, 인물들이 스케치의 배치와 달리 벽면 쪽에 치우쳐 있습니다.",
      "entities": "여성이 정장을 입고 있으며, 서의용 레퍼런스 인물이 경찰관 대신 문간에 서 있습니다. 지시문에 없는 파란 셔츠의 남성이 등장합니다.",
      "hard_violations": [
       "지시문에 없는 새로운 인물 등장",
       "프롬프트에 명시된 경찰관 누락",
       "지정된 인물의 역할 및 위치가 완전히 뒤바뀜"
      ],
      "physics": "인물들은 바닥에 발을 딛고 서 있으며 자세 자체는 물리적으로 가능하나 프롬프트의 상황과는 다릅니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "스케치에 지정된 미디엄 샷 프레이밍을 정확히 구현했으며, 서의용의 신원, 여성의 형광색 등산복, 배경 경찰관의 근무복과 간판 텍스트(강력1팀)까지 프롬프트의 요구사항을 충실히 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "서의용과 다른 형사의 역할 및 위치가 완전히 뒤바뀌었고, 여성의 복장이 정장으로 잘못 렌더링되었으며, 지정된 미디엄 샷 구도 역시 무시되어 프롬프트 준수도가 매우 낮습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "여성이 서의용의 가슴팍 옷깃을 쥐고 밀치며, 서의용은 오른쪽으로 중심을 잃고 밀려남. 뒤쪽 경찰관은 반쯤 열린 문을 잡고 시선을 두 사람에게 향함.",
      "built_space": "배경 이미지의 복도와 문 구조를 유지하면서 레이아웃 스케치의 미디엄 샷 크기와 인물 배치를 정확히 따름. 문 위 간판에 '강력1팀'이 올바르게 렌더링됨.",
      "entities": "여성은 형광색 등산복 상의와 검은색 바지를 입음(일치). 서의용은 참조 이미지의 인상착의(비니, 가죽자켓 등)와 완벽히 일치. 배경 남성은 청회색 셔츠의 경찰 근무복 착용(일치).",
      "hard_violations": [],
      "physics": "여성은 두 다리로 단단히 바닥을 디딘 채 체중을 실어 밀고 있으며, 서의용은 타격으로 인해 뒤로 젖혀지는 물리적 균형 상실 상태를 자연스럽게 보여줌. 경찰관은 손으로 문을 지지함."
     },
     {
      "label": "A",
      "direction": "여성이 벽에 기대어 있는 엉뚱한 남성(파란 셔츠)을 밀치고 있으며, 서의용은 문손잡이를 잡은 채 이들을 방관함.",
      "built_space": "프롬프트와 스케치가 요구한 미디엄 샷이 아닌 훨씬 넓은 와이드 샷으로 렌더링됨. 간판 텍스트가 지정된 내용과 다름.",
      "entities": "인물의 위치와 역할이 엇갈려 서의용이 배경에 서 있음. 여성은 지정된 아웃도어 복장이 아닌 회색 정장을 착용. 밀쳐지는 남성은 지정되지 않은 외모와 복장임.",
      "hard_violations": [
       "지정된 인물 위치 및 역할 완전 뒤바뀜 (서의용이 뒷배경 문 앞에 위치)",
       "프레이밍 지시 위반 (스케치의 미디엄 샷을 무시한 와이드 샷)",
       "여성 복장 위반 (형광색 등산복이 아닌 정장 착용)"
      ],
      "physics": "여성은 바닥을 딛고 서 있으며, 남성은 벽에 기대어 지지받고 있음. 문 앞의 인물은 손잡이를 잡고 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "스케치에 지정된 미디엄 샷 프레이밍을 정확히 구현했으며, 서의용의 신원, 여성의 형광색 등산복, 배경 경찰관의 근무복과 간판 텍스트(강력1팀)까지 프롬프트의 요구사항을 충실히 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "서의용과 다른 형사의 역할 및 위치가 완전히 뒤바뀌었고, 여성의 복장이 정장으로 잘못 렌더링되었으며, 지정된 미디엄 샷 구도 역시 무시되어 프롬프트 준수도가 매우 낮습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "여성이 서의용의 가슴팍 옷깃을 쥐고 밀치며, 서의용은 오른쪽으로 중심을 잃고 밀려남. 뒤쪽 경찰관은 반쯤 열린 문을 잡고 시선을 두 사람에게 향함.",
      "built_space": "배경 이미지의 복도와 문 구조를 유지하면서 레이아웃 스케치의 미디엄 샷 크기와 인물 배치를 정확히 따름. 문 위 간판에 '강력1팀'이 올바르게 렌더링됨.",
      "entities": "여성은 형광색 등산복 상의와 검은색 바지를 입음(일치). 서의용은 참조 이미지의 인상착의(비니, 가죽자켓 등)와 완벽히 일치. 배경 남성은 청회색 셔츠의 경찰 근무복 착용(일치).",
      "hard_violations": [],
      "physics": "여성은 두 다리로 단단히 바닥을 디딘 채 체중을 실어 밀고 있으며, 서의용은 타격으로 인해 뒤로 젖혀지는 물리적 균형 상실 상태를 자연스럽게 보여줌. 경찰관은 손으로 문을 지지함."
     },
     {
      "label": "B",
      "direction": "여성이 벽에 기대어 있는 엉뚱한 남성(파란 셔츠)을 밀치고 있으며, 서의용은 문손잡이를 잡은 채 이들을 방관함.",
      "built_space": "프롬프트와 스케치가 요구한 미디엄 샷이 아닌 훨씬 넓은 와이드 샷으로 렌더링됨. 간판 텍스트가 지정된 내용과 다름.",
      "entities": "인물의 위치와 역할이 엇갈려 서의용이 배경에 서 있음. 여성은 지정된 아웃도어 복장이 아닌 회색 정장을 착용. 밀쳐지는 남성은 지정되지 않은 외모와 복장임.",
      "hard_violations": [
       "지정된 인물 위치 및 역할 완전 뒤바뀜 (서의용이 뒷배경 문 앞에 위치)",
       "프레이밍 지시 위반 (스케치의 미디엄 샷을 무시한 와이드 샷)",
       "여성 복장 위반 (형광색 등산복이 아닌 정장 착용)"
      ],
      "physics": "여성은 바닥을 딛고 서 있으며, 남성은 벽에 기대어 지지받고 있음. 문 앞의 인물은 손잡이를 잡고 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S10sh4__bgfirst_bg.png",
   "bg_asset_id": "7a70f7ca-9234-4df3-bb88-991603c7d506",
   "bg_record_key": "S10sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S10sh4::cine": {
  "applied": true,
  "fingerprint": "aabae7d7e5799a04893df6783aa1ea48e0d46621d71f1c5b13c10ddfe1f15244",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S10sh4_sel.png",
  "source_sha256": "91418e4a86ab681414e17d84ecfc59595c17515131ec85b364aa5e09ea3a28e0",
  "file": "S10sh4_cine.png",
  "latency_ms": 11497
 },
 "S10sh8::signage": {
  "fp": "d707a19656701a5d",
  "inscriptions": [
   {
    "surface_native": "문 팻말",
    "text_native": "수사과장실",
    "reason_ko": "수사과장실 문 앞이라는 공간의 정체성을 시각적으로 명확히 전달하기 위해 문에 붙은 표지판의 명칭이 필요합니다."
   }
  ]
 },
 "S10sh8": {
  "input_fingerprint": "8d3d7c269f7be826",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 저 끝 수사과장실 팻말이 붙은 문 앞에 서서 문고리를 쥐고 돌린 상태로 고정된 전택수의 뒷모습\n\nLOCATION (lock): At the far end of the police-station corridor, immediately outside the locked investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height on the hallway axis, let the dolly-in pause in a fixed wide view far behind 전택수 as the handle reaches its locked stop. His back remains centered in the distant midground, one hand fixed on the handle, while the office door and its sign close the corridor's perspective line beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground; investigation chief's office door in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation chief's office door (Closed and locked) — Its corridor-facing side is visible behind 전택수; used as Terminal anchor of the hallway composition behind 전택수; door handle (Stopped in the locked position) — The handle projects toward the hallway and is seen beneath his gripping hand; used as Shows where 전택수's hand has met resistance; investigation chief's office sign (Attached to the door) — Its labeled face is visible to the camera on the office door; used as Identifies the room at the far end of the corridor; police station corridor (No additional state specified); used as Creates the long central axis separating 전택수 from the earlier disturbance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior maintains a subdued neutral-to-cool procedural mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 문 팻말: \"수사과장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 저 끝 수사과장실 팻말이 붙은 문 앞에 서서 문고리를 쥐고 돌린 상태로 고정된 전택수의 뒷모습\n\nLOCATION (lock): At the far end of the police-station corridor, immediately outside the locked investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height on the hallway axis, let the dolly-in pause in a fixed wide view far behind 전택수 as the handle reaches its locked stop. His back remains centered in the distant midground, one hand fixed on the handle, while the office door and its sign close the corridor's perspective line beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground; investigation chief's office door in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation chief's office door (Closed and locked) — Its corridor-facing side is visible behind 전택수; used as Terminal anchor of the hallway composition behind 전택수; door handle (Stopped in the locked position) — The handle projects toward the hallway and is seen beneath his gripping hand; used as Shows where 전택수's hand has met resistance; investigation chief's office sign (Attached to the door) — Its labeled face is visible to the camera on the office door; used as Identifies the room at the far end of the corridor; police station corridor (No additional state specified); used as Creates the long central axis separating 전택수 from the earlier disturbance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior maintains a subdued neutral-to-cool procedural mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 문 팻말: \"수사과장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 저 끝 수사과장실 팻말이 붙은 문 앞에 서서 문고리를 쥐고 돌린 상태로 고정된 전택수의 뒷모습\n\nLOCATION (lock): At the far end of the police-station corridor, immediately outside the locked investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height on the hallway axis, let the dolly-in pause in a fixed wide view far behind 전택수 as the handle reaches its locked stop. His back remains centered in the distant midground, one hand fixed on the handle, while the office door and its sign close the corridor's perspective line beyond him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground; investigation chief's office door in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation chief's office door (Closed and locked) — Its corridor-facing side is visible behind 전택수; used as Terminal anchor of the hallway composition behind 전택수; door handle (Stopped in the locked position) — The handle projects toward the hallway and is seen beneath his gripping hand; used as Shows where 전택수's hand has met resistance; investigation chief's office sign (Attached to the door) — Its labeled face is visible to the camera on the office door; used as Identifies the room at the far end of the corridor; police station corridor (No additional state specified); used as Creates the long central axis separating 전택수 from the earlier disturbance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime interior maintains a subdued neutral-to-cool procedural mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 문 팻말: \"수사과장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "인물은 복도 우측 측면에 있는 열린 문을 향해 서 있으며, 오른손으로 문고리를 잡고 있음.",
    "built_space": "복도 우측 벽에 문이 열려 있고, 복도의 원근선은 인물 너머 먼 배경까지 길게 이어짐.",
    "entities": "남색 재킷과 카키색 바지를 입은 전택수의 뒷모습. 열린 문에 '수사과장실' 팻말이 보임.",
    "hard_violations": [
     "복도 원근선을 닫는 정면 끝의 문이어야 한다는 공간 구도 지시 위반",
     "문이 닫혀 있고 잠겨 있어야 한다는(Closed and locked) 지시 위반(문이 열려 있음)"
    ],
    "physics": "두 발로 바닥에 안정적으로 서 있으며, 오른손으로 손잡이를 잡고 있음."
   },
   {
    "label": "B",
    "direction": "인물은 복도 끝 정면에 있는 닫힌 문을 향해 서 있으며, 오른손으로 문고리를 쥐고 있음.",
    "built_space": "복도의 원근선이 정면에 위치한 단일 문으로 완전히 닫히는 구조.",
    "entities": "남색 재킷, 카키색 바지, 무전기 등을 착용한 전택수의 뒷모습. 문 정면에 '수사과장실' 팻말이 정확히 표시됨.",
    "hard_violations": [],
    "physics": "두 발로 바닥을 딛고 서 있으며, 오른손으로 문고리를 단단히 쥐고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "복도 끝에 위치한 닫힌 문이라는 구도와 설정, '수사과장실' 팻말, 인물의 복장과 포즈를 프롬프트에 맞게 정확히 구현했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "문이 닫혀 있어야 한다는 지시를 어기고 문을 열어두었으며, 복도 끝이 아닌 측면에 문을 배치해 핵심 구도를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S10sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "오른손이 문고리를 쥐고 돌리는 대신 문 우측 가장자리의 평면을 짚고 있음.",
     "fix_en": "Redraw the character's right hand and the adjacent door surface to show a physical lever-style door handle projecting from the door, with the hand gripping and turning the lever downwards; preserve the character's body, clothing, the '수사과장실' sign, the rest of the door, and the corridor.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 장면 스틸과 달리 복도 왼쪽 벽의 창문들이 모두 사라지고 막힌 벽으로 묘사됨.",
     "fix_en": "Replace the solid left wall with continuous glass windows to match the previous shot still's architecture; maintain the character's pose, the door, and the right wall.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "캐릭터 레퍼런스의 재킷에 부착되어 있는 후드(모자)가 묘사되지 않음.",
     "fix_en": "Add the dark navy hood to the back of the character's jacket collar; preserve the character's head, the rest of the jacket, and the background.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "이전 장면 스틸의 복도(원거리 유리문·측면 창·기존 팻말)와 달리 회색 수사과장실 문이 복도를 막는 다른 구조다",
     "fix_en": "Rebuild the corridor layout to feature the continuing hallway and distant double glass doors shown in the previous shot still; maintain the character and his interaction with the door.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전택수가 먼 중경이 아니라 프레임 높이를 크게 차지해 원거리 와이드가 아니다",
     "fix_en": "Pull the camera back to create a distant wide shot where the character sits much smaller in the midground; preserve the lighting, the action, and the set elements.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "오른손이 문고리를 쥐고 돌리는 대신 문 우측 가장자리의 평면을 짚고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "인물의 오른손이 닿아 있는 위치에 쥐어야 할 문고리 자체가 존재하지 않음.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 장면 스틸과 달리 복도 왼쪽 벽의 창문들이 모두 사라지고 막힌 벽으로 묘사됨.",
     "severity": "major"
    },
    {
     "issue_ko": "캐릭터 레퍼런스의 재킷에 부착되어 있는 후드(모자)가 묘사되지 않음.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 장면 스틸의 복도(원거리 유리문·측면 창·기존 팻말)와 달리 회색 수사과장실 문이 복도를 막는 다른 구조다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 먼 중경이 아니라 프레임 높이를 크게 차지해 원거리 와이드가 아니다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 장면 스틸과 달리 복도 왼쪽 벽의 창문들이 모두 사라지고 막힌 벽으로 묘사됨.",
    "fix_en": "Replace the solid left wall with continuous glass windows to match the previous shot still's architecture; maintain the character's pose, the door, and the right wall.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "캐릭터 레퍼런스의 재킷에 부착되어 있는 후드(모자)가 묘사되지 않음.",
    "fix_en": "Add the dark navy hood to the back of the character's jacket collar; preserve the character's head, the rest of the jacket, and the background.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "이전 장면 스틸의 복도(원거리 유리문·측면 창·기존 팻말)와 달리 회색 수사과장실 문이 복도를 막는 다른 구조다",
    "fix_en": "Rebuild the corridor layout to feature the continuing hallway and distant double glass doors shown in the previous shot still; maintain the character and his interaction with the door.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "전택수가 먼 중경이 아니라 프레임 높이를 크게 차지해 원거리 와이드가 아니다",
    "fix_en": "Pull the camera back to create a distant wide shot where the character sits much smaller in the midground; preserve the lighting, the action, and the set elements.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the character's right hand and the adjacent door surface to show a physical lever-style door handle projecting from the door, with the hand gripping and turning the lever downwards; preserve the character's body, clothing, the '수사과장실' sign, the rest of the door, and the corridor.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "레퍼런스의 복도 구조를 반영하지 못해 로케이션 설정에 실패했으나, 지시된 가로형 레버 손잡이를 쥐고 있는 모습을 A보다 정확히 표현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "레퍼런스의 복도 구조를 반영하지 못했으며, 돌릴 수 없는 수직형 손잡이가 생성되어 프롬프트의 핵심 행동 지시를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "남자는 복도 끝의 문을 향해 서 있으며, 오른손은 문의 수직형 손잡이를 향하고 있습니다.",
      "built_space": "끝에 문이 있는 대칭형의 좁은 복도입니다. 레퍼런스 이미지에 나타난 넓고 비대칭적인 복도(왼쪽 창문 등) 구조와 전혀 일치하지 않습니다.",
      "entities": "지정된 복장을 입은 전택수의 뒷모습입니다. 문에는 '수사과장실' 팻말이 있으나, 손잡이는 돌릴 수 없는 수직형 바(bar) 형태입니다.",
      "hard_violations": [
       "레퍼런스와 전혀 다른 복도 구조 및 배경 (로케이션 락 위반)"
      ],
      "physics": "두 발로 바닥에 서 있으며, 오른손은 고정된 수직 손잡이를 잡고 있습니다."
     },
     {
      "label": "B",
      "direction": "남자는 복도 끝의 문을 향해 서 있으며, 오른손은 가로형 레버 손잡이를 쥐고 있습니다.",
      "built_space": "끝에 문이 있는 대칭형의 좁은 복도입니다. 레퍼런스 이미지에 나타난 넓고 비대칭적인 복도(왼쪽 창문 등) 구조와 전혀 일치하지 않습니다.",
      "entities": "지정된 복장을 입은 전택수의 뒷모습입니다. 문에 '수사과장실' 팻말이 부착되어 있고, 손잡이는 가로형 레버입니다.",
      "hard_violations": [
       "레퍼런스와 전혀 다른 복도 구조 및 배경 (로케이션 락 위반)"
      ],
      "physics": "두 발로 바닥에 서 있으며, 오른손으로 가로형 레버를 쥐고 누르는 형태를 취하고 있습니다."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "레퍼런스의 복도 구조를 반영하지 못해 로케이션 설정에 실패했으나, 지시된 가로형 레버 손잡이를 쥐고 있는 모습을 A보다 정확히 표현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "레퍼런스의 복도 구조를 반영하지 못했으며, 돌릴 수 없는 수직형 손잡이가 생성되어 프롬프트의 핵심 행동 지시를 위반했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "남자는 복도 끝의 문을 향해 서 있으며, 오른손은 문의 수직형 손잡이를 향하고 있습니다.",
      "built_space": "끝에 문이 있는 대칭형의 좁은 복도입니다. 레퍼런스 이미지에 나타난 넓고 비대칭적인 복도(왼쪽 창문 등) 구조와 전혀 일치하지 않습니다.",
      "entities": "지정된 복장을 입은 전택수의 뒷모습입니다. 문에는 '수사과장실' 팻말이 있으나, 손잡이는 돌릴 수 없는 수직형 바(bar) 형태입니다.",
      "hard_violations": [
       "레퍼런스와 전혀 다른 복도 구조 및 배경 (로케이션 락 위반)"
      ],
      "physics": "두 발로 바닥에 서 있으며, 오른손은 고정된 수직 손잡이를 잡고 있습니다."
     },
     {
      "label": "B",
      "direction": "남자는 복도 끝의 문을 향해 서 있으며, 오른손은 가로형 레버 손잡이를 쥐고 있습니다.",
      "built_space": "끝에 문이 있는 대칭형의 좁은 복도입니다. 레퍼런스 이미지에 나타난 넓고 비대칭적인 복도(왼쪽 창문 등) 구조와 전혀 일치하지 않습니다.",
      "entities": "지정된 복장을 입은 전택수의 뒷모습입니다. 문에 '수사과장실' 팻말이 부착되어 있고, 손잡이는 가로형 레버입니다.",
      "hard_violations": [
       "레퍼런스와 전혀 다른 복도 구조 및 배경 (로케이션 락 위반)"
      ],
      "physics": "두 발로 바닥에 서 있으며, 오른손으로 가로형 레버를 쥐고 누르는 형태를 취하고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1833,
      "verdict_ko": "문고리를 아래로 돌려 잡고 있는 모습을 정확히 구현하여 '돌린 상태로 고정된'이라는 세부 지시를 훌륭하게 충족했습니다."
     },
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "문고리가 수평을 유지하고 있어 프롬프트가 요구한 '돌린 상태'의 동작 지시를 반영하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 1.833
     },
     "adjusted": {
      "A": 1.714,
      "B": 1.833
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.167,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1833,
      "verdict_ko": "문고리를 아래로 돌려 잡고 있는 모습을 정확히 구현하여 '돌린 상태로 고정된'이라는 세부 지시를 훌륭하게 충족했습니다."
     },
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "문고리가 수평을 유지하고 있어 프롬프트가 요구한 '돌린 상태'의 동작 지시를 반영하지 못했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1837,
     "B": 1719
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh4"
  }
 },
 "S10sh8::cine": {
  "applied": true,
  "fingerprint": "736ea5ac15e17912980b7a9babc7162b7a0d93790a7585e97ada9fb79f6087d7",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S10sh8_sel.png",
  "source_sha256": "38ab2843fe0b646f92483c621a6dcb927f0df0c9732efe1322cac4c220467da9",
  "file": "S10sh8_cine.png",
  "latency_ms": 11946
 },
 "S10sh13::signage": {
  "fp": "173a97f05097c391",
  "inscriptions": [
   {
    "surface_native": "계단 벽면의 액자",
    "text_native": "정성을 다하는 국민의 경찰",
    "reason_ko": "나주경찰서 내부 계단 벽면에 당시 한국 경찰서에서 흔히 볼 수 있었던 공식 슬로건 액자를 배치하여 공간의 사실감을 높입니다."
   }
  ]
 },
 "S10sh13": {
  "input_fingerprint": "9e85a3f1f1b3386f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 계단 중턱에 멈춰 선 채 전택수를 향해 눈을 크게 뜨고 반가운 기색을 보이는 나주경찰서장\n\nLOCATION (lock): On the interior stairway landing adjoining the police-station corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 전택수's side at upper-chest height, finish the upward tilt in a medium, slightly low-angle view toward the middle of the staircase. 전택수's shoulder and partial profile remain at the lower-left foreground while 나주경찰서장 pauses off-axis on the stairs at upper right, eyes widened as he recognizes 전택수.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police chief on the staircase; 나주경찰서장 in the upper-right of the frame, midground, looks toward 전택수 in the hallway; staircase in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: police station staircase (No additional state specified) — Its steps rise diagonally behind and beneath 나주경찰서장; used as Elevates 나주경찰서장 above the hallway and provides the diagonal reveal line; hallway below the stairs (No additional state specified); used as Keeps 전택수 spatially connected to the staircase reveal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior ambience keeps the recognition beat restrained, neutral-to-cool, and moderately contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 계단 벽면의 액자: \"정성을 다하는 국민의 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 계단 중턱에 멈춰 선 채 전택수를 향해 눈을 크게 뜨고 반가운 기색을 보이는 나주경찰서장\n\nLOCATION (lock): On the interior stairway landing adjoining the police-station corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 전택수's side at upper-chest height, finish the upward tilt in a medium, slightly low-angle view toward the middle of the staircase. 전택수's shoulder and partial profile remain at the lower-left foreground while 나주경찰서장 pauses off-axis on the stairs at upper right, eyes widened as he recognizes 전택수.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police chief on the staircase; 나주경찰서장 in the upper-right of the frame, midground, looks toward 전택수 in the hallway; staircase in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: police station staircase (No additional state specified) — Its steps rise diagonally behind and beneath 나주경찰서장; used as Elevates 나주경찰서장 above the hallway and provides the diagonal reveal line; hallway below the stairs (No additional state specified); used as Keeps 전택수 spatially connected to the staircase reveal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior ambience keeps the recognition beat restrained, neutral-to-cool, and moderately contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 계단 벽면의 액자: \"정성을 다하는 국민의 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 계단 중턱에 멈춰 선 채 전택수를 향해 눈을 크게 뜨고 반가운 기색을 보이는 나주경찰서장\n\nLOCATION (lock): On the interior stairway landing adjoining the police-station corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 전택수's side at upper-chest height, finish the upward tilt in a medium, slightly low-angle view toward the middle of the staircase. 전택수's shoulder and partial profile remain at the lower-left foreground while 나주경찰서장 pauses off-axis on the stairs at upper right, eyes widened as he recognizes 전택수.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the lower-left of the frame, foreground, looks toward police chief on the staircase; 나주경찰서장 in the upper-right of the frame, midground, looks toward 전택수 in the hallway; staircase in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: police station staircase (No additional state specified) — Its steps rise diagonally behind and beneath 나주경찰서장; used as Elevates 나주경찰서장 above the hallway and provides the diagonal reveal line; hallway below the stairs (No additional state specified); used as Keeps 전택수 spatially connected to the staircase reveal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior ambience keeps the recognition beat restrained, neutral-to-cool, and moderately contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 계단 벽면의 액자: \"정성을 다하는 국민의 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.714,
    "B": 0.929
   },
   "violations": {
    "B": [
     "[gemini-pro] 레퍼런스에 없는 인물 소품(가슴 배지) 임의 추가",
     "[gemini-pro] 지시되지 않은 표지판 및 읽을 수 있는 텍스트 누출(우측 문 옆 간판)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 929
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "카메라 구도와 캐릭터 레퍼런스를 완벽히 준수했으며, 지정된 표지판 텍스트를 오류나 임의의 추가 요소 없이 정확하게 구현하여 가장 충실합니다."
   },
   {
    "label": "B",
    "score": 929,
    "verdict_ko": "레퍼런스에 없는 가슴 배지를 임의로 추가하고, 지시되지 않은 의미 불명의 텍스트 간판을 생성하여 주요 규칙을 위반했습니다.  ★위반: [gemini-pro] 레퍼런스에 없는 인물 소품(가슴 배지) 임의 추가 / [gemini-pro] 지시되지 않은 표지판 및 읽을 수 있는 텍스트 누출(우측 문 옆 간판)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S10sh8_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832602>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 '눈을 크게 뜨고'를 요구했으나, 화면 우측 계단 위의 경찰서장은 웃음을 지으며 눈이 가늘게 접혀 있습니다.",
     "fix_en": "Redraw the police chief's face to open his eyes wide in surprise instead of squinting, keeping his identity, uniform, the foreground character, and the background staircase exactly unchanged.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 샷 스틸의 복도(회색 문·수사과장실 표지·벽 안내판)와 고정 장소가 전혀 다른 계단실임",
     "fix_en": "Add a grey door matching the previous corridor to the lower wall section visible at the bottom of the stairs, keeping the characters, their current poses, and the main staircase unchanged.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "나주경찰서장이 계단 중턱에 멈춰 서지 않고 한 발을 아래 칸에 딛고 내려오는 자세임",
     "fix_en": "Redraw the police chief's lower body so both feet stand together firmly on the same step to appear paused, preserving his identity, uniform, the foreground character, and the staircase.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "왼쪽 전경의 전택수가 어깨와 부분 옆모습이 아니라 뒷통수로만 보임",
     "fix_en": "Redraw the foreground character's head turned slightly right to reveal a partial side profile instead of just the back of the head, keeping his placement, clothing, the police chief, and the background unchanged.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "나주경찰서장이 우측 상단 미디엄이 아니라 화면 중앙~우측에 전신으로 잡혀 있음",
     "fix_en": "Increase the scale of the police chief's upper body to fill the right side of the frame as a medium shot, keeping his identity, the foreground character, and the background unchanged.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 '눈을 크게 뜨고'를 요구했으나, 화면 우측 계단 위의 경찰서장은 웃음을 지으며 눈이 가늘게 접혀 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷 스틸의 복도(회색 문·수사과장실 표지·벽 안내판)와 고정 장소가 전혀 다른 계단실임",
     "severity": "major"
    },
    {
     "issue_ko": "나주경찰서장이 계단 중턱에 멈춰 서지 않고 한 발을 아래 칸에 딛고 내려오는 자세임",
     "severity": "major"
    },
    {
     "issue_ko": "나주경찰서장의 눈이 크게 떠지지 않고 웃으며 가늘어져 있음",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 전경의 전택수가 어깨와 부분 옆모습이 아니라 뒷통수로만 보임",
     "severity": "major"
    },
    {
     "issue_ko": "나주경찰서장이 우측 상단 미디엄이 아니라 화면 중앙~우측에 전신으로 잡혀 있음",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 '눈을 크게 뜨고'를 요구했으나, 화면 우측 계단 위의 경찰서장은 웃음을 지으며 눈이 가늘게 접혀 있습니다.",
    "fix_en": "Redraw the police chief's face to open his eyes wide in surprise instead of squinting, keeping his identity, uniform, the foreground character, and the background staircase exactly unchanged.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "이전 샷 스틸의 복도(회색 문·수사과장실 표지·벽 안내판)와 고정 장소가 전혀 다른 계단실임",
    "fix_en": "Add a grey door matching the previous corridor to the lower wall section visible at the bottom of the stairs, keeping the characters, their current poses, and the main staircase unchanged.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "나주경찰서장이 계단 중턱에 멈춰 서지 않고 한 발을 아래 칸에 딛고 내려오는 자세임",
    "fix_en": "Redraw the police chief's lower body so both feet stand together firmly on the same step to appear paused, preserving his identity, uniform, the foreground character, and the staircase.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "왼쪽 전경의 전택수가 어깨와 부분 옆모습이 아니라 뒷통수로만 보임",
    "fix_en": "Redraw the foreground character's head turned slightly right to reveal a partial side profile instead of just the back of the head, keeping his placement, clothing, the police chief, and the background unchanged.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "나주경찰서장이 우측 상단 미디엄이 아니라 화면 중앙~우측에 전신으로 잡혀 있음",
    "fix_en": "Increase the scale of the police chief's upper body to fill the right side of the frame as a medium shot, keeping his identity, the foreground character, and the background unchanged.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh8"
  }
 },
 "S10sh13::cine": {
  "applied": true,
  "fingerprint": "71f45a4e19253c7aa54cff0242b32b59f9d5ab29aa737338535ef3adbb0589b2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S10sh13_sel.png",
  "source_sha256": "be2142a08190b198593988270c80887e4cd5d8da43ce064a6026206633363ea2",
  "file": "S10sh13_cine.png",
  "latency_ms": 12051
 },
 "S11sh1::signage": {
  "fp": "d7c1ec39b3117ef3",
  "inscriptions": [
   {
    "surface_native": "흡연구역 표지판",
    "text_native": "흡연구역",
    "reason_ko": "경찰서 야외 휴게소에서 인물이 담배를 물고 있는 장면이므로 해당 공간이 흡연 구역임을 명시하는 표지판이 필요합니다."
   }
  ]
 },
 "era_assess::d6a4a9d65c241ef0": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 한국 경찰서 및 지구대 야외 휴게 공간과 벤치",
    "search_terms_native": [
     "지구대 야외 쉼터",
     "파출소 외관 벤치",
     "경찰서 야외 휴게실",
     "경찰서 자전거 보관소 벤치"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며, 영어 등 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "2001년에서 2017년 사이 한국 경찰서(지구대/파출소) 특유의 외관 색상, 경찰 로고(CI), 야외 철제/목재 벤치 및 재떨이 등의 배치 형태는 일반적인 서구식 경찰서나 현대식 시설과 확연히 다릅니다."
   }
  ]
 },
 "era_ref::425e33e65b63d8d2": {
  "subject": "2000년대-2010년대 한국 경찰서 및 지구대 야외 휴게 공간과 벤치",
  "terms": [
   "지구대 야외 쉼터",
   "파출소 외관 벤치",
   "경찰서 야외 휴게실",
   "경찰서 자전거 보관소 벤치"
  ],
  "queries": [
   [
    "2000년대 2010년대 지구대 야외 쉼터 파출소 외관 벤치",
    "2000년대 2010년대 경찰서 야외 휴게실 자전거 보관소 벤치"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://img0.yna.co.kr/photo/yna/YH/2011/06/17/PYH2011061706360034300_P4.jpg",
  "picked_reason_ko": "1번은 2000~2010년대 한국 지구대 앞에 조성된 화단과 목재 벤치형 야외 휴게 공간을 가장 선명하고 일상적인 모습으로 보여준다.",
  "sha256": "b9209faf4562a042af4ebe3779ccc73a77321b1886010ec4b1c1c10014dca403",
  "file": "eraref_425e33e65b63d8d2.png"
 },
 "S11sh1::bgfirst_bg": {
  "input_fingerprint": "3a84dc33bcdc430b",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 휴게 공간 의자에 앉아 입에 담배를 문 채 앞을 주시하는 서의용의 상체\n\nLOCATION (lock): In the outdoor staff rest area in front of the police station, beside its seating.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated chest height and slightly below 서의용's eye line, hold a medium-close static view from roughly thirty degrees off his front. His upper body occupies one side of the frame as he settles back in the chair with the cigarette held in his mouth, eyes fixed ahead; the near shoulder is left available for the following lateral reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rest-area chair (Occupied by 서의용) — The chair is seen obliquely beneath and behind his torso; used as Supports 서의용's seated posture while preserving the open space behind his shoulder; cigarette (Being smoked) — It projects outward from one side of 서의용's mouth; used as Small facial action detail held clearly against his mouth.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime rest area, rendered with restrained neutral-to-cool color and light documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국 경찰서 및 지구대 야외 휴게 공간과 벤치: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 휴게 공간 의자에 앉아 입에 담배를 문 채 앞을 주시하는 서의용의 상체\n\nLOCATION (lock): In the outdoor staff rest area in front of the police station, beside its seating.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated chest height and slightly below 서의용's eye line, hold a medium-close static view from roughly thirty degrees off his front. His upper body occupies one side of the frame as he settles back in the chair with the cigarette held in his mouth, eyes fixed ahead; the near shoulder is left available for the following lateral reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rest-area chair (Occupied by 서의용) — The chair is seen obliquely beneath and behind his torso; used as Supports 서의용's seated posture while preserving the open space behind his shoulder; cigarette (Being smoked) — It projects outward from one side of 서의용's mouth; used as Small facial action detail held clearly against his mouth.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime rest area, rendered with restrained neutral-to-cool color and light documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국 경찰서 및 지구대 야외 휴게 공간과 벤치: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S11sh1__bgfirst_bg.png",
  "asset_id": "fa5a6a76-663e-449b-b974-a79c1facbbc9",
  "input_asset_ids": [
   "d6924610-b609-4f79-80ab-ca8c8a6cb904",
   "df2bebe5-d663-41c0-b18e-5579baf9483f"
  ],
  "era_research": {
   "subject": "2000년대-2010년대 한국 경찰서 및 지구대 야외 휴게 공간과 벤치",
   "queries": [
    [
     "2000년대 2010년대 지구대 야외 쉼터 파출소 외관 벤치",
     "2000년대 2010년대 경찰서 야외 휴게실 자전거 보관소 벤치"
    ]
   ],
   "picked_url": "https://img0.yna.co.kr/photo/yna/YH/2011/06/17/PYH2011061706360034300_P4.jpg",
   "sha256": "b9209faf4562a042af4ebe3779ccc73a77321b1886010ec4b1c1c10014dca403",
   "file": "eraref_425e33e65b63d8d2.png"
  }
 },
 "S11sh1": {
  "input_fingerprint": "5eafa3a8b6a474b2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 휴게 공간 의자에 앉아 입에 담배를 문 채 앞을 주시하는 서의용의 상체\n\nLOCATION (lock): In the outdoor staff rest area in front of the police station, beside its seating. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height and slightly below 서의용's eye line, hold a medium-close static view from roughly thirty degrees off his front. His upper body occupies one side of the frame as he settles back in the chair with the cigarette held in his mouth, eyes fixed ahead; the near shoulder is left available for the following lateral reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rest-area chair (Occupied by 서의용) — The chair is seen obliquely beneath and behind his torso; used as Supports 서의용's seated posture while preserving the open space behind his shoulder; cigarette (Being smoked) — It projects outward from one side of 서의용's mouth; used as Small facial action detail held clearly against his mouth.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime rest area, rendered with restrained neutral-to-cool color and light documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 흡연구역 표지판: \"흡연구역\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 휴게 공간 의자에 앉아 입에 담배를 문 채 앞을 주시하는 서의용의 상체\n\nLOCATION (lock): In the outdoor staff rest area in front of the police station, beside its seating. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height and slightly below 서의용's eye line, hold a medium-close static view from roughly thirty degrees off his front. His upper body occupies one side of the frame as he settles back in the chair with the cigarette held in his mouth, eyes fixed ahead; the near shoulder is left available for the following lateral reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rest-area chair (Occupied by 서의용) — The chair is seen obliquely beneath and behind his torso; used as Supports 서의용's seated posture while preserving the open space behind his shoulder; cigarette (Being smoked) — It projects outward from one side of 서의용's mouth; used as Small facial action detail held clearly against his mouth.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime rest area, rendered with restrained neutral-to-cool color and light documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 흡연구역 표지판: \"흡연구역\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 휴게 공간 의자에 앉아 입에 담배를 문 채 앞을 주시하는 서의용의 상체\n\nLOCATION (lock): In the outdoor staff rest area in front of the police station, beside its seating. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height and slightly below 서의용's eye line, hold a medium-close static view from roughly thirty degrees off his front. His upper body occupies one side of the frame as he settles back in the chair with the cigarette held in his mouth, eyes fixed ahead; the near shoulder is left available for the following lateral reveal.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rest-area chair (Occupied by 서의용) — The chair is seen obliquely beneath and behind his torso; used as Supports 서의용's seated posture while preserving the open space behind his shoulder; cigarette (Being smoked) — It projects outward from one side of 서의용's mouth; used as Small facial action detail held clearly against his mouth.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime rest area, rendered with restrained neutral-to-cool color and light documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 흡연구역 표지판: \"흡연구역\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S11sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S11sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L14B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1417,
      "verdict_ko": "카메라 앵글과 여백 활용은 지시를 따랐으나, 로케이션의 평벤치에 등받이를 임의로 추가하는 구조적 위반이 발생함.  ★위반: [gemini-pro] 참조된 로케이션의 등받이가 없는 평벤치를 등받이가 있는 의자로 임의 변경함"
     },
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "로케이션 벤치 형태와 피사체는 정확히 구현했으나, 지시된 30도 측면 앵글이 아닌 정면 구도로 렌즈를 응시하는 연출 오류가 있음."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.667
     },
     "adjusted": {
      "A": 1.417,
      "B": 1.667
     },
     "violations": {
      "A": [
       "[gemini-pro] 참조된 로케이션의 등받이가 없는 평벤치를 등받이가 있는 의자로 임의 변경함"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.333,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1417,
      "verdict_ko": "카메라 앵글과 여백 활용은 지시를 따랐으나, 로케이션의 평벤치에 등받이를 임의로 추가하는 구조적 위반이 발생함.  ★위반: [gemini-pro] 참조된 로케이션의 등받이가 없는 평벤치를 등받이가 있는 의자로 임의 변경함"
     },
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "로케이션 벤치 형태와 피사체는 정확히 구현했으나, 지시된 30도 측면 앵글이 아닌 정면 구도로 렌즈를 응시하는 연출 오류가 있음."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "프레이밍이 지정된 측면 배치가 아닌 중앙에 위치한 점은 아쉬우나, 장소의 벤치 형태와 표지판, 인물 디테일을 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "인물을 한쪽에 배치한 프레이밍은 좋으나, 지정된 장소의 등받이 없는 벤치 구조를 변형한 치명적인 오류가 있음.  ★위반: [gemini-pro] 존재하지 않는 사물 생성 (위치 레퍼런스의 등받이 없는 벤치에 등받이와 팔걸이를 임의로 추가함)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 1.429
     },
     "adjusted": {
      "A": 1.714,
      "B": 1.179
     },
     "violations": {
      "B": [
       "[gemini-pro] 존재하지 않는 사물 생성 (위치 레퍼런스의 등받이 없는 벤치에 등받이와 팔걸이를 임의로 추가함)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "프레이밍이 지정된 측면 배치가 아닌 중앙에 위치한 점은 아쉬우나, 장소의 벤치 형태와 표지판, 인물 디테일을 정확하게 구현함."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "인물을 한쪽에 배치한 프레이밍은 좋으나, 지정된 장소의 등받이 없는 벤치 구조를 변형한 치명적인 오류가 있음.  ★위반: [gemini-pro] 존재하지 않는 사물 생성 (위치 레퍼런스의 등받이 없는 벤치에 등받이와 팔걸이를 임의로 추가함)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2596,
     "B": 3381
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 2596,
   "B": 3381
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "카메라 앵글과 여백 활용은 지시를 따랐으나, 로케이션의 평벤치에 등받이를 임의로 추가하는 구조적 위반이 발생함.  ★위반: [gemini-pro] 참조된 로케이션의 등받이가 없는 평벤치를 등받이가 있는 의자로 임의 변경함"
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "로케이션 벤치 형태와 피사체는 정확히 구현했으나, 지시된 30도 측면 앵글이 아닌 정면 구도로 렌즈를 응시하는 연출 오류가 있음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L14B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "인물이 입에 문 담배의 방향이 거꾸로 뒤집혀 오렌지색 필터 부분이 바깥을 향하고 있음.",
     "fix_en": "Reverse the orientation of the cigarette in the man's mouth so that the orange filter is held inside his lips and the white paper cylinder extends outward. Maintain the man's exact face, expression, clothing, posture, and the entire background and lighting unchanged.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "프롬프트는 앞을 주시하는 자연스러운 순간을 지시했으나, 인물이 카메라 렌즈를 직접 응시하고 있음.",
     "fix_en": "Shift the man's pupils so his gaze is directed off-camera, looking ahead into the distance rather than staring into the lens. Preserve his face, posture, clothing, the cigarette, and the background.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "인물이 입에 문 담배의 방향이 거꾸로 뒤집혀 오렌지색 필터 부분이 바깥을 향하고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "프롬프트는 앞을 주시하는 자연스러운 순간을 지시했으나, 인물이 카메라 렌즈를 직접 응시하고 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 서의용이 렌즈를 정면으로 응시하고 몸이 카메라에 거의 정면이라 앞을 주시하는 순간이 아니라 포즈 초상처럼 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 앞을 주시하는 자연스러운 순간을 지시했으나, 인물이 카메라 렌즈를 직접 응시하고 있음.",
    "fix_en": "Shift the man's pupils so his gaze is directed off-camera, looking ahead into the distance rather than staring into the lens. Preserve his face, posture, clothing, the cigarette, and the background.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Reverse the orientation of the cigarette in the man's mouth so that the orange filter is held inside his lips and the white paper cylinder extends outward. Maintain the man's exact face, expression, clothing, posture, and the entire background and lighting unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트의 '흡연 중인 담배' 지시에 맞게 담배 끝에 불이 붙어 있어 B보다 요구사항을 더 충실히 이행했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "전반적으로 우수하나, 담배 끝에 불이 붙어 있지 않아 '흡연 중'이라는 디테일에서 A에 밀립니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선은 정면을 향하고, 담배는 입에서 바깥쪽을 향해 똑바로 물려 있습니다.",
      "built_space": "경찰서 앞 야외 벤치에 위치하며, 뒤편에 건물과 표지판이 자연스럽게 자리 잡고 있습니다.",
      "entities": "서의용의 얼굴과 복장이 레퍼런스와 일치하며, 표지판의 '흡연구역' 텍스트가 정확히 표기되었습니다.",
      "hard_violations": [],
      "physics": "인물은 벤치에 안정적으로 앉아 체중을 지탱하고 있으며, 담배는 입술로 지지되고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선은 정면을 향하고, 담배는 입에서 바깥쪽을 향해 똑바로 물려 있습니다.",
      "built_space": "경찰서 앞 야외 벤치에 위치하며, 뒤편에 건물과 표지판이 자연스럽게 자리 잡고 있습니다.",
      "entities": "서의용의 얼굴과 복장이 레퍼런스와 일치하며, 표지판의 '흡연구역' 텍스트가 정확히 표기되었습니다.",
      "hard_violations": [],
      "physics": "인물은 벤치에 안정적으로 앉아 체중을 지탱하고 있으며, 담배는 입술로 지지되고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트의 '흡연 중인 담배' 지시에 맞게 담배 끝에 불이 붙어 있어 B보다 요구사항을 더 충실히 이행했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "전반적으로 우수하나, 담배 끝에 불이 붙어 있지 않아 '흡연 중'이라는 디테일에서 A에 밀립니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선은 정면을 향하고, 담배는 입에서 바깥쪽을 향해 똑바로 물려 있습니다.",
      "built_space": "경찰서 앞 야외 벤치에 위치하며, 뒤편에 건물과 표지판이 자연스럽게 자리 잡고 있습니다.",
      "entities": "서의용의 얼굴과 복장이 레퍼런스와 일치하며, 표지판의 '흡연구역' 텍스트가 정확히 표기되었습니다.",
      "hard_violations": [],
      "physics": "인물은 벤치에 안정적으로 앉아 체중을 지탱하고 있으며, 담배는 입술로 지지되고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선은 정면을 향하고, 담배는 입에서 바깥쪽을 향해 똑바로 물려 있습니다.",
      "built_space": "경찰서 앞 야외 벤치에 위치하며, 뒤편에 건물과 표지판이 자연스럽게 자리 잡고 있습니다.",
      "entities": "서의용의 얼굴과 복장이 레퍼런스와 일치하며, 표지판의 '흡연구역' 텍스트가 정확히 표기되었습니다.",
      "hard_violations": [],
      "physics": "인물은 벤치에 안정적으로 앉아 체중을 지탱하고 있으며, 담배는 입술로 지지되고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "캐릭터의 외모와 의상, 지정된 위치의 벤치, '흡연구역' 표지판 텍스트 및 미디엄 샷 프레이밍은 매우 훌륭하게 구현되었으나, 담배를 거꾸로(주황색 필터가 바깥을 향하도록) 물고 있어 프롭의 기능적 방향 지시를 크게 위반했습니다."
     },
     {
      "label": "B",
      "score": 2000,
      "verdict_ko": "A와 시각적으로 거의 동일하며 캐릭터, 배경, 간판 텍스트의 디테일은 우수하나, 마찬가지로 담배의 필터 부분이 바깥쪽을 향해 거꾸로 물려 있는 오류를 공유하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.857,
      "B": 2.0
     },
     "adjusted": {
      "A": 1.857,
      "B": 2.0
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.143,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "캐릭터의 외모와 의상, 지정된 위치의 벤치, '흡연구역' 표지판 텍스트 및 미디엄 샷 프레이밍은 매우 훌륭하게 구현되었으나, 담배를 거꾸로(주황색 필터가 바깥을 향하도록) 물고 있어 프롭의 기능적 방향 지시를 크게 위반했습니다."
     },
     {
      "label": "A",
      "score": 2000,
      "verdict_ko": "A와 시각적으로 거의 동일하며 캐릭터, 배경, 간판 텍스트의 디테일은 우수하나, 마찬가지로 담배의 필터 부분이 바깥쪽을 향해 거꾸로 물려 있는 오류를 공유하고 있습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2008,
     "B": 1864
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S11sh1__bgfirst_bg.png",
   "bg_asset_id": "fa5a6a76-663e-449b-b974-a79c1facbbc9",
   "bg_record_key": "S11sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S11sh1::cine": {
  "applied": true,
  "fingerprint": "1ac066f92c0beddf46f5537533e87cb2e7c648ddfeffe976be108a1bead5b8dd",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S11sh1_sel.png",
  "source_sha256": "52c0392836ad0ba4e044c2096ca0fe2ae01c6cec7c92d47c58059b704ca51264",
  "file": "S11sh1_cine.png",
  "latency_ms": 10797
 },
 "S12sh3::signage": {
  "fp": "3056dc2e659a52e2",
  "inscriptions": [
   {
    "surface_native": "강당 정면 벽면 표어",
    "text_native": "정성을 다하는 국민의 경찰",
    "reason_ko": "2000년대 대한민국 경찰서 강당 전면에 흔히 부착되어 있던 공식 슬로건을 배경에 배치하여 경찰서 강당이라는 공간적 배경의 사실성을 더하기 위함입니다."
   }
  ]
 },
 "era_assess::50d8591a0378c6ff": {
  "subjects": [
   {
    "subject_native": "대한민국 경찰서 강당 (2001년-2017년)",
    "search_terms_native": [
     "경찰서 강당 단상",
     "경찰서 대강당",
     "경찰서 시무식"
    ],
    "language_lock_native": "검색어는 오직 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 대한민국 경찰서 강당 단상의 참수리 문양, 태극기 및 경찰기 배치, 강당 의자와 인테리어 등의 세부 특징을 고증하기 위함입니다."
   }
  ]
 },
 "era_ref::7c3fabc84c6d560e": {
  "subject": "대한민국 경찰서 강당 (2001년-2017년)",
  "terms": [
   "경찰서 강당 단상",
   "경찰서 대강당",
   "경찰서 시무식"
  ],
  "queries": [
   [
    "경찰서 강당 단상 경찰서 대강당 경찰서 시무식 2001년 2017년",
    "경찰서 시무식 강당 2000년대",
    "경찰서 시무식 대강당 2010년대"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://ojsfile.ohmynews.com/STD_IMG_FILE/2017/0701/IE002183422_STD.jpg",
  "picked_reason_ko": "2017년 광주경찰서의 실제 강당 무대와 목재 마감, 연단, 행사 설비가 선명하게 보여 해당 시대의 일상적인 경찰서 강당 참고자료로 가장 적합하다.",
  "sha256": "0d625f04c48b9af35836c87dea73c7a59be4b01b7fca500d97777ae5d4d08f35",
  "file": "eraref_7c3fabc84c6d560e.png"
 },
 "S12sh3::bgfirst_bg": {
  "input_fingerprint": "898392f9f6bc2d78",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 굳은 표정으로 서의용을 곧게 응시하며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the front of the police-station assembly hall, beside the podium and facing the seated staff under indoor lighting.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At chest height near the front side of the room, hold a medium-close, slightly low three-quarter view that isolates 전택수's upper body without flattening him into a frontal portrait. His weight is set toward the rear row, mouth open mid-challenge and eyes locked on 서의용 beyond the frame, with only restrained meeting-room context remaining at the edges.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 서의용 in the rear row.\n- KEY BACKGROUND ELEMENTS: front area of the meeting room (No additional state specified); used as Retains minimal institutional context behind the isolated speaker; platform area (전택수 is positioned before it) — Its front-side edge recedes behind 전택수; used as Marks 전택수's speaking position without overtaking the portrait.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience supports a subdued neutral-to-cool palette, moderate contrast, and readable facial tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 경찰서 강당 (2001년-2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 굳은 표정으로 서의용을 곧게 응시하며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the front of the police-station assembly hall, beside the podium and facing the seated staff under indoor lighting.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At chest height near the front side of the room, hold a medium-close, slightly low three-quarter view that isolates 전택수's upper body without flattening him into a frontal portrait. His weight is set toward the rear row, mouth open mid-challenge and eyes locked on 서의용 beyond the frame, with only restrained meeting-room context remaining at the edges.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 서의용 in the rear row.\n- KEY BACKGROUND ELEMENTS: front area of the meeting room (No additional state specified); used as Retains minimal institutional context behind the isolated speaker; platform area (전택수 is positioned before it) — Its front-side edge recedes behind 전택수; used as Marks 전택수's speaking position without overtaking the portrait.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience supports a subdued neutral-to-cool palette, moderate contrast, and readable facial tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 경찰서 강당 (2001년-2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S12sh3__bgfirst_bg.png",
  "asset_id": "67e903c8-fed5-4d2b-9368-6057cf0cf35d",
  "input_asset_ids": [
   "6f0fa98e-0c22-41fc-b9e4-e012e92982fb",
   "79629a30-d527-4029-bfd8-1f3a180ee285"
  ],
  "era_research": {
   "subject": "대한민국 경찰서 강당 (2001년-2017년)",
   "queries": [
    [
     "경찰서 강당 단상 경찰서 대강당 경찰서 시무식 2001년 2017년",
     "경찰서 시무식 강당 2000년대",
     "경찰서 시무식 대강당 2010년대"
    ]
   ],
   "picked_url": "https://ojsfile.ohmynews.com/STD_IMG_FILE/2017/0701/IE002183422_STD.jpg",
   "sha256": "0d625f04c48b9af35836c87dea73c7a59be4b01b7fca500d97777ae5d4d08f35",
   "file": "eraref_7c3fabc84c6d560e.png"
  }
 },
 "S12sh3": {
  "input_fingerprint": "8d09c6e337048c92",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 굳은 표정으로 서의용을 곧게 응시하며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the front of the police-station assembly hall, beside the podium and facing the seated staff under indoor lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height near the front side of the room, hold a medium-close, slightly low three-quarter view that isolates 전택수's upper body without flattening him into a frontal portrait. His weight is set toward the rear row, mouth open mid-challenge and eyes locked on 서의용 beyond the frame, with only restrained meeting-room context remaining at the edges.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 서의용 in the rear row.\n- KEY BACKGROUND ELEMENTS: front area of the meeting room (No additional state specified); used as Retains minimal institutional context behind the isolated speaker; platform area (전택수 is positioned before it) — Its front-side edge recedes behind 전택수; used as Marks 전택수's speaking position without overtaking the portrait.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience supports a subdued neutral-to-cool palette, moderate contrast, and readable facial tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 강당 정면 벽면 표어: \"정성을 다하는 국민의 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 굳은 표정으로 서의용을 곧게 응시하며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the front of the police-station assembly hall, beside the podium and facing the seated staff under indoor lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height near the front side of the room, hold a medium-close, slightly low three-quarter view that isolates 전택수's upper body without flattening him into a frontal portrait. His weight is set toward the rear row, mouth open mid-challenge and eyes locked on 서의용 beyond the frame, with only restrained meeting-room context remaining at the edges.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 서의용 in the rear row.\n- KEY BACKGROUND ELEMENTS: front area of the meeting room (No additional state specified); used as Retains minimal institutional context behind the isolated speaker; platform area (전택수 is positioned before it) — Its front-side edge recedes behind 전택수; used as Marks 전택수's speaking position without overtaking the portrait.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience supports a subdued neutral-to-cool palette, moderate contrast, and readable facial tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 강당 정면 벽면 표어: \"정성을 다하는 국민의 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 굳은 표정으로 서의용을 곧게 응시하며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the front of the police-station assembly hall, beside the podium and facing the seated staff under indoor lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height near the front side of the room, hold a medium-close, slightly low three-quarter view that isolates 전택수's upper body without flattening him into a frontal portrait. His weight is set toward the rear row, mouth open mid-challenge and eyes locked on 서의용 beyond the frame, with only restrained meeting-room context remaining at the edges.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 서의용 in the rear row.\n- KEY BACKGROUND ELEMENTS: front area of the meeting room (No additional state specified); used as Retains minimal institutional context behind the isolated speaker; platform area (전택수 is positioned before it) — Its front-side edge recedes behind 전택수; used as Marks 전택수's speaking position without overtaking the portrait.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience supports a subdued neutral-to-cool palette, moderate contrast, and readable facial tension.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 강당 정면 벽면 표어: \"정성을 다하는 국민의 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S12sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S12sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L15B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 미디엄 클로즈업 프레이밍으로 상체를 정확히 포착했으며, 명시된 짧은 머리를 드러내고 강당 정면의 스크린 구조를 온전히 유지하여 가장 완성도가 높습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레이밍이 너무 넓어 상체 고립에 실패했고, 레퍼런스의 핵심 구조물인 정면 스크린을 삭제했으며 모자를 써서 지정된 헤어스타일을 가렸습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 프레임 밖 우측을 향하고 있으며, 입을 벌린 채 명시된 대상을 강하게 응시함.",
      "built_space": "강당 내부 프론트 영역. 청중석 의자들이 정면을 향해 있음. 연단이 단상 밖 바닥 좌측에 위치함. 배경 중앙에 있어야 할 프로젝터 스크린이 사라지고 벽면에 표어가 적혀 있음.",
      "entities": "전택수의 얼굴과 의상은 일치하나 모자를 착용하여 프롬프트의 '짧은 검은색 머리'가 보이지 않음. 표어 '정성을 다하는 국민의 경찰'이 명확히 표기됨.",
      "hard_violations": [
       "지정된 로케이션의 고정 구조물(정면 중앙 스크린) 임의 삭제"
      ],
      "physics": "바닥에 두 발로 서서 체중을 지탱하고 있으며 자세에 물리적 어색함이 없음."
     },
     {
      "label": "B",
      "direction": "시선은 프레임 밖 우측 전방을 향하고 있으며, 입을 벌려 말하는 듯한 표정으로 목표를 응시함.",
      "built_space": "강당 내부 프론트 영역. 정면 벽면에 스크린이 온전히 존재하며 그 위쪽 구조물에 표어가 배치됨. 인물이 연단 옆이 아닌 연단 바로 뒤에 서 있음.",
      "entities": "전택수의 얼굴, 복장이 일치하며 모자를 벗어 명시된 헤어스타일이 노출됨. 표어 '정성을 다하는 국민의 경찰' 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "연단 뒤 바닥에 서서 안정적인 상체 자세를 유지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 미디엄 클로즈업 프레이밍으로 상체를 정확히 포착했으며, 명시된 짧은 머리를 드러내고 강당 정면의 스크린 구조를 온전히 유지하여 가장 완성도가 높습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레이밍이 너무 넓어 상체 고립에 실패했고, 레퍼런스의 핵심 구조물인 정면 스크린을 삭제했으며 모자를 써서 지정된 헤어스타일을 가렸습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 프레임 밖 우측을 향하고 있으며, 입을 벌린 채 명시된 대상을 강하게 응시함.",
      "built_space": "강당 내부 프론트 영역. 청중석 의자들이 정면을 향해 있음. 연단이 단상 밖 바닥 좌측에 위치함. 배경 중앙에 있어야 할 프로젝터 스크린이 사라지고 벽면에 표어가 적혀 있음.",
      "entities": "전택수의 얼굴과 의상은 일치하나 모자를 착용하여 프롬프트의 '짧은 검은색 머리'가 보이지 않음. 표어 '정성을 다하는 국민의 경찰'이 명확히 표기됨.",
      "hard_violations": [
       "지정된 로케이션의 고정 구조물(정면 중앙 스크린) 임의 삭제"
      ],
      "physics": "바닥에 두 발로 서서 체중을 지탱하고 있으며 자세에 물리적 어색함이 없음."
     },
     {
      "label": "B",
      "direction": "시선은 프레임 밖 우측 전방을 향하고 있으며, 입을 벌려 말하는 듯한 표정으로 목표를 응시함.",
      "built_space": "강당 내부 프론트 영역. 정면 벽면에 스크린이 온전히 존재하며 그 위쪽 구조물에 표어가 배치됨. 인물이 연단 옆이 아닌 연단 바로 뒤에 서 있음.",
      "entities": "전택수의 얼굴, 복장이 일치하며 모자를 벗어 명시된 헤어스타일이 노출됨. 표어 '정성을 다하는 국민의 경찰' 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "연단 뒤 바닥에 서서 안정적인 상체 자세를 유지하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1667,
      "verdict_ko": "지정된 샷 크기(상체 위주의 미디엄 샷)를 정확히 준수하였으며, 지문이 요구한 짧은 머리카락 묘사와 표어 텍스트가 매우 우수하게 구현되었습니다."
     },
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "공간 내 위치와 표어 구현은 좋으나, 프레임이 피사체의 하반신까지 넓게 잡혀 우선순위인 프레이밍 스케일에서 감점되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.667,
      "B": 1.714
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.333,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "지정된 샷 크기(상체 위주의 미디엄 샷)를 정확히 준수하였으며, 지문이 요구한 짧은 머리카락 묘사와 표어 텍스트가 매우 우수하게 구현되었습니다."
     },
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "공간 내 위치와 표어 구현은 좋으나, 프레임이 피사체의 하반신까지 넓게 잡혀 우선순위인 프레이밍 스케일에서 감점되었습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1718,
     "B": 1674
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 밖 우측을 향하고 있으며, 입을 벌린 채 명시된 대상을 강하게 응시함.",
    "built_space": "강당 내부 프론트 영역. 청중석 의자들이 정면을 향해 있음. 연단이 단상 밖 바닥 좌측에 위치함. 배경 중앙에 있어야 할 프로젝터 스크린이 사라지고 벽면에 표어가 적혀 있음.",
    "entities": "전택수의 얼굴과 의상은 일치하나 모자를 착용하여 프롬프트의 '짧은 검은색 머리'가 보이지 않음. 표어 '정성을 다하는 국민의 경찰'이 명확히 표기됨.",
    "hard_violations": [
     "지정된 로케이션의 고정 구조물(정면 중앙 스크린) 임의 삭제"
    ],
    "physics": "바닥에 두 발로 서서 체중을 지탱하고 있으며 자세에 물리적 어색함이 없음."
   },
   {
    "label": "B",
    "direction": "시선은 프레임 밖 우측 전방을 향하고 있으며, 입을 벌려 말하는 듯한 표정으로 목표를 응시함.",
    "built_space": "강당 내부 프론트 영역. 정면 벽면에 스크린이 온전히 존재하며 그 위쪽 구조물에 표어가 배치됨. 인물이 연단 옆이 아닌 연단 바로 뒤에 서 있음.",
    "entities": "전택수의 얼굴, 복장이 일치하며 모자를 벗어 명시된 헤어스타일이 노출됨. 표어 '정성을 다하는 국민의 경찰' 텍스트가 정확함.",
    "hard_violations": [],
    "physics": "연단 뒤 바닥에 서서 안정적인 상체 자세를 유지하고 있음."
   }
  ],
  "totals": {
   "A": 1718,
   "B": 1674
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 미디엄 클로즈업 프레이밍으로 상체를 정확히 포착했으며, 명시된 짧은 머리를 드러내고 강당 정면의 스크린 구조를 온전히 유지하여 가장 완성도가 높습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "프레이밍이 너무 넓어 상체 고립에 실패했고, 레퍼런스의 핵심 구조물인 정면 스크린을 삭제했으며 모자를 써서 지정된 헤어스타일을 가렸습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L15B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "레이아웃 스케치에 지정된 오른팔을 들어 가리키는 포즈가 무시되고 양팔을 내린 경직된 차려 자세로 묘사됨.",
     "fix_en": "Raise the character's right arm to point forward. Preserve his clothing, facial expression, and the background elements.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "지정된 원본 배경에서 중앙에 위치했던 프로젝터 스크린이 완전히 삭제됨.",
     "fix_en": "Restore the large white projector screen on the front wall. Preserve the character's identity, pose, and lighting.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "원본 배경의 카메라 투시와 원근감이 유지되지 않아, 배경의 단상과 의자 크기 및 배치가 임의로 변형됨.",
     "fix_en": "Restore the exact chair layout, podium position, and camera perspective of the original background. Preserve the character.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 우측 깃발에 인쇄된 글자가 원본과 달리 알아볼 수 없게 뭉개져 있음.",
     "fix_en": "Correct the garbled text on the right side flag. Preserve the character, the wall text, and the rest of the room.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "전택수가 거의 정면으로 렌즈를 향해 서서 요구된 약간 로우 3/4 뷰가 아니며 초상화처럼 보인다",
     "fix_en": "Change the camera angle to a slightly low 3/4 view of the character. Preserve his clothing, identity, and the background.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전택수 시선이 프레임 밖 후열 서의용이 아니라 거의 카메라를 향한다",
     "fix_en": "Direct the character's eyes off-camera to the right instead of looking at the lens. Preserve his facial structure and pose.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "상체만 고립시키는 미디엄 클로스가 아니라 허벅지까지 보이는 넓은 샷으로 회의실 배경이 过多하다",
     "fix_en": "Adjust framing to a tighter medium-close shot isolating the upper body. Preserve the character and the background elements shown.",
     "severity": "major",
     "observation_index": 6,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "레이아웃 스케치에 지정된 오른팔을 들어 가리키는 포즈가 무시되고 양팔을 내린 경직된 차려 자세로 묘사됨.",
     "severity": "major"
    },
    {
     "issue_ko": "지정된 원본 배경에서 중앙에 위치했던 프로젝터 스크린이 완전히 삭제됨.",
     "severity": "major"
    },
    {
     "issue_ko": "원본 배경의 카메라 투시와 원근감이 유지되지 않아, 배경의 단상과 의자 크기 및 배치가 임의로 변형됨.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 우측 깃발에 인쇄된 글자가 원본과 달리 알아볼 수 없게 뭉개져 있음.",
     "severity": "minor"
    },
    {
     "issue_ko": "전택수가 거의 정면으로 렌즈를 향해 서서 요구된 약간 로우 3/4 뷰가 아니며 초상화처럼 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수 시선이 프레임 밖 후열 서의용이 아니라 거의 카메라를 향한다",
     "severity": "major"
    },
    {
     "issue_ko": "상체만 고립시키는 미디엄 클로스가 아니라 허벅지까지 보이는 넓은 샷으로 회의실 배경이 过多하다",
     "severity": "major"
    },
    {
     "issue_ko": "레이아웃 스케치의 오른손 가리키는 포즈가 없고 양손이 내려가 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 7,
  "fix_severity_skipped": [
   {
    "issue_ko": "레이아웃 스케치에 지정된 오른팔을 들어 가리키는 포즈가 무시되고 양팔을 내린 경직된 차려 자세로 묘사됨.",
    "fix_en": "Raise the character's right arm to point forward. Preserve his clothing, facial expression, and the background elements.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "지정된 원본 배경에서 중앙에 위치했던 프로젝터 스크린이 완전히 삭제됨.",
    "fix_en": "Restore the large white projector screen on the front wall. Preserve the character's identity, pose, and lighting.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "원본 배경의 카메라 투시와 원근감이 유지되지 않아, 배경의 단상과 의자 크기 및 배치가 임의로 변형됨.",
    "fix_en": "Restore the exact chair layout, podium position, and camera perspective of the original background. Preserve the character.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "화면 우측 깃발에 인쇄된 글자가 원본과 달리 알아볼 수 없게 뭉개져 있음.",
    "fix_en": "Correct the garbled text on the right side flag. Preserve the character, the wall text, and the rest of the room.",
    "severity": "minor",
    "observation_index": 3
   },
   {
    "issue_ko": "전택수가 거의 정면으로 렌즈를 향해 서서 요구된 약간 로우 3/4 뷰가 아니며 초상화처럼 보인다",
    "fix_en": "Change the camera angle to a slightly low 3/4 view of the character. Preserve his clothing, identity, and the background.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "전택수 시선이 프레임 밖 후열 서의용이 아니라 거의 카메라를 향한다",
    "fix_en": "Direct the character's eyes off-camera to the right instead of looking at the lens. Preserve his facial structure and pose.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "상체만 고립시키는 미디엄 클로스가 아니라 허벅지까지 보이는 넓은 샷으로 회의실 배경이 过多하다",
    "fix_en": "Adjust framing to a tighter medium-close shot isolating the upper body. Preserve the character and the background elements shown.",
    "severity": "major",
    "observation_index": 6,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S12sh3__bgfirst_bg.png",
   "bg_asset_id": "67e903c8-fed5-4d2b-9368-6057cf0cf35d",
   "bg_record_key": "S12sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S12sh3::cine": {
  "applied": true,
  "fingerprint": "be7e6bdc8167ef7a0324d002cdf677cddcaf579c30d7caa9c56b957528ff0aca",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S12sh3_sel.png",
  "source_sha256": "8e87ea5c6f490e2c5f3a98db6ec670e82da947d89b01995f1808b67b1ffe938c",
  "file": "S12sh3_cine.png",
  "latency_ms": 13224
 },
 "S12sh4::signage": {
  "fp": "5b7d2093d2beeace",
  "inscriptions": [
   {
    "surface_native": "출입문 표지판",
    "text_native": "대강당",
    "reason_ko": "경찰서 대강당 뒷좌석이라는 공간적 배경을 자연스럽게 보여주기 위해 출입문에 붙은 장소 안내 표지판을 묘사합니다."
   }
  ]
 },
 "S12sh4": {
  "input_fingerprint": "6ec2463661e7cc27",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 눈을 크게 뜬 채 굳어있는 서의용과, 그를 곁눈질로 쳐다보는 강주철의 모습\n\nLOCATION (lock): In the rear seating row of the police-station assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From in front of the rear row and slightly above seated eye level, hold a widened medium two-shot on an oblique axis. 서의용 sits rigid at frame left with widened eyes aimed toward 전택수 off-screen, while 강주철 occupies frame right with his torso still forward but eyes cutting sideways toward 서의용; their different orientations keep the reaction natural rather than posed.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rear-row chairs (Occupied by seated officers) — The chair backs run obliquely across the lower background behind the two men; used as Keeps both seated reactions on the same rear-row plane; meeting-room rear area (No additional state specified); used as Provides restrained depth while maintaining attention on the paired reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience maintains a subdued neutral-to-cool palette and moderate contrast across both reactions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 출입문 표지판: \"대강당\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 눈을 크게 뜬 채 굳어있는 서의용과, 그를 곁눈질로 쳐다보는 강주철의 모습\n\nLOCATION (lock): In the rear seating row of the police-station assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From in front of the rear row and slightly above seated eye level, hold a widened medium two-shot on an oblique axis. 서의용 sits rigid at frame left with widened eyes aimed toward 전택수 off-screen, while 강주철 occupies frame right with his torso still forward but eyes cutting sideways toward 서의용; their different orientations keep the reaction natural rather than posed.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rear-row chairs (Occupied by seated officers) — The chair backs run obliquely across the lower background behind the two men; used as Keeps both seated reactions on the same rear-row plane; meeting-room rear area (No additional state specified); used as Provides restrained depth while maintaining attention on the paired reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience maintains a subdued neutral-to-cool palette and moderate contrast across both reactions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 출입문 표지판: \"대강당\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 눈을 크게 뜬 채 굳어있는 서의용과, 그를 곁눈질로 쳐다보는 강주철의 모습\n\nLOCATION (lock): In the rear seating row of the police-station assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From in front of the rear row and slightly above seated eye level, hold a widened medium two-shot on an oblique axis. 서의용 sits rigid at frame left with widened eyes aimed toward 전택수 off-screen, while 강주철 occupies frame right with his torso still forward but eyes cutting sideways toward 서의용; their different orientations keep the reaction natural rather than posed.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: rear-row chairs (Occupied by seated officers) — The chair backs run obliquely across the lower background behind the two men; used as Keeps both seated reactions on the same rear-row plane; meeting-room rear area (No additional state specified); used as Provides restrained depth while maintaining attention on the paired reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural morning interior ambience maintains a subdued neutral-to-cool palette and moderate contrast across both reactions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 출입문 표지판: \"대강당\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.417,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 지정된 인물의 좌우 위치가 뒤바뀜 (서의용이 왼쪽, 강주철이 오른쪽이어야 함)"
    ],
    "B": [
     "[gemini-pro] 지정된 인물의 좌우 위치가 뒤바뀜 (서의용이 왼쪽, 강주철이 오른쪽이어야 함)",
     "[gemini-pro] 카메라 방향 오류 (강당 뒷벽을 배경으로 해야 하나 단상이 있는 앞쪽을 배경으로 렌더링함)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1417,
   "B": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1417,
    "verdict_ko": "두 인물의 좌우 배치가 프롬프트의 지시와 완전히 반대로 적용되어 치명적인 오류가 발생했습니다.  ★위반: [gemini-pro] 지정된 인물의 좌우 위치가 뒤바뀜 (서의용이 왼쪽, 강주철이 오른쪽이어야 함)"
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "인물의 좌우 배치가 뒤바뀌었으며, 카메라가 강당 뒷벽이 아닌 앞쪽 단상을 향하고 있어 배경 설정마저 위반했습니다.  ★위반: [gemini-pro] 지정된 인물의 좌우 위치가 뒤바뀜 (서의용이 왼쪽, 강주철이 오른쪽이어야 함) / [gemini-pro] 카메라 방향 오류 (강당 뒷벽을 배경으로 해야 하나 단상이 있는 앞쪽을 배경으로 렌더링함)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S12sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 서의용(검은 가죽 재킷)을 화면 왼쪽, 강주철(주황색 재킷)을 화면 오른쪽에 배치하도록 지시했으나, 두 인물의 위치가 서로 뒤바뀌어 있습니다.",
     "fix_en": "Change the left man to wear a black leather jacket and blue beanie, and the right man to wear an orange jacket and grey beanie, swapping their identities. Preserve the room, chairs, background figures, camera angle, and lighting.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 왼쪽 인물(강주철 복장)의 재킷 가슴 부분에 레퍼런스에 존재하지 않는 '수사'라는 텍스트가 임의로 추가되었습니다.",
     "fix_en": "Remove the yellow text from the left man's orange jacket, leaving only the blank jacket material in that area. Preserve both men's faces, poses, all other clothing, the background, and the lighting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "화면 오른쪽 인물(서의용 복장)의 시선이 프롬프트가 지시한 화면 밖(off-screen)이 아닌 옆 사람을 향하고 있습니다.",
     "fix_en": "Turn the right man's eyes to look toward the left off-screen area instead of at the man next to him. Preserve his face, pose, clothing, the left man, and the background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "배경에 앉은 인물들이 이전 컷(Previous Shot) 등장인물의 의상과 모자('수사' 문구 포함)를 그대로 착용하고 있어, 이전 컷 인물의 의상을 이관하지 말라는 지시를 위반했습니다.",
     "fix_en": "Change the clothes of the background people to generic plainclothes without caps, removing the blue jackets and hats. Preserve the foreground men, their clothing, the chairs, and the room.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 서의용(검은 가죽 재킷)을 화면 왼쪽, 강주철(주황색 재킷)을 화면 오른쪽에 배치하도록 지시했으나, 두 인물의 위치가 서로 뒤바뀌어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 왼쪽 인물(강주철 복장)의 재킷 가슴 부분에 레퍼런스에 존재하지 않는 '수사'라는 텍스트가 임의로 추가되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 오른쪽 인물(서의용 복장)의 시선이 프롬프트가 지시한 화면 밖(off-screen)이 아닌 옆 사람을 향하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경에 앉은 인물들이 이전 컷(Previous Shot) 등장인물의 의상과 모자('수사' 문구 포함)를 그대로 착용하고 있어, 이전 컷 인물의 의상을 이관하지 말라는 지시를 위반했습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "서의용이 화면 오른쪽, 강주철이 왼쪽에 앉아 지정된 좌우 배치와 반대이다",
     "severity": "critical"
    },
    {
     "issue_ko": "눈을 크게 뜨고 굳은 쪽은 강주철이고 곁눈질하는 쪽은 서의용으로 연기가 뒤바뀌었다",
     "severity": "critical"
    },
    {
     "issue_ko": "두 사람이 정면으로 나란히 앉아 사선 축의 리액션 투샷이 아니다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 인물 재킷 가슴에 장면에 없는 '수사' 글자가 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 오른쪽 인물(서의용 복장)의 시선이 프롬프트가 지시한 화면 밖(off-screen)이 아닌 옆 사람을 향하고 있습니다.",
    "fix_en": "Turn the right man's eyes to look toward the left off-screen area instead of at the man next to him. Preserve his face, pose, clothing, the left man, and the background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "배경에 앉은 인물들이 이전 컷(Previous Shot) 등장인물의 의상과 모자('수사' 문구 포함)를 그대로 착용하고 있어, 이전 컷 인물의 의상을 이관하지 말라는 지시를 위반했습니다.",
    "fix_en": "Change the clothes of the background people to generic plainclothes without caps, removing the blue jackets and hats. Preserve the foreground men, their clothing, the chairs, and the room.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the left man to wear a black leather jacket and blue beanie, and the right man to wear an orange jacket and grey beanie, swapping their identities. Preserve the room, chairs, background figures, camera angle, and lighting.\n- Remove the yellow text from the left man's orange jacket, leaving only the blank jacket material in that area. Preserve both men's faces, poses, all other clothing, the background, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 167,
      "verdict_ko": "뒷열 좌석에서의 카메라 구도와 두 인물의 시선 처리 등 연출 지시를 훌륭하게 따랐으나, 두 인물의 의상이 서로 뒤바뀌고 옷에 잘못된 텍스트가 유출되어 정체성과 디테일에서 큰 감점 요소가 됨.  ★위반: [gemini-pro] 두 인물의 의상과 모자가 서로 뒤바뀌어 지정된 인물의 정체성이 훼손됨 / [gemini-pro] 왼쪽 인물의 주황색 점퍼 가슴에 이전 샷에서 유출된 '수사' 텍스트가 새겨짐 (Leaked text) / [gemini-pro] 이전 샷 인물의 특정 의상(수사 모자, 경찰 로고 재킷)을 배경 엑스트라들이 그대로 입고 있음 / [openrouter:x-ai/grok-4.6] 프레임 왼쪽 지정석에 서의용 대신 강주철이 앉음 / [openrouter:x-ai/grok-4.6] 프레임 오른쪽 지정석에 강주철 대신 서의용이 앉음 / [openrouter:x-ai/grok-4.6] 샷 텍스트가 지명하지 않은 경찰 3명을 배경에 추가"
     },
     {
      "label": "B",
      "score": 583,
      "verdict_ko": "인물의 얼굴과 의상 배치는 정확하게 구현했으나, 지정된 뒷열 좌석과 카메라 구도를 완전히 무시한 채 금지된 이전 샷의 구도를 그대로 복사하여 서 있는 모습으로 연출함.  ★위반: [gemini-pro] 카메라가 지시된 뒷열 앞이 아닌 강당 전면에 위치함 (Camera viewpoint violation) / [gemini-pro] 인물들이 지정된 뒷열 의자에 앉지 않고 단상 앞 바닥에 서 있음 (Pose and Location violation) / [gemini-pro] 명시적으로 금지된 이전 샷 레퍼런스의 카메라 구도와 배경을 그대로 복사함"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.333
     },
     "adjusted": {
      "A": 0.167,
      "B": 0.583
     },
     "violations": {
      "A": [
       "[gemini-pro] 두 인물의 의상과 모자가 서로 뒤바뀌어 지정된 인물의 정체성이 훼손됨",
       "[gemini-pro] 왼쪽 인물의 주황색 점퍼 가슴에 이전 샷에서 유출된 '수사' 텍스트가 새겨짐 (Leaked text)",
       "[gemini-pro] 이전 샷 인물의 특정 의상(수사 모자, 경찰 로고 재킷)을 배경 엑스트라들이 그대로 입고 있음",
       "[openrouter:x-ai/grok-4.6] 프레임 왼쪽 지정석에 서의용 대신 강주철이 앉음",
       "[openrouter:x-ai/grok-4.6] 프레임 오른쪽 지정석에 강주철 대신 서의용이 앉음",
       "[openrouter:x-ai/grok-4.6] 샷 텍스트가 지명하지 않은 경찰 3명을 배경에 추가"
      ],
      "B": [
       "[gemini-pro] 카메라가 지시된 뒷열 앞이 아닌 강당 전면에 위치함 (Camera viewpoint violation)",
       "[gemini-pro] 인물들이 지정된 뒷열 의자에 앉지 않고 단상 앞 바닥에 서 있음 (Pose and Location violation)",
       "[gemini-pro] 명시적으로 금지된 이전 샷 레퍼런스의 카메라 구도와 배경을 그대로 복사함"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.333,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 167,
      "verdict_ko": "뒷열 좌석에서의 카메라 구도와 두 인물의 시선 처리 등 연출 지시를 훌륭하게 따랐으나, 두 인물의 의상이 서로 뒤바뀌고 옷에 잘못된 텍스트가 유출되어 정체성과 디테일에서 큰 감점 요소가 됨.  ★위반: [gemini-pro] 두 인물의 의상과 모자가 서로 뒤바뀌어 지정된 인물의 정체성이 훼손됨 / [gemini-pro] 왼쪽 인물의 주황색 점퍼 가슴에 이전 샷에서 유출된 '수사' 텍스트가 새겨짐 (Leaked text) / [gemini-pro] 이전 샷 인물의 특정 의상(수사 모자, 경찰 로고 재킷)을 배경 엑스트라들이 그대로 입고 있음 / [openrouter:x-ai/grok-4.6] 프레임 왼쪽 지정석에 서의용 대신 강주철이 앉음 / [openrouter:x-ai/grok-4.6] 프레임 오른쪽 지정석에 강주철 대신 서의용이 앉음 / [openrouter:x-ai/grok-4.6] 샷 텍스트가 지명하지 않은 경찰 3명을 배경에 추가"
     },
     {
      "label": "B",
      "score": 583,
      "verdict_ko": "인물의 얼굴과 의상 배치는 정확하게 구현했으나, 지정된 뒷열 좌석과 카메라 구도를 완전히 무시한 채 금지된 이전 샷의 구도를 그대로 복사하여 서 있는 모습으로 연출함.  ★위반: [gemini-pro] 카메라가 지시된 뒷열 앞이 아닌 강당 전면에 위치함 (Camera viewpoint violation) / [gemini-pro] 인물들이 지정된 뒷열 의자에 앉지 않고 단상 앞 바닥에 서 있음 (Pose and Location violation) / [gemini-pro] 명시적으로 금지된 이전 샷 레퍼런스의 카메라 구도와 배경을 그대로 복사함"
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "카메라 프레이밍, 장소(뒷좌석), 인물들의 시선과 행동을 지시문과 가깝게 구현했으나, 두 인물의 좌우 위치와 역할을 정반대로 뒤바꿔 배치하는 치명적인 스테이징 위반(Hard Violation)이 발생했습니다."
     },
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "인물들의 좌우 위치만 맞을 뿐, '뒷좌석에 앉아있다'는 명시적인 스테이징을 무시하고 강당 앞쪽에 서서 정면을 응시하게 만들어 프롬프트의 지시를 완전히 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 인물 모두 카메라 렌즈를 정면으로 응시하고 있음.",
      "built_space": "강당의 앞쪽 공간. 인물들 뒤로 여러 줄의 빈 의자가 보이며, 배경에 연단과 태극기, 경찰 깃발이 배치되어 있음.",
      "entities": "화면 왼쪽은 서의용(가죽 재킷, 파란색 비니), 오른쪽은 강주철(오렌지색 재킷, 회색 비니)로 신원과 복장이 참조 이미지와 일치함.",
      "hard_violations": [
       "지정된 위치(뒷좌석)와 자세(앉아있음)를 완전히 무시하고 이전 샷 참조 이미지의 구도를 모방해 강당 앞쪽에 서 있음 (잘못된 스테이징)."
      ],
      "physics": "두 인물 모두 바닥에 서서 발로 체중을 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "왼쪽 인물은 화면 오른쪽 밖을 향해 눈을 크게 뜨고 있으며, 오른쪽 인물은 왼쪽 인물 쪽으로 곁눈질을 하고 있음.",
      "built_space": "강당의 뒷좌석. 인물들 바로 뒷줄에 3명의 경찰관이 앉아 있으며, 우측 상단 벽면에 '대강당'이라고 적힌 표지판이 있음.",
      "entities": "화면 왼쪽에 강주철(오렌지색 재킷), 오른쪽에 서의용(가죽 재킷)이 배치되어 프롬프트에서 지시한 좌우 위치(왼쪽 서의용, 오른쪽 강주철)와 완전히 반대로 렌더링됨.",
      "hard_violations": [
       "프롬프트의 지시(왼쪽 서의용, 오른쪽 강주철)와 정반대로 인물들이 지정되지 않은 자리에 잘못 앉아 있음 (위치 및 역할 뒤바뀜)."
      ],
      "physics": "두 인물 모두 의자에 앉아 좌석과 등받이에 체중을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "카메라 프레이밍, 장소(뒷좌석), 인물들의 시선과 행동을 지시문과 가깝게 구현했으나, 두 인물의 좌우 위치와 역할을 정반대로 뒤바꿔 배치하는 치명적인 스테이징 위반(Hard Violation)이 발생했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "인물들의 좌우 위치만 맞을 뿐, '뒷좌석에 앉아있다'는 명시적인 스테이징을 무시하고 강당 앞쪽에 서서 정면을 응시하게 만들어 프롬프트의 지시를 완전히 실패했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "두 인물 모두 카메라 렌즈를 정면으로 응시하고 있음.",
      "built_space": "강당의 앞쪽 공간. 인물들 뒤로 여러 줄의 빈 의자가 보이며, 배경에 연단과 태극기, 경찰 깃발이 배치되어 있음.",
      "entities": "화면 왼쪽은 서의용(가죽 재킷, 파란색 비니), 오른쪽은 강주철(오렌지색 재킷, 회색 비니)로 신원과 복장이 참조 이미지와 일치함.",
      "hard_violations": [
       "지정된 위치(뒷좌석)와 자세(앉아있음)를 완전히 무시하고 이전 샷 참조 이미지의 구도를 모방해 강당 앞쪽에 서 있음 (잘못된 스테이징)."
      ],
      "physics": "두 인물 모두 바닥에 서서 발로 체중을 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "왼쪽 인물은 화면 오른쪽 밖을 향해 눈을 크게 뜨고 있으며, 오른쪽 인물은 왼쪽 인물 쪽으로 곁눈질을 하고 있음.",
      "built_space": "강당의 뒷좌석. 인물들 바로 뒷줄에 3명의 경찰관이 앉아 있으며, 우측 상단 벽면에 '대강당'이라고 적힌 표지판이 있음.",
      "entities": "화면 왼쪽에 강주철(오렌지색 재킷), 오른쪽에 서의용(가죽 재킷)이 배치되어 프롬프트에서 지시한 좌우 위치(왼쪽 서의용, 오른쪽 강주철)와 완전히 반대로 렌더링됨.",
      "hard_violations": [
       "프롬프트의 지시(왼쪽 서의용, 오른쪽 강주철)와 정반대로 인물들이 지정되지 않은 자리에 잘못 앉아 있음 (위치 및 역할 뒤바뀜)."
      ],
      "physics": "두 인물 모두 의자에 앉아 좌석과 등받이에 체중을 지탱하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 169,
     "B": 583
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S12sh3"
  }
 },
 "S12sh4::cine": {
  "applied": true,
  "fingerprint": "2af9f642c42e1660f70e5b2fd659d18e9e0437cec94dfd407cb035575b7e2bd2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S12sh4_sel.png",
  "source_sha256": "dab2f68529882f95f80dafb77ae1fb3cee86f6c4e5c159e80cb92083952037a3",
  "file": "S12sh4_cine.png",
  "latency_ms": 13221
 },
 "S13sh3::signage": {
  "fp": "f89fcca458599a1e",
  "inscriptions": [
   {
    "surface_native": "사무실 벽면 표지판",
    "text_native": "강력1팀",
    "reason_ko": "형사들이 근무하는 강력계 사무실 내부라는 공간적 배경을 사실적으로 나타내기 위해 벽면에 팀 표시가 필요합니다."
   }
  ]
 },
 "era_assess::6332f2c632867db2": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 대한민국 경찰서 강력반 사무실 및 책상",
    "search_terms_native": [
     "경찰서 강력반 사무실",
     "형사과 강력계 책상",
     "경찰서 강력계 내부",
     "2000년대 경찰서"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 경찰서 강력반 특유의 철제 캐비닛, 녹색 책상 매트, 한국식 서류철, 형사과 고유의 지저분하고 밀집된 배치 등은 일반적인 AI 모델이 서구식 현대 형사 사무실로 오해하여 잘못 그리기 쉽습니다."
   }
  ]
 },
 "era_ref::5f0dfc1939e2a579": {
  "subject": "2000년대~2010년대 대한민국 경찰서 강력반 사무실 및 책상",
  "terms": [
   "경찰서 강력반 사무실",
   "형사과 강력계 책상",
   "경찰서 강력계 내부",
   "2000년대 경찰서"
  ],
  "queries": [
   [
    "2000년대 대한민국 경찰서 강력반 사무실 형사과 강력계 책상 내부",
    "2010년대 대한민국 경찰서 형사과 강력계 사무실 책상"
   ]
  ],
  "candidates": 3,
  "picked_index": 2,
  "picked_url": "https://www.yeongnam.com/mnt/file/200706/20070615.010351512200001i4.jpg",
  "picked_reason_ko": "강력반 전용 공간은 아니지만, 해당 시기 대한민국 경찰서에서 실제 사용된 평범한 사무공간과 책상·컴퓨터·전화·서류 집기를 가장 명확하고 사실적으로 보여준다.",
  "sha256": "bf5901212d195c56d7d5081d9fd55364996781bfe50bdc332beb81b4ecdf4116",
  "file": "eraref_5f0dfc1939e2a579.png"
 },
 "S13sh3::bgfirst_bg": {
  "input_fingerprint": "d239d29f94ceae0e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 강주철이 든 핸드폰 화면을 들여다보며 눈을 동그랗게 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the team leader's desk in the busy violent-crimes office.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From close beside 강주철 at seated shoulder height, the camera completes its inward move on 서의용 in a tight three-quarter view, with 강주철's shoulder softened at the near edge. 서의용's widened eyes occupy the upper center as he leans toward the mobile phone held just beneath his eyeline, its screen retained along the lower edge of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 강주철의 핸드폰 화면 (Held out with a photograph displayed) — The active screen face tilts upward toward 서의용 and is seen obliquely along the lower frame edge, carrying the photograph he is studying; used as The visible screen anchors 서의용's lowered eyeline without obscuring his reaction; 강주철의 책상 (In use within the busy office) — Its near edge runs diagonally behind the phone and toward 서의용; used as Provides a restrained spatial boundary behind the close reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast keeps the reaction grounded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 대한민국 경찰서 강력반 사무실 및 책상: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 강주철이 든 핸드폰 화면을 들여다보며 눈을 동그랗게 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the team leader's desk in the busy violent-crimes office.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From close beside 강주철 at seated shoulder height, the camera completes its inward move on 서의용 in a tight three-quarter view, with 강주철's shoulder softened at the near edge. 서의용's widened eyes occupy the upper center as he leans toward the mobile phone held just beneath his eyeline, its screen retained along the lower edge of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 강주철의 핸드폰 화면 (Held out with a photograph displayed) — The active screen face tilts upward toward 서의용 and is seen obliquely along the lower frame edge, carrying the photograph he is studying; used as The visible screen anchors 서의용's lowered eyeline without obscuring his reaction; 강주철의 책상 (In use within the busy office) — Its near edge runs diagonally behind the phone and toward 서의용; used as Provides a restrained spatial boundary behind the close reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast keeps the reaction grounded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 대한민국 경찰서 강력반 사무실 및 책상: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S13sh3__bgfirst_bg.png",
  "asset_id": "1f838a6d-5c90-42ea-a102-6318ae706256",
  "input_asset_ids": [
   "ce7e9797-38b2-449b-a5d8-3b5f01646ba7",
   "96216ca9-2e74-4830-90fc-136fe702e783"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 대한민국 경찰서 강력반 사무실 및 책상",
   "queries": [
    [
     "2000년대 대한민국 경찰서 강력반 사무실 형사과 강력계 책상 내부",
     "2010년대 대한민국 경찰서 형사과 강력계 사무실 책상"
    ]
   ],
   "picked_url": "https://www.yeongnam.com/mnt/file/200706/20070615.010351512200001i4.jpg",
   "sha256": "bf5901212d195c56d7d5081d9fd55364996781bfe50bdc332beb81b4ecdf4116",
   "file": "eraref_5f0dfc1939e2a579.png"
  }
 },
 "S13sh3": {
  "input_fingerprint": "d0b3b08455b571a9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철이 든 핸드폰 화면을 들여다보며 눈을 동그랗게 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the team leader's desk in the busy violent-crimes office. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close beside 강주철 at seated shoulder height, the camera completes its inward move on 서의용 in a tight three-quarter view, with 강주철's shoulder softened at the near edge. 서의용's widened eyes occupy the upper center as he leans toward the mobile phone held just beneath his eyeline, its screen retained along the lower edge of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 강주철의 핸드폰 화면 (Held out with a photograph displayed) — The active screen face tilts upward toward 서의용 and is seen obliquely along the lower frame edge, carrying the photograph he is studying; used as The visible screen anchors 서의용's lowered eyeline without obscuring his reaction; 강주철의 책상 (In use within the busy office) — Its near edge runs diagonally behind the phone and toward 서의용; used as Provides a restrained spatial boundary behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast keeps the reaction grounded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 강주철 right now — 강주철's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 강주철 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 표지판: \"강력1팀\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철이 든 핸드폰 화면을 들여다보며 눈을 동그랗게 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the team leader's desk in the busy violent-crimes office. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close beside 강주철 at seated shoulder height, the camera completes its inward move on 서의용 in a tight three-quarter view, with 강주철's shoulder softened at the near edge. 서의용's widened eyes occupy the upper center as he leans toward the mobile phone held just beneath his eyeline, its screen retained along the lower edge of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 강주철의 핸드폰 화면 (Held out with a photograph displayed) — The active screen face tilts upward toward 서의용 and is seen obliquely along the lower frame edge, carrying the photograph he is studying; used as The visible screen anchors 서의용's lowered eyeline without obscuring his reaction; 강주철의 책상 (In use within the busy office) — Its near edge runs diagonally behind the phone and toward 서의용; used as Provides a restrained spatial boundary behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast keeps the reaction grounded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 강주철 right now — 강주철's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 강주철 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 표지판: \"강력1팀\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철이 든 핸드폰 화면을 들여다보며 눈을 동그랗게 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the team leader's desk in the busy violent-crimes office. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close beside 강주철 at seated shoulder height, the camera completes its inward move on 서의용 in a tight three-quarter view, with 강주철's shoulder softened at the near edge. 서의용's widened eyes occupy the upper center as he leans toward the mobile phone held just beneath his eyeline, its screen retained along the lower edge of the frame.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 강주철의 핸드폰 화면 (Held out with a photograph displayed) — The active screen face tilts upward toward 서의용 and is seen obliquely along the lower frame edge, carrying the photograph he is studying; used as The visible screen anchors 서의용's lowered eyeline without obscuring his reaction; 강주철의 책상 (In use within the busy office) — Its near edge runs diagonally behind the phone and toward 서의용; used as Provides a restrained spatial boundary behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast keeps the reaction grounded.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 강주철 right now — 강주철's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 강주철 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 표지판: \"강력1팀\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S13sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S13sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L16B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 오버더숄더 앵글과 서의용의 레퍼런스 복장(가죽 재킷, 후드)을 정확히 구현했으며, 사무실 배경과 표지판 텍스트도 훌륭하게 반영되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "정면의 서의용과 전경의 강주철의 복장이 서로 뒤바뀌어 레퍼런스 인물 지시를 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "서의용의 시선이 전경의 인물이 들고 있는 핸드폰 화면을 정확히 향하고 있음.",
      "built_space": "사무실 내부. 복사기와 캐비닛이 배경에 위치하며 우측에 책상이 배치됨.",
      "entities": "서의용은 레퍼런스에 맞는 가죽 재킷과 후드, 신분증을 착용함. 핸드폰 화면에 이미지가 보이며 '강력1팀' 표지판이 벽에 있음.",
      "hard_violations": [],
      "physics": "핸드폰을 쥔 손의 파지가 자연스러우며 인물들의 자세가 안정적임."
     },
     {
      "label": "B",
      "direction": "정면 인물의 시선이 핸드폰 화면을 향하고 있음.",
      "built_space": "사무실 내부. 캐비닛과 창문, 책상이 배치되어 있음.",
      "entities": "정면 인물(서의용)이 파란색 옷을 입고, 전경 인물(강주철)이 서의용의 레퍼런스 복장을 입고 있어 인물 정보가 뒤바뀜. 핸드폰 화면에 사진이 있고 '강력1팀' 표지판이 있음.",
      "hard_violations": [],
      "physics": "핸드폰을 들고 있는 손의 형태와 인물들의 자세를 지지하는 물리적 요소가 정상적임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 오버더숄더 앵글과 서의용의 레퍼런스 복장(가죽 재킷, 후드)을 정확히 구현했으며, 사무실 배경과 표지판 텍스트도 훌륭하게 반영되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "정면의 서의용과 전경의 강주철의 복장이 서로 뒤바뀌어 레퍼런스 인물 지시를 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서의용의 시선이 전경의 인물이 들고 있는 핸드폰 화면을 정확히 향하고 있음.",
      "built_space": "사무실 내부. 복사기와 캐비닛이 배경에 위치하며 우측에 책상이 배치됨.",
      "entities": "서의용은 레퍼런스에 맞는 가죽 재킷과 후드, 신분증을 착용함. 핸드폰 화면에 이미지가 보이며 '강력1팀' 표지판이 벽에 있음.",
      "hard_violations": [],
      "physics": "핸드폰을 쥔 손의 파지가 자연스러우며 인물들의 자세가 안정적임."
     },
     {
      "label": "B",
      "direction": "정면 인물의 시선이 핸드폰 화면을 향하고 있음.",
      "built_space": "사무실 내부. 캐비닛과 창문, 책상이 배치되어 있음.",
      "entities": "정면 인물(서의용)이 파란색 옷을 입고, 전경 인물(강주철)이 서의용의 레퍼런스 복장을 입고 있어 인물 정보가 뒤바뀜. 핸드폰 화면에 사진이 있고 '강력1팀' 표지판이 있음.",
      "hard_violations": [],
      "physics": "핸드폰을 들고 있는 손의 형태와 인물들의 자세를 지지하는 물리적 요소가 정상적임."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "서의용의 외모와 의상, 위치 레퍼런스 및 촬영 구도를 정확히 구현했으나, 핸드폰 화면의 사진이 실루엣으로 처리된 점이 아쉽습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "메인 피사체의 외모가 일치하지 않으며, 서의용의 레퍼런스 의상을 앞쪽 인물에게 입히고 핸드폰 화면 안에 레퍼런스를 띄우는 등 정체성을 심각하게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "프레임 중앙의 남성이 시선을 낮춰 앞쪽의 핸드폰 화면을 똑바로 응시하고 있음.",
      "built_space": "사무실을 배경으로 책상과 서류 캐비닛이 배치되어 있으며, 오른쪽 벽면에 표지판이 평면적으로 붙어 있음.",
      "entities": "반응하는 인물이 서의용 레퍼런스와 전혀 다른 외모와 의상을 가짐. 레퍼런스의 의상(가죽 재킷, 후드)은 전경의 강주철에게 잘못 적용되었으며, 핸드폰 화면 안에 레퍼런스 이미지가 그대로 출력됨. '강력1팀' 텍스트는 명확함.",
      "hard_violations": [
       "지정된 인물(서의용)의 외모와 의상을 메인 피사체가 아닌 다른 곳(전경 인물 및 핸드폰 화면)에 잘못 할당하여 인물 정체성 위반"
      ],
      "physics": "전경의 손이 핸드폰 본체를 쥐고 있으며, 두 인물 모두 무게 중심과 자세가 물리적으로 자연스럽게 지지됨."
     },
     {
      "label": "B",
      "direction": "서의용이 시선을 아래로 향해 앞쪽의 핸드폰 화면을 바라보고 있음.",
      "built_space": "위치 레퍼런스를 훌륭하게 반영한 사무실. 왼쪽에 대형 복사기, 뒤쪽에 캐비닛과 서류철, 앞쪽에 전화기가 놓인 책상이 입체적으로 배치됨.",
      "entities": "서의용의 얼굴, 짧은 머리, 가죽 재킷, 회색 후드 의상 및 목걸이(신분증)가 레퍼런스와 정확히 일치함. '강력1팀' 표지판도 뒤쪽에 정상적으로 배치됨. 핸드폰 화면의 사진이 뚜렷한 인물 사진이 아닌 어두운 실루엣으로 렌더링됨.",
      "hard_violations": [],
      "physics": "전경의 손이 핸드폰 가장자리를 안정적으로 쥐고 있으며, 서의용의 몸이 책상 앞 공간에서 적절히 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "서의용의 외모와 의상, 위치 레퍼런스 및 촬영 구도를 정확히 구현했으나, 핸드폰 화면의 사진이 실루엣으로 처리된 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "메인 피사체의 외모가 일치하지 않으며, 서의용의 레퍼런스 의상을 앞쪽 인물에게 입히고 핸드폰 화면 안에 레퍼런스를 띄우는 등 정체성을 심각하게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "프레임 중앙의 남성이 시선을 낮춰 앞쪽의 핸드폰 화면을 똑바로 응시하고 있음.",
      "built_space": "사무실을 배경으로 책상과 서류 캐비닛이 배치되어 있으며, 오른쪽 벽면에 표지판이 평면적으로 붙어 있음.",
      "entities": "반응하는 인물이 서의용 레퍼런스와 전혀 다른 외모와 의상을 가짐. 레퍼런스의 의상(가죽 재킷, 후드)은 전경의 강주철에게 잘못 적용되었으며, 핸드폰 화면 안에 레퍼런스 이미지가 그대로 출력됨. '강력1팀' 텍스트는 명확함.",
      "hard_violations": [
       "지정된 인물(서의용)의 외모와 의상을 메인 피사체가 아닌 다른 곳(전경 인물 및 핸드폰 화면)에 잘못 할당하여 인물 정체성 위반"
      ],
      "physics": "전경의 손이 핸드폰 본체를 쥐고 있으며, 두 인물 모두 무게 중심과 자세가 물리적으로 자연스럽게 지지됨."
     },
     {
      "label": "A",
      "direction": "서의용이 시선을 아래로 향해 앞쪽의 핸드폰 화면을 바라보고 있음.",
      "built_space": "위치 레퍼런스를 훌륭하게 반영한 사무실. 왼쪽에 대형 복사기, 뒤쪽에 캐비닛과 서류철, 앞쪽에 전화기가 놓인 책상이 입체적으로 배치됨.",
      "entities": "서의용의 얼굴, 짧은 머리, 가죽 재킷, 회색 후드 의상 및 목걸이(신분증)가 레퍼런스와 정확히 일치함. '강력1팀' 표지판도 뒤쪽에 정상적으로 배치됨. 핸드폰 화면의 사진이 뚜렷한 인물 사진이 아닌 어두운 실루엣으로 렌더링됨.",
      "hard_violations": [],
      "physics": "전경의 손이 핸드폰 가장자리를 안정적으로 쥐고 있으며, 서의용의 몸이 책상 앞 공간에서 적절히 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "서의용의 시선이 전경의 인물이 들고 있는 핸드폰 화면을 정확히 향하고 있음.",
    "built_space": "사무실 내부. 복사기와 캐비닛이 배경에 위치하며 우측에 책상이 배치됨.",
    "entities": "서의용은 레퍼런스에 맞는 가죽 재킷과 후드, 신분증을 착용함. 핸드폰 화면에 이미지가 보이며 '강력1팀' 표지판이 벽에 있음.",
    "hard_violations": [],
    "physics": "핸드폰을 쥔 손의 파지가 자연스러우며 인물들의 자세가 안정적임."
   },
   {
    "label": "B",
    "direction": "정면 인물의 시선이 핸드폰 화면을 향하고 있음.",
    "built_space": "사무실 내부. 캐비닛과 창문, 책상이 배치되어 있음.",
    "entities": "정면 인물(서의용)이 파란색 옷을 입고, 전경 인물(강주철)이 서의용의 레퍼런스 복장을 입고 있어 인물 정보가 뒤바뀜. 핸드폰 화면에 사진이 있고 '강력1팀' 표지판이 있음.",
    "hard_violations": [],
    "physics": "핸드폰을 들고 있는 손의 형태와 인물들의 자세를 지지하는 물리적 요소가 정상적임."
   }
  ],
  "totals": {
   "A": 15,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "요구된 오버더숄더 앵글과 서의용의 레퍼런스 복장(가죽 재킷, 후드)을 정확히 구현했으며, 사무실 배경과 표지판 텍스트도 훌륭하게 반영되었습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "정면의 서의용과 전경의 강주철의 복장이 서로 뒤바뀌어 레퍼런스 인물 지시를 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L16B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트가 명시한 '사진' 대신, 핸드폰 화면에 사람의 단순한 검은색 실루엣만 표시되어 있습니다.",
     "fix_en": "Replace the black silhouette on the phone screen with a clear color photograph. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "화면 중앙 서의용의 얼굴이 캐릭터 레퍼런스와 이목구비·골격이 다른 사람으로 보인다",
     "fix_en": "Adjust the central man's facial features to match the character reference identity. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "서의용 얼굴이 타이트 클로즈업으로 상부 중앙을 채우지 못하고 상반신과 책상·사무실 배경이 과도하게 보인다",
     "fix_en": "Scale up the central man to a tight close-up so his eyes occupy the upper center. Preserve the people present, their clothing, the set, and the light.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    },
    {
     "issue_ko": "하단 핸드폰 화면이 서의용 쪽으로 비스듬히 위를 향하지 않고 카메라에 거의 정면으로 보인다",
     "fix_en": "Tilt the phone backward so the screen faces upward toward the man's eyes. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우측 전화기 버튼의 숫자와 파란색 바인더 표지의 텍스트가 알아볼 수 없게 뭉개져 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "프롬프트가 명시한 '사진' 대신, 핸드폰 화면에 사람의 단순한 검은색 실루엣만 표시되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙 서의용의 얼굴이 캐릭터 레퍼런스와 이목구비·골격이 다른 사람으로 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "서의용이 레퍼런스 의상인 남색 비니를 쓰지 않아 짧은 머리가 그대로 드러나 있다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 근경의 강주철이 스케치처럼 머리·어깨로 크게 잡히지 않고 어두운 가장자리만 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "서의용 얼굴이 타이트 클로즈업으로 상부 중앙을 채우지 못하고 상반신과 책상·사무실 배경이 과도하게 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "하단 핸드폰 화면이 서의용 쪽으로 비스듬히 위를 향하지 않고 카메라에 거의 정면으로 보인다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트가 명시한 '사진' 대신, 핸드폰 화면에 사람의 단순한 검은색 실루엣만 표시되어 있습니다.",
    "fix_en": "Replace the black silhouette on the phone screen with a clear color photograph. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "화면 중앙 서의용의 얼굴이 캐릭터 레퍼런스와 이목구비·골격이 다른 사람으로 보인다",
    "fix_en": "Adjust the central man's facial features to match the character reference identity. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "서의용 얼굴이 타이트 클로즈업으로 상부 중앙을 채우지 못하고 상반신과 책상·사무실 배경이 과도하게 보인다",
    "fix_en": "Scale up the central man to a tight close-up so his eyes occupy the upper center. Preserve the people present, their clothing, the set, and the light.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   },
   {
    "issue_ko": "하단 핸드폰 화면이 서의용 쪽으로 비스듬히 위를 향하지 않고 카메라에 거의 정면으로 보인다",
    "fix_en": "Tilt the phone backward so the screen faces upward toward the man's eyes. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S13sh3__bgfirst_bg.png",
   "bg_asset_id": "1f838a6d-5c90-42ea-a102-6318ae706256",
   "bg_record_key": "S13sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S13sh3::cine": {
  "applied": true,
  "fingerprint": "c897a32f3da064f978d81c1a36ff40597f6dd976cfd273772ee9af0e283da62c",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S13sh3_sel.png",
  "source_sha256": "e8638a56916efe896b809c756e202acbb00072990d3fe05c7c2d3ec94690357a",
  "file": "S13sh3_cine.png",
  "latency_ms": 11210
 },
 "S13sh6::signage": {
  "fp": "ea9f5a0eec5036f7",
  "inscriptions": [
   {
    "surface_native": "사무실 벽면 표지판",
    "text_native": "강력1반",
    "reason_ko": "대한민국의 경찰서 강력반 사무실 내부라는 설정을 명확히 보여주고 극의 사실감을 높이기 위해 필요한 부서 표지입니다."
   }
  ]
 },
 "S13sh6": {
  "input_fingerprint": "58a9c4d73d45ee70",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수화기를 본체 위에 올려놓은 채 강주철을 향해 시선을 고정한 서의용\n\nLOCATION (lock): At a desk telephone inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At desk height beside the telephone base, the camera finishes its advance and holds at a slight low angle, using the replaced receiver in the lower foreground as a diagonal lead toward 서의용's three-quarter profile. 서의용's hand has just released the receiver while his gaze remains fixed on 강주철, who is retained more softly across the office space at frame right.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, looks toward 강주철 across the office; 강주철 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 전화기 본체와 수화기 (Receiver replaced on its base) — The receiver rests on the base, whose long axis points diagonally from the lower foreground toward 서의용; used as Large lower-foreground lead that connects the ended call to 서의용's reaction; 사무실 책상 (In use in the office) — The desk edge recedes from lower left toward the background right; used as Separates the telephone foreground from 강주철's softer position beyond.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime ambient illumination maintains subdued neutral-to-cool tones and clear but unexaggerated facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 표지판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수화기를 본체 위에 올려놓은 채 강주철을 향해 시선을 고정한 서의용\n\nLOCATION (lock): At a desk telephone inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At desk height beside the telephone base, the camera finishes its advance and holds at a slight low angle, using the replaced receiver in the lower foreground as a diagonal lead toward 서의용's three-quarter profile. 서의용's hand has just released the receiver while his gaze remains fixed on 강주철, who is retained more softly across the office space at frame right.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, looks toward 강주철 across the office; 강주철 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 전화기 본체와 수화기 (Receiver replaced on its base) — The receiver rests on the base, whose long axis points diagonally from the lower foreground toward 서의용; used as Large lower-foreground lead that connects the ended call to 서의용's reaction; 사무실 책상 (In use in the office) — The desk edge recedes from lower left toward the background right; used as Separates the telephone foreground from 강주철's softer position beyond.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime ambient illumination maintains subdued neutral-to-cool tones and clear but unexaggerated facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 표지판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수화기를 본체 위에 올려놓은 채 강주철을 향해 시선을 고정한 서의용\n\nLOCATION (lock): At a desk telephone inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At desk height beside the telephone base, the camera finishes its advance and holds at a slight low angle, using the replaced receiver in the lower foreground as a diagonal lead toward 서의용's three-quarter profile. 서의용's hand has just released the receiver while his gaze remains fixed on 강주철, who is retained more softly across the office space at frame right.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, looks toward 강주철 across the office; 강주철 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 전화기 본체와 수화기 (Receiver replaced on its base) — The receiver rests on the base, whose long axis points diagonally from the lower foreground toward 서의용; used as Large lower-foreground lead that connects the ended call to 서의용's reaction; 사무실 책상 (In use in the office) — The desk edge recedes from lower left toward the background right; used as Separates the telephone foreground from 강주철's softer position beyond.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime ambient illumination maintains subdued neutral-to-cool tones and clear but unexaggerated facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 표지판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "서의용의 시선이 배경의 강주철이 아닌 전경의 전화기 수화기를 향하고 있음.",
    "built_space": "사무실 내부의 책상, 전화기 본체, 복사기 등이 카메라 시점에 맞춰 자연스럽게 배치됨.",
    "entities": "서의용의 인상착의(비니, 가죽 재킷 등)가 레퍼런스와 정확히 일치하며 배경에 강주철이 위치함. 벽면의 '강력1반' 텍스트가 정확함.",
    "hard_violations": [],
    "physics": "서의용의 손이 수화기를 짚고 있으며, 신체와 사물들이 지면과 가구에 의해 안정적으로 지지됨."
   },
   {
    "label": "B",
    "direction": "서의용의 시선이 배경 우측의 강주철을 정확히 향하고 있음.",
    "built_space": "책상과 전화기가 전경을 차지하며 사무실 가구들이 올바른 심도와 위치에 배치됨.",
    "entities": "서의용과 강주철의 배치는 좋으나, 화면 좌측 하단 전경에 프롬프트에 명시되지 않은 인물의 팔(파란 소매)이 등장함. 표지판 텍스트는 정확함.",
    "hard_violations": [
     "지시문에 없는 인물의 신체 부위 등장 (전경 좌측 하단의 파란색 소매를 입은 팔과 손)"
    ],
    "physics": "전경에 등장한 정체불명의 팔이 책상 면에 닿아 지지되고 있으며, 다른 인물들의 자세는 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 6,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "서의용의 시선이 강주철이 아닌 전화기를 향해 행동 지시를 일부 어겼으나, 금지된 요소나 치명적인 위반이 없어 최종 선택됨."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "시선 처리와 전반적인 구도는 우수하나, 화면 전경에 지시문에 없는 타인의 팔(파란색 소매)이 등장하는 치명적인 위반이 발생함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S13sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "서의용의 시선이 배경에 있는 강주철을 향하지 않고, 아래쪽에 있는 전화기를 내려다보고 있음.",
     "fix_en": "Redirect the left man's eyes and head angle so his gaze locks onto the standing man in the background right. Preserve both men's identities, clothing, overall postures, the desk telephone, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "서의용의 외양은 이전 장면(Previous Shot Still)과 동일하게 고정되어야 하므로 모자를 쓰지 않아야 하나, 파란색 비니를 착용하고 있음.",
     "fix_en": "Remove the blue beanie and render short black hair on the left man. Preserve his face, jacket, the background man, the office set, lighting, and framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "서의용 오른손이 전경 수화기를 아직 짚고 있어 막 내려놓은 상태가 아니다",
     "fix_en": "Redraw the left man's right hand slightly lifted above the telephone receiver, fingers detached from the plastic. Preserve his arm, jacket, the telephone, the desk, the background man, and framing.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "서의용의 시선이 배경에 있는 강주철을 향하지 않고, 아래쪽에 있는 전화기를 내려다보고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용의 외양은 이전 장면(Previous Shot Still)과 동일하게 고정되어야 하므로 모자를 쓰지 않아야 하나, 파란색 비니를 착용하고 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 중경의 서의용이 이전 샷에 없던 파란 비니를 써서 머리·의상 잠금과 다르다",
     "severity": "major"
    },
    {
     "issue_ko": "서의용의 시선이 오른쪽 배경의 강주철이 아니라 카메라를 향한다",
     "severity": "major"
    },
    {
     "issue_ko": "서의용 오른손이 전경 수화기를 아직 짚고 있어 막 내려놓은 상태가 아니다",
     "severity": "major"
    },
    {
     "issue_ko": "서의용이 3/4 측면이 아니라 거의 정면으로 렌즈를 보고 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "서의용의 외양은 이전 장면(Previous Shot Still)과 동일하게 고정되어야 하므로 모자를 쓰지 않아야 하나, 파란색 비니를 착용하고 있음.",
    "fix_en": "Remove the blue beanie and render short black hair on the left man. Preserve his face, jacket, the background man, the office set, lighting, and framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "서의용 오른손이 전경 수화기를 아직 짚고 있어 막 내려놓은 상태가 아니다",
    "fix_en": "Redraw the left man's right hand slightly lifted above the telephone receiver, fingers detached from the plastic. Preserve his arm, jacket, the telephone, the desk, the background man, and framing.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redirect the left man's eyes and head angle so his gaze locks onto the standing man in the background right. Preserve both men's identities, clothing, overall postures, the desk telephone, lighting, and framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 샷 룩 유지 지시를 어기고 캐릭터 레퍼런스의 비니를 그대로 착용한 점은 감점 요소이나, 수화기에 손을 얹은 채 강주철을 응시하는 측면 얼굴과 시선 방향을 프롬프트대로 완벽하게 구현하여 크게 우수합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "서의용이 강주철을 보지 않고 아래쪽 전화기를 내려다보고 있어 핵심 행동 지시를 위반했으며, B와 마찬가지로 이전 샷에서 없던 비니를 착용하여 룩 고정 지시를 어겼습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "서의용의 고개가 돌아가 있으며 시선이 프레임 우측 배경에 있는 강주철을 정확히 향하고 있음. 강주철 역시 서의용을 마주 보고 있음.",
      "built_space": "책상 위에 전화기가 올바른 방향으로 놓여 있으며, 우측 배경에 강주철이 위치함. 사무실 집기와 구조가 이전 샷과 일관됨.",
      "entities": "서의용, 강주철, 전화기, '강력1반' 표지판 모두 정확히 묘사됨. 단, 서의용이 이전 샷(모자 미착용)과 달리 비니를 착용하고 있음.",
      "hard_violations": [],
      "physics": "서의용의 손이 전화기 수화기 위에 중력에 맞게 자연스럽게 지지되어 있음."
     },
     {
      "label": "A",
      "direction": "서의용의 시선이 강주철이 아닌 아래쪽의 전화기 수화기를 향하고 있음. 강주철은 서의용을 바라봄.",
      "built_space": "책상 위에 전화기가 올바른 방향으로 놓여 있으며, 우측 배경에 강주철이 위치함. 사무실 집기와 구조가 이전 샷과 일관됨.",
      "entities": "서의용, 강주철, 전화기, '강력1반' 표지판이 존재함. B와 동일하게 이전 샷 룩에 없던 비니를 착용하고 있음.",
      "hard_violations": [],
      "physics": "서의용의 손이 전화기 수화기 위에 중력에 맞게 자연스럽게 지지되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 샷 룩 유지 지시를 어기고 캐릭터 레퍼런스의 비니를 그대로 착용한 점은 감점 요소이나, 수화기에 손을 얹은 채 강주철을 응시하는 측면 얼굴과 시선 방향을 프롬프트대로 완벽하게 구현하여 크게 우수합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "서의용이 강주철을 보지 않고 아래쪽 전화기를 내려다보고 있어 핵심 행동 지시를 위반했으며, B와 마찬가지로 이전 샷에서 없던 비니를 착용하여 룩 고정 지시를 어겼습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "서의용의 고개가 돌아가 있으며 시선이 프레임 우측 배경에 있는 강주철을 정확히 향하고 있음. 강주철 역시 서의용을 마주 보고 있음.",
      "built_space": "책상 위에 전화기가 올바른 방향으로 놓여 있으며, 우측 배경에 강주철이 위치함. 사무실 집기와 구조가 이전 샷과 일관됨.",
      "entities": "서의용, 강주철, 전화기, '강력1반' 표지판 모두 정확히 묘사됨. 단, 서의용이 이전 샷(모자 미착용)과 달리 비니를 착용하고 있음.",
      "hard_violations": [],
      "physics": "서의용의 손이 전화기 수화기 위에 중력에 맞게 자연스럽게 지지되어 있음."
     },
     {
      "label": "A",
      "direction": "서의용의 시선이 강주철이 아닌 아래쪽의 전화기 수화기를 향하고 있음. 강주철은 서의용을 바라봄.",
      "built_space": "책상 위에 전화기가 올바른 방향으로 놓여 있으며, 우측 배경에 강주철이 위치함. 사무실 집기와 구조가 이전 샷과 일관됨.",
      "entities": "서의용, 강주철, 전화기, '강력1반' 표지판이 존재함. B와 동일하게 이전 샷 룩에 없던 비니를 착용하고 있음.",
      "hard_violations": [],
      "physics": "서의용의 손이 전화기 수화기 위에 중력에 맞게 자연스럽게 지지되어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "이전 샷의 복장(비니 미착용) 상태를 유지하지 못한 공통적인 오류가 있으나, 수화기를 내려놓고 강주철을 향해 시선을 고정하는 프롬프트의 핵심 연출을 정확하게 구현하여 훨씬 우수합니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "서의용의 시선이 강주철이 아닌 아래쪽의 수화기를 향하고 있어 프롬프트가 요구한 핵심적인 시선 방향 지시를 완전히 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "서의용의 시선이 배경 우측에 서 있는 강주철의 얼굴을 정확히 향하고 있음. 손은 전경의 수화기 위에 올려져 있음.",
      "built_space": "사무실 책상이 전경의 전화기와 배경의 강주철 사이를 적절히 분리하고 있으며, 전화기가 프레임 우하단에서 대각선 방향으로 자연스럽게 배치됨.",
      "entities": "서의용과 강주철이 올바르게 등장하며 벽면 표지판에 '강력1반' 텍스트가 명확함. 다만 이전 샷(비니 미착용)의 복장 고정 지시를 어기고 캐릭터 레퍼런스의 비니를 착용함.",
      "hard_violations": [],
      "physics": "전화기 본체는 책상 위에, 수화기는 본체 위에 안정적으로 놓여 있으며, 서의용의 손이 수화기에 닿아 있는 물리적 접촉이 자연스러움."
     },
     {
      "label": "B",
      "direction": "서의용의 시선이 강주철을 향하지 않고 아래쪽에 놓인 전화기 수화기를 내려다보고 있어 샷의 방향성 지시를 위반함.",
      "built_space": "사무실 책상이 전경의 전화기와 배경의 강주철 사이를 분리하며 배치되었고, 카메라 앵글 및 전화기 배치가 지시와 일치함.",
      "entities": "서의용과 강주철이 등장하고 '강력1반' 표지판 텍스트가 명확하나, A와 마찬가지로 이전 샷의 복장 고정 지시(비니 미착용)를 어기고 비니를 착용함.",
      "hard_violations": [],
      "physics": "전화기 본체와 수화기의 지지 상태, 손의 접촉 등 물리적 묘사는 안정적임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 샷의 복장(비니 미착용) 상태를 유지하지 못한 공통적인 오류가 있으나, 수화기를 내려놓고 강주철을 향해 시선을 고정하는 프롬프트의 핵심 연출을 정확하게 구현하여 훨씬 우수합니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "서의용의 시선이 강주철이 아닌 아래쪽의 수화기를 향하고 있어 프롬프트가 요구한 핵심적인 시선 방향 지시를 완전히 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "서의용의 시선이 배경 우측에 서 있는 강주철의 얼굴을 정확히 향하고 있음. 손은 전경의 수화기 위에 올려져 있음.",
      "built_space": "사무실 책상이 전경의 전화기와 배경의 강주철 사이를 적절히 분리하고 있으며, 전화기가 프레임 우하단에서 대각선 방향으로 자연스럽게 배치됨.",
      "entities": "서의용과 강주철이 올바르게 등장하며 벽면 표지판에 '강력1반' 텍스트가 명확함. 다만 이전 샷(비니 미착용)의 복장 고정 지시를 어기고 캐릭터 레퍼런스의 비니를 착용함.",
      "hard_violations": [],
      "physics": "전화기 본체는 책상 위에, 수화기는 본체 위에 안정적으로 놓여 있으며, 서의용의 손이 수화기에 닿아 있는 물리적 접촉이 자연스러움."
     },
     {
      "label": "A",
      "direction": "서의용의 시선이 강주철을 향하지 않고 아래쪽에 놓인 전화기 수화기를 내려다보고 있어 샷의 방향성 지시를 위반함.",
      "built_space": "사무실 책상이 전경의 전화기와 배경의 강주철 사이를 분리하며 배치되었고, 카메라 앵글 및 전화기 배치가 지시와 일치함.",
      "entities": "서의용과 강주철이 등장하고 '강력1반' 표지판 텍스트가 명확하나, A와 마찬가지로 이전 샷의 복장 고정 지시(비니 미착용)를 어기고 비니를 착용함.",
      "hard_violations": [],
      "physics": "전화기 본체와 수화기의 지지 상태, 손의 접촉 등 물리적 묘사는 안정적임."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S13sh3"
  }
 },
 "S13sh6::cine": {
  "applied": true,
  "fingerprint": "9cf1c9f7a173ee10a76f0971becc1af976d45bf13d0b4f64dbd644b5e465c231",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S13sh6_sel.png",
  "source_sha256": "7d526f2af3b4533400ae9a18db32d1f4b1c86b59241d97eef1f1f99bb3f2598c",
  "file": "S13sh6_cine.png",
  "latency_ms": 10686
 },
 "S14sh4::signage": {
  "fp": "e228f2167f6c8be3",
  "inscriptions": [
   {
    "surface_native": "수사보고서 표지",
    "text_native": "드들강 피살 사건 수사보고서",
    "reason_ko": "전택수의 손이 가리키고 있는 서류가 '드들강 사건'의 공식 수사보고서임을 나타내기 위해 표지 제목이 필요합니다."
   }
  ]
 },
 "era_assess::50b9c945ecc977d7": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 경찰서 수사과장실 소파 및 테이블",
    "search_terms_native": [
     "경찰서 과장실 소파",
     "형사과장실 내부",
     "경찰서 응접실 테이블",
     "2000년대 경찰서"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 검색해야 하며, 영어 등 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 경찰서 간부실 특유의 무겁고 중후한 가죽 소파와 유리 상판 목재 테이블 세트는 일반적인 현대식 사무실 소파나 서구식 집기 디자인과 확연히 다릅니다."
   },
   {
    "subject_native": "2000년대 한국 경찰 수사보고서 서류철",
    "search_terms_native": [
     "경찰 수사보고서 표지",
     "사건기록 서류철",
     "경찰 수사기록"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 검색해야 하며, 영어 등 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 경찰에서 사용하는 황색 표지의 사건 기록철과 '수사보고' 서체, 관인 날인 등의 고유 양식은 범용적인 서류 봉투나 바인더 이미지와 크게 다릅니다."
   }
  ]
 },
 "era_ref::7818a8fbb75fc95a": {
  "subject": "2000년대 한국 경찰서 수사과장실 소파 및 테이블",
  "terms": [
   "경찰서 과장실 소파",
   "형사과장실 내부",
   "경찰서 응접실 테이블",
   "2000년대 경찰서"
  ],
  "queries": [
   [
    "2000년대 한국 경찰서 형사과장실 내부 소파 응접실 테이블",
    "2000년대 한국 경찰서 수사과장실 소파 테이블"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://cdn.gnnews.co.kr/news/photo/second/201204/img_131293_1.jpg",
  "picked_reason_ko": "실제 한국 경찰관서의 과장실·응접 공간으로 보이며, 2000년대에 흔한 청색 비닐 소파와 낮은 직사각형 테이블의 재료·형태·배치를 가장 선명하게 확인할 수 있다.",
  "sha256": "ee0c2cd8b5ddbcec6bbb78f12adcd6984267ae265809e58bc0303c6541594466",
  "file": "eraref_7818a8fbb75fc95a.png"
 },
 "S14sh4::bgfirst_bg": {
  "input_fingerprint": "bf84217560cb0ce6",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블 위의 드들강 사건 보고서 표지 위로 검지손가락을 얹은 전택수의 손\n\nLOCATION (lock): At the sofa table inside the investigation chief's office, with the cold-case report laid out between the seats.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just above table height beside 전택수, a steep diagonal close view completes the dolly toward his hand and isolates the instant his index finger comes to rest on the case-report cover. The forearm enters from the upper edge while the report remains centered with a narrow band of tabletop around it for scale and context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 표지 (Resting among the unresolved-case reports on the table) — The front cover faces upward and is readable to camera, showing that it is the 드들강 case report; used as Primary evidence detail beneath 전택수's resting index finger; 테이블 상판 (Supporting the case documents); used as Supplies a narrow contextual border around the report and supports the hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral naturalistic daytime illumination renders the report and hand legibly without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 경찰서 수사과장실 소파 및 테이블: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블 위의 드들강 사건 보고서 표지 위로 검지손가락을 얹은 전택수의 손\n\nLOCATION (lock): At the sofa table inside the investigation chief's office, with the cold-case report laid out between the seats.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just above table height beside 전택수, a steep diagonal close view completes the dolly toward his hand and isolates the instant his index finger comes to rest on the case-report cover. The forearm enters from the upper edge while the report remains centered with a narrow band of tabletop around it for scale and context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 표지 (Resting among the unresolved-case reports on the table) — The front cover faces upward and is readable to camera, showing that it is the 드들강 case report; used as Primary evidence detail beneath 전택수's resting index finger; 테이블 상판 (Supporting the case documents); used as Supplies a narrow contextual border around the report and supports the hand.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral naturalistic daytime illumination renders the report and hand legibly without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 경찰서 수사과장실 소파 및 테이블: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S14sh4__bgfirst_bg.png",
  "asset_id": "1985b753-cccb-45d3-a955-6c9a72d91f9b",
  "input_asset_ids": [
   "66cc917f-fa45-43c0-88aa-8abcb32c8311",
   "36cfacb8-cfbf-4c8b-a3d1-35be6baa2cb8"
  ],
  "era_research": {
   "subject": "2000년대 한국 경찰서 수사과장실 소파 및 테이블",
   "queries": [
    [
     "2000년대 한국 경찰서 형사과장실 내부 소파 응접실 테이블",
     "2000년대 한국 경찰서 수사과장실 소파 테이블"
    ]
   ],
   "picked_url": "https://cdn.gnnews.co.kr/news/photo/second/201204/img_131293_1.jpg",
   "sha256": "ee0c2cd8b5ddbcec6bbb78f12adcd6984267ae265809e58bc0303c6541594466",
   "file": "eraref_7818a8fbb75fc95a.png"
  }
 },
 "S14sh4": {
  "input_fingerprint": "64e0fcd5e0562a4b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위의 드들강 사건 보고서 표지 위로 검지손가락을 얹은 전택수의 손\n\nLOCATION (lock): At the sofa table inside the investigation chief's office, with the cold-case report laid out between the seats. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above table height beside 전택수, a steep diagonal close view completes the dolly toward his hand and isolates the instant his index finger comes to rest on the case-report cover. The forearm enters from the upper edge while the report remains centered with a narrow band of tabletop around it for scale and context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 표지 (Resting among the unresolved-case reports on the table) — The front cover faces upward and is readable to camera, showing that it is the 드들강 case report; used as Primary evidence detail beneath 전택수's resting index finger; 테이블 상판 (Supporting the case documents); used as Supplies a narrow contextual border around the report and supports the hand.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral naturalistic daytime illumination renders the report and hand legibly without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Drdeul River case report remains laid out among the unsolved-case status documents under Taeksu's pointing hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사보고서 표지: \"드들강 피살 사건 수사보고서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위의 드들강 사건 보고서 표지 위로 검지손가락을 얹은 전택수의 손\n\nLOCATION (lock): At the sofa table inside the investigation chief's office, with the cold-case report laid out between the seats. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above table height beside 전택수, a steep diagonal close view completes the dolly toward his hand and isolates the instant his index finger comes to rest on the case-report cover. The forearm enters from the upper edge while the report remains centered with a narrow band of tabletop around it for scale and context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 표지 (Resting among the unresolved-case reports on the table) — The front cover faces upward and is readable to camera, showing that it is the 드들강 case report; used as Primary evidence detail beneath 전택수's resting index finger; 테이블 상판 (Supporting the case documents); used as Supplies a narrow contextual border around the report and supports the hand.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral naturalistic daytime illumination renders the report and hand legibly without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Drdeul River case report remains laid out among the unsolved-case status documents under Taeksu's pointing hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사보고서 표지: \"드들강 피살 사건 수사보고서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위의 드들강 사건 보고서 표지 위로 검지손가락을 얹은 전택수의 손\n\nLOCATION (lock): At the sofa table inside the investigation chief's office, with the cold-case report laid out between the seats. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above table height beside 전택수, a steep diagonal close view completes the dolly toward his hand and isolates the instant his index finger comes to rest on the case-report cover. The forearm enters from the upper edge while the report remains centered with a narrow band of tabletop around it for scale and context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 표지 (Resting among the unresolved-case reports on the table) — The front cover faces upward and is readable to camera, showing that it is the 드들강 case report; used as Primary evidence detail beneath 전택수's resting index finger; 테이블 상판 (Supporting the case documents); used as Supplies a narrow contextual border around the report and supports the hand.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral naturalistic daytime illumination renders the report and hand legibly without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Drdeul River case report remains laid out among the unsolved-case status documents under Taeksu's pointing hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사보고서 표지: \"드들강 피살 사건 수사보고서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S14sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S14sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L17B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 좁은 인서트 샷 프레이밍을 정확히 따랐으며, 금지된 추가 텍스트 없이 단일 지정 텍스트와 손가락이 닿는 물리적 액션을 훌륭히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레이밍이 지시보다 넓고 손가락이 표지에서 떠 있으며, 금지된 추가 텍스트가 다수 노출되어 지침을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "검지손가락이 중앙의 보고서를 향하고 있으나 표지에 닿지 않고 허공에 떠 있음.",
      "built_space": "어두운 목재 테이블과 검은 가죽 소파가 배치되어 레퍼런스의 사무실 환경과 일치함.",
      "entities": "레퍼런스와 일치하는 남색 재킷 및 체크 셔츠 소매. 여러 보고서 표지에 '강력 사건 보고서' 등 지정되지 않은 한글 텍스트가 선명함.",
      "hard_violations": [
       "지시되지 않은 읽을 수 있는 텍스트(다른 보고서 제목들) 임의 추가"
      ],
      "physics": "손과 팔은 테이블 상공에 떠 있으며, 뻗은 팔의 근력으로 자세를 유지하고 있음."
     },
     {
      "label": "B",
      "direction": "검지손가락이 지정된 수사보고서 표지 위를 정확하게 짚고 있음.",
      "built_space": "목재 테이블, 검은 가죽 소파, 수납장 등 레퍼런스의 요소들이 프레임 내에 알맞게 자리함.",
      "entities": "전택수의 지정된 복장(남색 재킷, 체크 셔츠)이 일치하며, 지시된 텍스트만 단독으로 인쇄된 보고서가 중앙에 놓임.",
      "hard_violations": [],
      "physics": "손가락 끝이 책상 위의 보고서 표지에 닿아 명확하게 하중을 지지받고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 좁은 인서트 샷 프레이밍을 정확히 따랐으며, 금지된 추가 텍스트 없이 단일 지정 텍스트와 손가락이 닿는 물리적 액션을 훌륭히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레이밍이 지시보다 넓고 손가락이 표지에서 떠 있으며, 금지된 추가 텍스트가 다수 노출되어 지침을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "검지손가락이 중앙의 보고서를 향하고 있으나 표지에 닿지 않고 허공에 떠 있음.",
      "built_space": "어두운 목재 테이블과 검은 가죽 소파가 배치되어 레퍼런스의 사무실 환경과 일치함.",
      "entities": "레퍼런스와 일치하는 남색 재킷 및 체크 셔츠 소매. 여러 보고서 표지에 '강력 사건 보고서' 등 지정되지 않은 한글 텍스트가 선명함.",
      "hard_violations": [
       "지시되지 않은 읽을 수 있는 텍스트(다른 보고서 제목들) 임의 추가"
      ],
      "physics": "손과 팔은 테이블 상공에 떠 있으며, 뻗은 팔의 근력으로 자세를 유지하고 있음."
     },
     {
      "label": "B",
      "direction": "검지손가락이 지정된 수사보고서 표지 위를 정확하게 짚고 있음.",
      "built_space": "목재 테이블, 검은 가죽 소파, 수납장 등 레퍼런스의 요소들이 프레임 내에 알맞게 자리함.",
      "entities": "전택수의 지정된 복장(남색 재킷, 체크 셔츠)이 일치하며, 지시된 텍스트만 단독으로 인쇄된 보고서가 중앙에 놓임.",
      "hard_violations": [],
      "physics": "손가락 끝이 책상 위의 보고서 표지에 닿아 명확하게 하중을 지지받고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 텍스트를 정확히 표출하고 인서트 클로즈업 샷의 의도를 잘 살렸으며 불필요한 요소가 없어 지침에 부합함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트에서 엄격히 금지한 '지정되지 않은 읽을 수 있는 텍스트'가 주변 문서들에 다수 생성되어 지침을 크게 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "검지손가락이 테이블 위에 놓인 보고서 표지의 텍스트를 정확히 짚고 있음.",
      "built_space": "기준 사진과 일치하는 어두운 목재 테이블 상판과 배경의 검은색 가죽 소파가 보임.",
      "entities": "전택수의 인상착의(남색 재킷, 체크 셔츠)와 일치하는 팔과 손, '드들강 피살 사건 수사보고서'가 정확히 적힌 단일 보고서.",
      "hard_violations": [],
      "physics": "손가락 끝이 테이블 위 보고서에 맞닿아 팔을 자연스럽게 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "검지손가락이 중앙에 놓인 보고서 표지의 여백을 가리키고 있음.",
      "built_space": "기준 사진과 일치하는 목재 테이블과 양옆의 검은색 가죽 소파가 확인됨.",
      "entities": "전택수의 의상과 일치하는 소매, 지정된 문구가 적힌 중앙의 보고서 및 주변에 정렬된 다수의 다른 문서들.",
      "hard_violations": [
       "지시하지 않은 임의의 텍스트('강력 사건 보고서', '미제 사건 보고서' 등)가 다수 생성됨 (leaked text 위반)"
      ],
      "physics": "손가락 끝이 보고서 표지에 닿아 팔의 자세를 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 텍스트를 정확히 표출하고 인서트 클로즈업 샷의 의도를 잘 살렸으며 불필요한 요소가 없어 지침에 부합함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트에서 엄격히 금지한 '지정되지 않은 읽을 수 있는 텍스트'가 주변 문서들에 다수 생성되어 지침을 크게 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "검지손가락이 테이블 위에 놓인 보고서 표지의 텍스트를 정확히 짚고 있음.",
      "built_space": "기준 사진과 일치하는 어두운 목재 테이블 상판과 배경의 검은색 가죽 소파가 보임.",
      "entities": "전택수의 인상착의(남색 재킷, 체크 셔츠)와 일치하는 팔과 손, '드들강 피살 사건 수사보고서'가 정확히 적힌 단일 보고서.",
      "hard_violations": [],
      "physics": "손가락 끝이 테이블 위 보고서에 맞닿아 팔을 자연스럽게 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "검지손가락이 중앙에 놓인 보고서 표지의 여백을 가리키고 있음.",
      "built_space": "기준 사진과 일치하는 목재 테이블과 양옆의 검은색 가죽 소파가 확인됨.",
      "entities": "전택수의 의상과 일치하는 소매, 지정된 문구가 적힌 중앙의 보고서 및 주변에 정렬된 다수의 다른 문서들.",
      "hard_violations": [
       "지시하지 않은 임의의 텍스트('강력 사건 보고서', '미제 사건 보고서' 등)가 다수 생성됨 (leaked text 위반)"
      ],
      "physics": "손가락 끝이 보고서 표지에 닿아 팔의 자세를 지지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "검지손가락이 중앙의 보고서를 향하고 있으나 표지에 닿지 않고 허공에 떠 있음.",
    "built_space": "어두운 목재 테이블과 검은 가죽 소파가 배치되어 레퍼런스의 사무실 환경과 일치함.",
    "entities": "레퍼런스와 일치하는 남색 재킷 및 체크 셔츠 소매. 여러 보고서 표지에 '강력 사건 보고서' 등 지정되지 않은 한글 텍스트가 선명함.",
    "hard_violations": [
     "지시되지 않은 읽을 수 있는 텍스트(다른 보고서 제목들) 임의 추가"
    ],
    "physics": "손과 팔은 테이블 상공에 떠 있으며, 뻗은 팔의 근력으로 자세를 유지하고 있음."
   },
   {
    "label": "B",
    "direction": "검지손가락이 지정된 수사보고서 표지 위를 정확하게 짚고 있음.",
    "built_space": "목재 테이블, 검은 가죽 소파, 수납장 등 레퍼런스의 요소들이 프레임 내에 알맞게 자리함.",
    "entities": "전택수의 지정된 복장(남색 재킷, 체크 셔츠)이 일치하며, 지시된 텍스트만 단독으로 인쇄된 보고서가 중앙에 놓임.",
    "hard_violations": [],
    "physics": "손가락 끝이 책상 위의 보고서 표지에 닿아 명확하게 하중을 지지받고 있음."
   }
  ],
  "totals": {
   "A": 6,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 좁은 인서트 샷 프레이밍을 정확히 따랐으며, 금지된 추가 텍스트 없이 단일 지정 텍스트와 손가락이 닿는 물리적 액션을 훌륭히 구현했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "프레이밍이 지시보다 넓고 손가락이 표지에서 떠 있으며, 금지된 추가 텍스트가 다수 노출되어 지침을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L17B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "문서 위에 얹힌 손이 우측 상단에서 뻗어 나온 오른팔에 연결되어 있음에도 불구하고, 엄지손가락이 오른쪽에 위치하여 해부학적으로 왼손의 형태를 띠는 치명적인 오류가 있습니다.",
     "fix_en": "Redraw the pointing hand as an anatomically correct right hand, placing the thumb on the left side (closest to the inner wrist) and the curled middle, ring, and pinky fingers on the right side. Keep the plaid cuff, blue jacket, wooden table, case report with its printed text, and the background exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "문서 위에 얹힌 손이 우측 상단에서 뻗어 나온 오른팔에 연결되어 있음에도 불구하고, 엄지손가락이 오른쪽에 위치하여 해부학적으로 왼손의 형태를 띠는 치명적인 오류가 있습니다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the pointing hand as an anatomically correct right hand, placing the thumb on the left side (closest to the inner wrist) and the curled middle, ring, and pinky fingers on the right side. Keep the plaid cuff, blue jacket, wooden table, case report with its printed text, and the background exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "해부학적으로 자연스러운 손을 묘사했으며, 지정된 텍스트, 프레이밍, 의상 지침을 완벽하게 충족한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "지정된 텍스트와 프레이밍은 잘 따랐으나, 손가락(중지)의 해부학적 구조가 뭉개져 하드 위반에 해당합니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (손가락 중지가 기형적으로 짧고 뭉개짐)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.429,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.179,
      "B": 1.75
     },
     "violations": {
      "A": [
       "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (손가락 중지가 기형적으로 짧고 뭉개짐)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "해부학적으로 자연스러운 손을 묘사했으며, 지정된 텍스트, 프레이밍, 의상 지침을 완벽하게 충족한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "지정된 텍스트와 프레이밍은 잘 따랐으나, 손가락(중지)의 해부학적 구조가 뭉개져 하드 위반에 해당합니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (손가락 중지가 기형적으로 짧고 뭉개짐)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A의 기형적인 손가락 구조를 정상적으로 수정하였으며, 지정된 텍스트와 카메라 프레이밍을 모두 정확히 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "검지와 중지가 하나로 융합된 심각한 해부학적 오류가 있어 실격임."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "검지손가락이 테이블 위의 보고서 표지를 향해 닿아 있음.",
      "built_space": "아웃포커싱된 소파와 수납장이 배경에 있으며, 테이블 상판 위에 서류가 놓여 있음.",
      "entities": "전택수의 의상(체크 셔츠와 남색 재킷)을 입은 팔, 지정된 문구가 정확히 적힌 보고서 표지.",
      "hard_violations": [],
      "physics": "보고서는 테이블 상판에, 손은 보고서 표지 위에 물리적으로 자연스럽게 지지됨."
     },
     {
      "label": "A",
      "direction": "검지손가락이 테이블 위의 보고서 표지를 향해 닿아 있음.",
      "built_space": "아웃포커싱된 소파와 수납장이 배경에 있으며, 테이블 상판 위에 서류가 놓여 있음.",
      "entities": "전택수의 의상(체크 셔츠와 남색 재킷)을 입은 팔, 지정된 문구가 정확히 적힌 보고서 표지.",
      "hard_violations": [
       "물리적으로 불가능한 해부학 구조 (중지와 검지손가락이 기형적으로 융합됨)"
      ],
      "physics": "보고서는 테이블 상판에, 손은 보고서 표지 위에 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "A의 기형적인 손가락 구조를 정상적으로 수정하였으며, 지정된 텍스트와 카메라 프레이밍을 모두 정확히 구현함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "검지와 중지가 하나로 융합된 심각한 해부학적 오류가 있어 실격임."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "검지손가락이 테이블 위의 보고서 표지를 향해 닿아 있음.",
      "built_space": "아웃포커싱된 소파와 수납장이 배경에 있으며, 테이블 상판 위에 서류가 놓여 있음.",
      "entities": "전택수의 의상(체크 셔츠와 남색 재킷)을 입은 팔, 지정된 문구가 정확히 적힌 보고서 표지.",
      "hard_violations": [],
      "physics": "보고서는 테이블 상판에, 손은 보고서 표지 위에 물리적으로 자연스럽게 지지됨."
     },
     {
      "label": "B",
      "direction": "검지손가락이 테이블 위의 보고서 표지를 향해 닿아 있음.",
      "built_space": "아웃포커싱된 소파와 수납장이 배경에 있으며, 테이블 상판 위에 서류가 놓여 있음.",
      "entities": "전택수의 의상(체크 셔츠와 남색 재킷)을 입은 팔, 지정된 문구가 정확히 적힌 보고서 표지.",
      "hard_violations": [
       "물리적으로 불가능한 해부학 구조 (중지와 검지손가락이 기형적으로 융합됨)"
      ],
      "physics": "보고서는 테이블 상판에, 손은 보고서 표지 위에 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1186,
     "B": 1753
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S14sh4__bgfirst_bg.png",
   "bg_asset_id": "1985b753-cccb-45d3-a955-6c9a72d91f9b",
   "bg_record_key": "S14sh4::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S14sh4::cine": {
  "applied": true,
  "fingerprint": "d4a81a57691fcf84de747155b43dabdcf6a09d30ece90d3a9d0b0fc7fe89014b",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S14sh4_sel.png",
  "source_sha256": "b216f826f3bad425fdb36dabde84ce40ea5efb1d42b0185aaf650aadb89bed95",
  "file": "S14sh4_cine.png",
  "latency_ms": 11450
 },
 "S14sh7::signage": {
  "fp": "c1fcb908c2535067",
  "inscriptions": [
   {
    "surface_native": "사무실 명패",
    "text_native": "수사과장",
    "reason_ko": "수사과장실 소파 접견 공간이라는 극 중 배경을 사실적으로 묘사하기 위해 벽면이나 문에 부착된 직책 표지판이 필요합니다."
   }
  ]
 },
 "S14sh7": {
  "input_fingerprint": "16ca800b226f5713",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 서의용과 강주철 쪽으로 단호한 시선을 던지는 전택수\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From seated chest height close to 전택수's three-quarter side, the completed upward tilt holds slightly below his eyeline as he raises his head from the report. 전택수 occupies the left-center foreground with one forearm still oriented toward the table, while 서의용 and 강주철 remain separately posed across the conversational space, receiving his firm proposal rather than facing the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward the two detectives; 서의용 in the middle-right of the frame, background; 강주철 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 (Selected from the unresolved-case status report) — Its front cover lies upward on the table below 전택수's arm, with the case identification facing the seated group; used as Keeps the reopened case physically present beneath the exchange; 소파와 테이블 (Occupied for the discussion) — The table recedes between opposing seats, placing 전택수 on one side and the detectives on the other; used as Defines the shared seated discussion axis between 전택수 and the detectives.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior illumination with a neutral-to-cool palette and moderate contrast supports the procedural gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 서의용과 강주철 쪽으로 단호한 시선을 던지는 전택수\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From seated chest height close to 전택수's three-quarter side, the completed upward tilt holds slightly below his eyeline as he raises his head from the report. 전택수 occupies the left-center foreground with one forearm still oriented toward the table, while 서의용 and 강주철 remain separately posed across the conversational space, receiving his firm proposal rather than facing the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward the two detectives; 서의용 in the middle-right of the frame, background; 강주철 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 (Selected from the unresolved-case status report) — Its front cover lies upward on the table below 전택수's arm, with the case identification facing the seated group; used as Keeps the reopened case physically present beneath the exchange; 소파와 테이블 (Occupied for the discussion) — The table recedes between opposing seats, placing 전택수 on one side and the detectives on the other; used as Defines the shared seated discussion axis between 전택수 and the detectives.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior illumination with a neutral-to-cool palette and moderate contrast supports the procedural gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 서의용과 강주철 쪽으로 단호한 시선을 던지는 전택수\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From seated chest height close to 전택수's three-quarter side, the completed upward tilt holds slightly below his eyeline as he raises his head from the report. 전택수 occupies the left-center foreground with one forearm still oriented toward the table, while 서의용 and 강주철 remain separately posed across the conversational space, receiving his firm proposal rather than facing the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward the two detectives; 서의용 in the middle-right of the frame, background; 강주철 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 드들강 사건 보고서 (Selected from the unresolved-case status report) — Its front cover lies upward on the table below 전택수's arm, with the case identification facing the seated group; used as Keeps the reopened case physically present beneath the exchange; 소파와 테이블 (Occupied for the discussion) — The table recedes between opposing seats, placing 전택수 on one side and the detectives on the other; used as Defines the shared seated discussion axis between 전택수 and the detectives.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained daytime interior illumination with a neutral-to-cool palette and moderate contrast supports the procedural gravity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.3
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.05
   },
   "violations": {
    "B": [
     "[gemini-pro] 복제된 인물 (배경에 앉은 두 형사의 얼굴이 완전히 똑같이 생김)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 1050
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "요구된 카메라 앵글과 프레이밍을 정확히 지켰으며, 전택수의 복장 디테일과 소품의 텍스트(명패, 보고서)까지 흠잡을 데 없이 완벽하게 구현된 결과물입니다."
   },
   {
    "label": "B",
    "score": 1050,
    "verdict_ko": "배경에 앉은 두 형사의 얼굴이 동일 인물처럼 복제되었고, 전택수의 재킷에 적힌 영문 스펠링(POLICS)에 오타가 발생하여 감점되었습니다.  ★위반: [gemini-pro] 복제된 인물 (배경에 앉은 두 형사의 얼굴이 완전히 똑같이 생김)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S14sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "테이블 위 보고서의 제목 텍스트가 프롬프트의 지시와 달리 맞은편 형사들이 아닌 전택수 방향을 향해 정방향으로 놓여 있습니다.",
     "fix_en": "Rotate the report 180 degrees so the text faces the detectives in the background. Keep the characters, poses, lighting, and framing exactly the same.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "전택수의 우측 어깨(화면 좌측) 패치가 레퍼런스의 경찰 마크가 아닌 정체불명의 흰색 기호로 잘못 그려져 의상 설정과 어긋납니다.",
     "fix_en": "Change the patch on the foreground character's left-screen shoulder to a standard police emblem. Keep the characters, poses, lighting, and framing exactly the same.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "전택수의 벨트가 레퍼런스에 있는 전술용 벨트가 아닌 일반 가죽 벨트로 다르게 그려졌습니다.",
     "fix_en": "Replace the leather belt with an olive tactical belt featuring a plastic buckle. Keep the characters, poses, lighting, and framing exactly the same.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "화면 우측 중간에 앉은 회색 정장 남성의 깍지 낀 손가락 형태가 서로 뭉개져 해부학적으로 부정확합니다.",
     "fix_en": "Redraw the clasped hands of the man in the grey suit to have anatomically correct, distinct fingers. Keep the characters, poses, lighting, and framing exactly the same.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "드들강 사건 보고서가 전택수 팔 아래가 아니라 테이블 오른쪽에 멀리 떨어져 있다",
     "fix_en": "Move the report folder to the left so it rests underneath the foreground character's forearm. Keep the characters, poses, lighting, and framing exactly the same.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "전택수 재킷 가슴 명찰 글자가 SALIE S로 깨져 있다",
     "fix_en": "Blur the text on the right chest patch so the letters are unreadable. Keep the characters, poses, lighting, and framing exactly the same.",
     "severity": "minor",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "테이블 위 보고서의 제목 텍스트가 프롬프트의 지시와 달리 맞은편 형사들이 아닌 전택수 방향을 향해 정방향으로 놓여 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수의 우측 어깨(화면 좌측) 패치가 레퍼런스의 경찰 마크가 아닌 정체불명의 흰색 기호로 잘못 그려져 의상 설정과 어긋납니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수의 벨트가 레퍼런스에 있는 전술용 벨트가 아닌 일반 가죽 벨트로 다르게 그려졌습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "화면 우측 중간에 앉은 회색 정장 남성의 깍지 낀 손가락 형태가 서로 뭉개져 해부학적으로 부정확합니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "드들강 사건 보고서가 전택수 팔 아래가 아니라 테이블 오른쪽에 멀리 떨어져 있다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수 재킷 가슴 명찰 글자가 SALIE S로 깨져 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 6,
  "fix_severity_skipped": [
   {
    "issue_ko": "테이블 위 보고서의 제목 텍스트가 프롬프트의 지시와 달리 맞은편 형사들이 아닌 전택수 방향을 향해 정방향으로 놓여 있습니다.",
    "fix_en": "Rotate the report 180 degrees so the text faces the detectives in the background. Keep the characters, poses, lighting, and framing exactly the same.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "전택수의 우측 어깨(화면 좌측) 패치가 레퍼런스의 경찰 마크가 아닌 정체불명의 흰색 기호로 잘못 그려져 의상 설정과 어긋납니다.",
    "fix_en": "Change the patch on the foreground character's left-screen shoulder to a standard police emblem. Keep the characters, poses, lighting, and framing exactly the same.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "전택수의 벨트가 레퍼런스에 있는 전술용 벨트가 아닌 일반 가죽 벨트로 다르게 그려졌습니다.",
    "fix_en": "Replace the leather belt with an olive tactical belt featuring a plastic buckle. Keep the characters, poses, lighting, and framing exactly the same.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "화면 우측 중간에 앉은 회색 정장 남성의 깍지 낀 손가락 형태가 서로 뭉개져 해부학적으로 부정확합니다.",
    "fix_en": "Redraw the clasped hands of the man in the grey suit to have anatomically correct, distinct fingers. Keep the characters, poses, lighting, and framing exactly the same.",
    "severity": "minor",
    "observation_index": 3
   },
   {
    "issue_ko": "드들강 사건 보고서가 전택수 팔 아래가 아니라 테이블 오른쪽에 멀리 떨어져 있다",
    "fix_en": "Move the report folder to the left so it rests underneath the foreground character's forearm. Keep the characters, poses, lighting, and framing exactly the same.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "전택수 재킷 가슴 명찰 글자가 SALIE S로 깨져 있다",
    "fix_en": "Blur the text on the right chest patch so the letters are unreadable. Keep the characters, poses, lighting, and framing exactly the same.",
    "severity": "minor",
    "observation_index": 5
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh4"
  }
 },
 "S14sh7::cine": {
  "applied": true,
  "fingerprint": "b8d12100f23ef3d6fd7cb947bfd5c23731d1f1daa2d5ac1161e7e2cdd8d33c60",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S14sh7_sel.png",
  "source_sha256": "dccd852720b93acaa0b09d5b15e33c43a9a928822834ec7d539f53d4bafa720e",
  "file": "S14sh7_cine.png",
  "latency_ms": 10671
 },
 "S15sh4::signage": {
  "fp": "fae8e7201bbd9d0d",
  "inscriptions": [
   {
    "surface_native": "종이 상자 앞면 라벨",
    "text_native": "수사기록",
    "reason_ko": "경찰서 서류 보관실에 보관된 상자 속 서류들이 경찰의 공식 수사기록임을 보여주기 위해 상자 라벨에 한글 표기가 필요합니다."
   }
  ]
 },
 "era_assess::6a9d3677e4c3e15f": {
  "subjects": [
   {
    "subject_native": "대한민국 경찰서 문서고 (2000년대)",
    "search_terms_native": [
     "경찰서 문서고",
     "관공서 모빌랙",
     "정부보존상자",
     "문서고 서가"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "한국 경찰서와 관공서 문서고는 규격화된 철제 모빌랙(이동식 서가)과 한국 특유의 종이 정부보존상자를 사용하므로 일반적인 서구식 파일 캐비닛과 형태가 전혀 다릅니다."
   }
  ]
 },
 "era_ref::bf2cc5a419ca8f71": {
  "subject": "대한민국 경찰서 문서고 (2000년대)",
  "terms": [
   "경찰서 문서고",
   "관공서 모빌랙",
   "정부보존상자",
   "문서고 서가"
  ],
  "queries": [
   [
    "2000년대 대한민국 경찰서 문서고 관공서 모빌랙 정부보존상자",
    "경찰서 문서고 서가 이동식 모빌랙 문서보존상자"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://dimg.donga.com/wps/NEWS/IMAGE/2018/08/29/91709914.5.jpg",
  "picked_reason_ko": "한국어 분류표가 부착된 이동식 문서 보존서가 통로 양쪽에 선명하게 보여, 2000년대 대한민국 경찰서 문서고의 구조와 설비를 가장 잘 참고할 수 있다.",
  "sha256": "868164ed20719cec269caf05b5294f58cac1638e1cae08a482fb8372f93d305c",
  "file": "eraref_bf2cc5a419ca8f71.png"
 },
 "S15sh4::bgfirst_bg": {
  "input_fingerprint": "df26d56ff8fe3115",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 강주철의 손에 들린 낡은 종이 상자 안, 빽빽하게 들어찬 서류 뭉치들의 모습\n\nLOCATION (lock): Between the document shelves inside the police records room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Directly above and slightly forward of 강주철's hands, the final inward move settles into a steep top-down insert of the open box, with his hands bracing opposite edges. The tightly packed report bundles occupy the center while a narrow portion of the archive aisle and shelving remains visible around the box as scale reference.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 낡은 종이 서류 상자 (Open and filled with reports) — The open top faces upward toward the camera, with 강주철's hands holding opposite sides; used as Primary reveal containing the recovered case records; 서류 뭉치들 (Tightly packed inside the box) — The document edges and partially visible cover faces point upward inside the box; no single report face is isolated for reading; used as Creates dense internal lines and shows the volume of archived material; 자료실 책장 (Holding archived case records) — Shelf fronts run along the outer sides of the frame around the box; used as Provides contextual edges around the overhead insert.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime archive-room illumination keeps the subdued palette and documentary clarity without embellishment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 경찰서 문서고 (2000년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 강주철의 손에 들린 낡은 종이 상자 안, 빽빽하게 들어찬 서류 뭉치들의 모습\n\nLOCATION (lock): Between the document shelves inside the police records room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Directly above and slightly forward of 강주철's hands, the final inward move settles into a steep top-down insert of the open box, with his hands bracing opposite edges. The tightly packed report bundles occupy the center while a narrow portion of the archive aisle and shelving remains visible around the box as scale reference.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 낡은 종이 서류 상자 (Open and filled with reports) — The open top faces upward toward the camera, with 강주철's hands holding opposite sides; used as Primary reveal containing the recovered case records; 서류 뭉치들 (Tightly packed inside the box) — The document edges and partially visible cover faces point upward inside the box; no single report face is isolated for reading; used as Creates dense internal lines and shows the volume of archived material; 자료실 책장 (Holding archived case records) — Shelf fronts run along the outer sides of the frame around the box; used as Provides contextual edges around the overhead insert.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime archive-room illumination keeps the subdued palette and documentary clarity without embellishment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 경찰서 문서고 (2000년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S15sh4__bgfirst_bg.png",
  "asset_id": "1ebbd80a-6266-42b1-a783-cb5ad300ecdf",
  "input_asset_ids": [
   "56c4be71-abbe-4eab-a977-33be300df45a",
   "7b34d5b4-7db6-40ce-bc0c-c76574a63d76"
  ],
  "era_research": {
   "subject": "대한민국 경찰서 문서고 (2000년대)",
   "queries": [
    [
     "2000년대 대한민국 경찰서 문서고 관공서 모빌랙 정부보존상자",
     "경찰서 문서고 서가 이동식 모빌랙 문서보존상자"
    ]
   ],
   "picked_url": "https://dimg.donga.com/wps/NEWS/IMAGE/2018/08/29/91709914.5.jpg",
   "sha256": "868164ed20719cec269caf05b5294f58cac1638e1cae08a482fb8372f93d305c",
   "file": "eraref_bf2cc5a419ca8f71.png"
  }
 },
 "S15sh4": {
  "input_fingerprint": "512d3d9c11611030",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철의 손에 들린 낡은 종이 상자 안, 빽빽하게 들어찬 서류 뭉치들의 모습\n\nLOCATION (lock): Between the document shelves inside the police records room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Directly above and slightly forward of 강주철's hands, the final inward move settles into a steep top-down insert of the open box, with his hands bracing opposite edges. The tightly packed report bundles occupy the center while a narrow portion of the archive aisle and shelving remains visible around the box as scale reference.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 낡은 종이 서류 상자 (Open and filled with reports) — The open top faces upward toward the camera, with 강주철's hands holding opposite sides; used as Primary reveal containing the recovered case records; 서류 뭉치들 (Tightly packed inside the box) — The document edges and partially visible cover faces point upward inside the box; no single report face is isolated for reading; used as Creates dense internal lines and shows the volume of archived material; 자료실 책장 (Holding archived case records) — Shelf fronts run along the outer sides of the frame around the box; used as Provides contextual edges around the overhead insert.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime archive-room illumination keeps the subdued palette and documentary clarity without embellishment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old document box remains open in Jucheol's hands and densely packed with case reports.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 강주철 right now — 강주철's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 강주철 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 종이 상자 앞면 라벨: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철의 손에 들린 낡은 종이 상자 안, 빽빽하게 들어찬 서류 뭉치들의 모습\n\nLOCATION (lock): Between the document shelves inside the police records room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Directly above and slightly forward of 강주철's hands, the final inward move settles into a steep top-down insert of the open box, with his hands bracing opposite edges. The tightly packed report bundles occupy the center while a narrow portion of the archive aisle and shelving remains visible around the box as scale reference.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 낡은 종이 서류 상자 (Open and filled with reports) — The open top faces upward toward the camera, with 강주철's hands holding opposite sides; used as Primary reveal containing the recovered case records; 서류 뭉치들 (Tightly packed inside the box) — The document edges and partially visible cover faces point upward inside the box; no single report face is isolated for reading; used as Creates dense internal lines and shows the volume of archived material; 자료실 책장 (Holding archived case records) — Shelf fronts run along the outer sides of the frame around the box; used as Provides contextual edges around the overhead insert.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime archive-room illumination keeps the subdued palette and documentary clarity without embellishment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old document box remains open in Jucheol's hands and densely packed with case reports.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 강주철 right now — 강주철's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 강주철 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 종이 상자 앞면 라벨: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철의 손에 들린 낡은 종이 상자 안, 빽빽하게 들어찬 서류 뭉치들의 모습\n\nLOCATION (lock): Between the document shelves inside the police records room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Directly above and slightly forward of 강주철's hands, the final inward move settles into a steep top-down insert of the open box, with his hands bracing opposite edges. The tightly packed report bundles occupy the center while a narrow portion of the archive aisle and shelving remains visible around the box as scale reference.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 낡은 종이 서류 상자 (Open and filled with reports) — The open top faces upward toward the camera, with 강주철's hands holding opposite sides; used as Primary reveal containing the recovered case records; 서류 뭉치들 (Tightly packed inside the box) — The document edges and partially visible cover faces point upward inside the box; no single report face is isolated for reading; used as Creates dense internal lines and shows the volume of archived material; 자료실 책장 (Holding archived case records) — Shelf fronts run along the outer sides of the frame around the box; used as Provides contextual edges around the overhead insert.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime archive-room illumination keeps the subdued palette and documentary clarity without embellishment.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old document box remains open in Jucheol's hands and densely packed with case reports.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 강주철 right now — 강주철's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 강주철 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 종이 상자 앞면 라벨: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S15sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S15sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:695371>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L18B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:695371>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 '인서트 클로즈업' 샷 크기를 정확히 지켜 상자와 손, 내부 서류에 프레임을 집중하여 프롬프트를 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "상자와 텍스트의 묘사는 좋으나, 인서트 클로즈업의 의도와 달리 피사체의 상체와 하반신 전체가 넓게 포함되어 프레이밍 우선순위에서 크게 감점되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 수직 아래를 향해 강주철의 손에 들린 열린 종이상자와 내부 서류를 똑바로 내려다보고 있습니다.",
      "built_space": "양옆으로 금속 재질의 자료실 책장이 늘어서 있고 가운데 좁은 통로가 보입니다.",
      "entities": "강주철의 손과 소매(주황색/회색 배색의 재킷 일치), '수사기록' 라벨이 붙은 낡은 종이상자, 빽빽하게 채워진 서류 뭉치들이 모두 프롬프트와 일치합니다.",
      "hard_violations": [],
      "physics": "종이 상자는 강주철의 양손에 의해 안정적으로 허공에 들려 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "카메라는 아래를 향해 종이상자와 내부 서류뿐만 아니라 인물의 하반신까지 함께 내려다보고 있습니다.",
      "built_space": "양옆으로 금속 자료실 책장이 있으며, 통로의 바닥이 명확하게 보입니다.",
      "entities": "강주철의 신체(손, 주황색 재킷, 녹색 셔츠와 바지 일치), '수사기록' 라벨이 붙은 상자, 서류 뭉치들이 존재하나 프레임에 인물의 하체가 과도하게 노출되었습니다.",
      "hard_violations": [],
      "physics": "종이 상자는 강주철의 양손에 쥐어져 안정적으로 들려 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 '인서트 클로즈업' 샷 크기를 정확히 지켜 상자와 손, 내부 서류에 프레임을 집중하여 프롬프트를 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "상자와 텍스트의 묘사는 좋으나, 인서트 클로즈업의 의도와 달리 피사체의 상체와 하반신 전체가 넓게 포함되어 프레이밍 우선순위에서 크게 감점되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 수직 아래를 향해 강주철의 손에 들린 열린 종이상자와 내부 서류를 똑바로 내려다보고 있습니다.",
      "built_space": "양옆으로 금속 재질의 자료실 책장이 늘어서 있고 가운데 좁은 통로가 보입니다.",
      "entities": "강주철의 손과 소매(주황색/회색 배색의 재킷 일치), '수사기록' 라벨이 붙은 낡은 종이상자, 빽빽하게 채워진 서류 뭉치들이 모두 프롬프트와 일치합니다.",
      "hard_violations": [],
      "physics": "종이 상자는 강주철의 양손에 의해 안정적으로 허공에 들려 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "카메라는 아래를 향해 종이상자와 내부 서류뿐만 아니라 인물의 하반신까지 함께 내려다보고 있습니다.",
      "built_space": "양옆으로 금속 자료실 책장이 있으며, 통로의 바닥이 명확하게 보입니다.",
      "entities": "강주철의 신체(손, 주황색 재킷, 녹색 셔츠와 바지 일치), '수사기록' 라벨이 붙은 상자, 서류 뭉치들이 존재하나 프레임에 인물의 하체가 과도하게 노출되었습니다.",
      "hard_violations": [],
      "physics": "종이 상자는 강주철의 양손에 쥐어져 안정적으로 들려 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "복장과 서류의 질감 표현은 우수하나, 상자의 윗면과 앞면이 카메라를 향해 동시에 평면적으로 보이는 기하학적 모순이 발생하여 실격입니다.  ★위반: [gemini-pro] 물리적으로 불가능한 기하학적 구조 (상자의 윗면 개방부와 앞면 라벨이 동시에 정면을 향함)"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "상자의 형태가 2D 평면처럼 왜곡되는 물리적 불가능성을 보이며, 서류의 질감과 조명 등 전반적인 디테일이 부족합니다.  ★위반: [gemini-pro] 물리적으로 불가능한 기하학적 구조 (원근법을 무시한 2D 평면 형태의 상자 왜곡)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.5,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[gemini-pro] 물리적으로 불가능한 기하학적 구조 (상자의 윗면 개방부와 앞면 라벨이 동시에 정면을 향함)"
      ],
      "B": [
       "[gemini-pro] 물리적으로 불가능한 기하학적 구조 (원근법을 무시한 2D 평면 형태의 상자 왜곡)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "복장과 서류의 질감 표현은 우수하나, 상자의 윗면과 앞면이 카메라를 향해 동시에 평면적으로 보이는 기하학적 모순이 발생하여 실격입니다.  ★위반: [gemini-pro] 물리적으로 불가능한 기하학적 구조 (상자의 윗면 개방부와 앞면 라벨이 동시에 정면을 향함)"
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "상자의 형태가 2D 평면처럼 왜곡되는 물리적 불가능성을 보이며, 서류의 질감과 조명 등 전반적인 디테일이 부족합니다.  ★위반: [gemini-pro] 물리적으로 불가능한 기하학적 구조 (원근법을 무시한 2D 평면 형태의 상자 왜곡)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1508,
     "B": 1505
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 수직 아래를 향해 강주철의 손에 들린 열린 종이상자와 내부 서류를 똑바로 내려다보고 있습니다.",
    "built_space": "양옆으로 금속 재질의 자료실 책장이 늘어서 있고 가운데 좁은 통로가 보입니다.",
    "entities": "강주철의 손과 소매(주황색/회색 배색의 재킷 일치), '수사기록' 라벨이 붙은 낡은 종이상자, 빽빽하게 채워진 서류 뭉치들이 모두 프롬프트와 일치합니다.",
    "hard_violations": [],
    "physics": "종이 상자는 강주철의 양손에 의해 안정적으로 허공에 들려 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 아래를 향해 종이상자와 내부 서류뿐만 아니라 인물의 하반신까지 함께 내려다보고 있습니다.",
    "built_space": "양옆으로 금속 자료실 책장이 있으며, 통로의 바닥이 명확하게 보입니다.",
    "entities": "강주철의 신체(손, 주황색 재킷, 녹색 셔츠와 바지 일치), '수사기록' 라벨이 붙은 상자, 서류 뭉치들이 존재하나 프레임에 인물의 하체가 과도하게 노출되었습니다.",
    "hard_violations": [],
    "physics": "종이 상자는 강주철의 양손에 쥐어져 안정적으로 들려 있습니다."
   }
  ],
  "totals": {
   "A": 1508,
   "B": 1505
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "지정된 '인서트 클로즈업' 샷 크기를 정확히 지켜 상자와 손, 내부 서류에 프레임을 집중하여 프롬프트를 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "상자와 텍스트의 묘사는 좋으나, 인서트 클로즈업의 의도와 달리 피사체의 상체와 하반신 전체가 넓게 포함되어 프레이밍 우선순위에서 크게 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L18B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "원본 배경 이미지의 시점 및 프롬프트가 지시한 '가파른 탑다운 인서트' 앵글을 무시하고, 상자의 외부 앞면이 보이도록 카메라 각도와 전체 원근감이 크게 변경되었습니다.",
     "fix_en": "Shift the box to a steep top-down perspective, replacing its visible front face with the floor and background shelves, strictly preserving the hands, sleeves, files, and background.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "상자 앞면의 '수사기록' 라벨이 표면의 질감이나 원근에 녹아들지 않고 프레임 위에 단순 합성된 평면적인 그래픽 스티커처럼 보입니다.",
     "fix_en": "Embed the label realistically into the cardboard surface by adding physical texture and matching the ambient lighting, while preserving the box, hands, files, and background.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "원본 배경 이미지의 시점 및 프롬프트가 지시한 '가파른 탑다운 인서트' 앵글을 무시하고, 상자의 외부 앞면이 보이도록 카메라 각도와 전체 원근감이 크게 변경되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "상자 앞면의 '수사기록' 라벨이 표면의 질감이나 원근에 녹아들지 않고 프레임 위에 단순 합성된 평면적인 그래픽 스티커처럼 보입니다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "상자 앞면의 '수사기록' 라벨이 표면의 질감이나 원근에 녹아들지 않고 프레임 위에 단순 합성된 평면적인 그래픽 스티커처럼 보입니다.",
    "fix_en": "Embed the label realistically into the cardboard surface by adding physical texture and matching the ambient lighting, while preserving the box, hands, files, and background.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Shift the box to a steep top-down perspective, replacing its visible front face with the floor and background shelves, strictly preserving the hands, sleeves, files, and background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트에서 요구한 상자 앞면의 '수사기록' 라벨을 정확하게 렌더링했으며, 인물의 의상과 프레이밍 모두 완벽하게 일치합니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "구도와 인물의 의상 등은 잘 반영되었으나, 프롬프트에서 요구된 '수사기록' 라벨 텍스트가 상자에 누락되어 감점되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 인물의 시점에서 아래로 향하여 열린 서류 상자와 그 안의 빽빽한 서류들을 내려다보고 있습니다.",
      "built_space": "자료실의 양옆 책장 사이 복도에 위치하며, 주어진 배경과 원근감이 정확히 일치합니다.",
      "entities": "강주철의 주황색/회색 재킷 소매를 입은 양손, 낡은 종이 서류 상자, 상자 안의 서류 뭉치들, 그리고 상자 앞면의 '수사기록' 라벨이 모두 정확히 구현되었습니다.",
      "hard_violations": [],
      "physics": "양손이 상자의 양옆을 단단히 쥐고 지탱하는 모습이 자연스럽게 연출되었습니다."
     },
     {
      "label": "B",
      "direction": "카메라는 인물의 시점에서 열린 서류 상자를 내려다보고 있습니다.",
      "built_space": "자료실 양옆 책장 사이의 좁은 복도 공간이 배경과 정확하게 일치합니다.",
      "entities": "주황색/회색 재킷을 입은 양손, 종이 서류 상자, 서류 뭉치들은 존재하나, 요구된 상자 앞면의 라벨 텍스트가 없습니다.",
      "hard_violations": [],
      "physics": "양손이 상자를 잡고 물리적으로 올바르게 지탱하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트에서 요구한 상자 앞면의 '수사기록' 라벨을 정확하게 렌더링했으며, 인물의 의상과 프레이밍 모두 완벽하게 일치합니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "구도와 인물의 의상 등은 잘 반영되었으나, 프롬프트에서 요구된 '수사기록' 라벨 텍스트가 상자에 누락되어 감점되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 인물의 시점에서 아래로 향하여 열린 서류 상자와 그 안의 빽빽한 서류들을 내려다보고 있습니다.",
      "built_space": "자료실의 양옆 책장 사이 복도에 위치하며, 주어진 배경과 원근감이 정확히 일치합니다.",
      "entities": "강주철의 주황색/회색 재킷 소매를 입은 양손, 낡은 종이 서류 상자, 상자 안의 서류 뭉치들, 그리고 상자 앞면의 '수사기록' 라벨이 모두 정확히 구현되었습니다.",
      "hard_violations": [],
      "physics": "양손이 상자의 양옆을 단단히 쥐고 지탱하는 모습이 자연스럽게 연출되었습니다."
     },
     {
      "label": "B",
      "direction": "카메라는 인물의 시점에서 열린 서류 상자를 내려다보고 있습니다.",
      "built_space": "자료실 양옆 책장 사이의 좁은 복도 공간이 배경과 정확하게 일치합니다.",
      "entities": "주황색/회색 재킷을 입은 양손, 종이 서류 상자, 서류 뭉치들은 존재하나, 요구된 상자 앞면의 라벨 텍스트가 없습니다.",
      "hard_violations": [],
      "physics": "양손이 상자를 잡고 물리적으로 올바르게 지탱하고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "지시된 '수사기록' 라벨을 상자 앞면에 정확히 렌더링하였으며, 인물의 복장 디테일과 프레이밍 모두 완벽하게 구현되었습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "전반적인 프레이밍과 배경 구성은 훌륭하나, 필수 사항인 상자 앞면의 '수사기록' 라벨이 누락되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라 시선은 열린 상자 안을 수직으로 내려다보고 있으며, 양손은 상자의 양쪽 가장자리를 향해 뻗어 있음.",
      "built_space": "양옆으로 서류가 가득 꽂힌 철제 책장이 늘어선 보관실 통로 중심.",
      "entities": "주황색/회색 소매를 입은 강주철의 팔과 손, 서류 뭉치가 빽빽하게 들어찬 종이 상자. 라벨 텍스트 없음.",
      "hard_violations": [],
      "physics": "양손이 상자 양쪽 테두리를 단단히 쥐고 무게를 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "카메라 시선은 열린 상자 안을 수직으로 내려다보고 있으며, 양손은 상자의 양쪽 가장자리를 향해 뻗어 있음.",
      "built_space": "양옆으로 서류가 가득 꽂힌 철제 책장이 늘어선 보관실 통로 중심.",
      "entities": "주황색 점퍼(소매 벨크로 디테일 포함)를 입은 강주철의 팔과 손, 서류 뭉치가 빽빽하게 들어찬 종이 상자, 상자 앞면의 '수사기록' 라벨.",
      "hard_violations": [],
      "physics": "양손이 상자 양쪽 테두리를 단단히 쥐고 무게를 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시된 '수사기록' 라벨을 상자 앞면에 정확히 렌더링하였으며, 인물의 복장 디테일과 프레이밍 모두 완벽하게 구현되었습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "전반적인 프레이밍과 배경 구성은 훌륭하나, 필수 사항인 상자 앞면의 '수사기록' 라벨이 누락되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "카메라 시선은 열린 상자 안을 수직으로 내려다보고 있으며, 양손은 상자의 양쪽 가장자리를 향해 뻗어 있음.",
      "built_space": "양옆으로 서류가 가득 꽂힌 철제 책장이 늘어선 보관실 통로 중심.",
      "entities": "주황색/회색 소매를 입은 강주철의 팔과 손, 서류 뭉치가 빽빽하게 들어찬 종이 상자. 라벨 텍스트 없음.",
      "hard_violations": [],
      "physics": "양손이 상자 양쪽 테두리를 단단히 쥐고 무게를 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "카메라 시선은 열린 상자 안을 수직으로 내려다보고 있으며, 양손은 상자의 양쪽 가장자리를 향해 뻗어 있음.",
      "built_space": "양옆으로 서류가 가득 꽂힌 철제 책장이 늘어선 보관실 통로 중심.",
      "entities": "주황색 점퍼(소매 벨크로 디테일 포함)를 입은 강주철의 팔과 손, 서류 뭉치가 빽빽하게 들어찬 종이 상자, 상자 앞면의 '수사기록' 라벨.",
      "hard_violations": [],
      "physics": "양손이 상자 양쪽 테두리를 단단히 쥐고 무게를 지탱하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 20,
     "B": 14
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S15sh4__bgfirst_bg.png",
   "bg_asset_id": "1ebbd80a-6266-42b1-a783-cb5ad300ecdf",
   "bg_record_key": "S15sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S15sh4::cine": {
  "applied": true,
  "fingerprint": "0520b301dbe0eea1b519618902f905f4be5a2dbccc836a6422769f47a653c612",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S15sh4_sel.png",
  "source_sha256": "7b16a74c0f7a378a65778a593616019f06397aeab51751575f1382eff0f9bcf8",
  "file": "S15sh4_cine.png",
  "latency_ms": 10286
 },
 "S15sh5::signage": {
  "fp": "44d10f1c8decb1d2",
  "inscriptions": [
   {
    "surface_native": "누런 서류 봉투",
    "text_native": "수사기록",
    "reason_ko": "경찰서 자료실에서 인물이 들여다보는 서류 봉투가 공식 수사 문서임을 나타내기 위해 봉투 표면에 '수사기록'이라는 표기가 필요합니다."
   }
  ]
 },
 "S15sh5": {
  "input_fingerprint": "28a88bbb6e26875a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 누런 서류 봉투 한 장을 손에 들고 미간을 잔뜩 찌푸린 채 내려다보는 서의용의 얼굴\n\nLOCATION (lock): In an aisle between shelves in the police records room, beside the opened archive box. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close beside 서의용 and slightly below his eye height, the tracking rise ends on an offset three-quarter view of his downturned face. His deeply furrowed brow holds the upper center while the yellowed document envelope remains naturally sized along the lower edge in his raised hand, preserving the source of his troubled concentration.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 누런 서류 봉투 (Lifted from the recovered records and held open for inspection) — The broad front face angles upward toward 서의용 and partly toward camera, without introducing unreadable invented markings; used as Retained at the lower frame edge as the immediate object of 서의용's scrutiny; 자료실 책장 (Filled with stored records) — The shelf fronts recede obliquely behind 서의용; used as Forms a compressed archive backdrop without distracting from his expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime ambient illumination preserves restrained neutral-to-cool color and subdued contrast across his troubled face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong has removed a yellowed document envelope from the records box and continues holding and examining it.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 누런 서류 봉투: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 누런 서류 봉투 한 장을 손에 들고 미간을 잔뜩 찌푸린 채 내려다보는 서의용의 얼굴\n\nLOCATION (lock): In an aisle between shelves in the police records room, beside the opened archive box. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close beside 서의용 and slightly below his eye height, the tracking rise ends on an offset three-quarter view of his downturned face. His deeply furrowed brow holds the upper center while the yellowed document envelope remains naturally sized along the lower edge in his raised hand, preserving the source of his troubled concentration.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 누런 서류 봉투 (Lifted from the recovered records and held open for inspection) — The broad front face angles upward toward 서의용 and partly toward camera, without introducing unreadable invented markings; used as Retained at the lower frame edge as the immediate object of 서의용's scrutiny; 자료실 책장 (Filled with stored records) — The shelf fronts recede obliquely behind 서의용; used as Forms a compressed archive backdrop without distracting from his expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime ambient illumination preserves restrained neutral-to-cool color and subdued contrast across his troubled face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong has removed a yellowed document envelope from the records box and continues holding and examining it.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 누런 서류 봉투: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 누런 서류 봉투 한 장을 손에 들고 미간을 잔뜩 찌푸린 채 내려다보는 서의용의 얼굴\n\nLOCATION (lock): In an aisle between shelves in the police records room, beside the opened archive box. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close beside 서의용 and slightly below his eye height, the tracking rise ends on an offset three-quarter view of his downturned face. His deeply furrowed brow holds the upper center while the yellowed document envelope remains naturally sized along the lower edge in his raised hand, preserving the source of his troubled concentration.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 누런 서류 봉투 (Lifted from the recovered records and held open for inspection) — The broad front face angles upward toward 서의용 and partly toward camera, without introducing unreadable invented markings; used as Retained at the lower frame edge as the immediate object of 서의용's scrutiny; 자료실 책장 (Filled with stored records) — The shelf fronts recede obliquely behind 서의용; used as Forms a compressed archive backdrop without distracting from his expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime ambient illumination preserves restrained neutral-to-cool color and subdued contrast across his troubled face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong has removed a yellowed document envelope from the records box and continues holding and examining it.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 누런 서류 봉투: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "프롬프트에 명시된 열린 보관 상자를 정확히 배치하였으며, 인물의 의상과 시선 처리, 지정된 문구 삽입 등 모든 요구사항을 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "인물의 표정과 시선, 봉투의 문구는 잘 표현되었으나, 장소 지정에서 필수 요소인 '열린 보관 상자'가 화면에 등장하지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S15sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "서류 봉투 중앙에 적힌 '수사기록' 텍스트가 종이의 질감, 굴곡, 조명을 전혀 반영하지 않은 채 화면 위에 디지털 그래픽으로 평면적으로 덧씌워진 것처럼 보입니다.",
     "fix_en": "Render the '수사기록' text as physical ink absorbed into the paper, matching the envelope's lighting, texture, and curvature. Preserve Euiyong's face, expression, hair, clothing, hand, the envelope's size and placement, and the bookshelf background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 하단에서 봉투를 잡고 있는 오른손의 해부학적 구조가 어색하며, 엄지손가락 바로 아래에 비정상적인 손가락 마디(또는 살덩이)가 봉투 모서리와 융합되어 있습니다.",
     "fix_en": "Redraw the hand holding the envelope to have correct human anatomy, ensuring natural fingers and no fused flesh. Preserve Euiyong's face, clothing, the envelope, the printed text, and the background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "프레임 중앙 서의용 얼굴이 캐릭터 레퍼런스와 다른 사람이다",
     "fix_en": "Adjust Euiyong's facial features to match the specific identity in the character reference. Preserve his frown, gaze direction, lighting, clothing, hand, envelope, and the background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "하단 누런 서류 봉투가 열어 검사하는 상태가 아니라 닫혀 있다",
     "fix_en": "Redraw the top edge of the yellow envelope so it appears held open for inspection. Preserve the printed text, Euiyong's hand, face, expression, clothing, and the background.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "머리 위 레퍼런스 남색 비니가 없고 짧은 머리가 그대로 드러나 있다",
     "fix_en": "Place the navy beanie from the character reference on Euiyong's head. Preserve his face, expression, lighting, jacket, hoodie, hand, envelope, and the background.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "서류 봉투 중앙에 적힌 '수사기록' 텍스트가 종이의 질감, 굴곡, 조명을 전혀 반영하지 않은 채 화면 위에 디지털 그래픽으로 평면적으로 덧씌워진 것처럼 보입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 하단에서 봉투를 잡고 있는 오른손의 해부학적 구조가 어색하며, 엄지손가락 바로 아래에 비정상적인 손가락 마디(또는 살덩이)가 봉투 모서리와 융합되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "프레임 중앙 서의용 얼굴이 캐릭터 레퍼런스와 다른 사람이다",
     "severity": "major"
    },
    {
     "issue_ko": "하단 누런 서류 봉투가 열어 검사하는 상태가 아니라 닫혀 있다",
     "severity": "major"
    },
    {
     "issue_ko": "머리 위 레퍼런스 남색 비니가 없고 짧은 머리가 그대로 드러나 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 하단에서 봉투를 잡고 있는 오른손의 해부학적 구조가 어색하며, 엄지손가락 바로 아래에 비정상적인 손가락 마디(또는 살덩이)가 봉투 모서리와 융합되어 있습니다.",
    "fix_en": "Redraw the hand holding the envelope to have correct human anatomy, ensuring natural fingers and no fused flesh. Preserve Euiyong's face, clothing, the envelope, the printed text, and the background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "프레임 중앙 서의용 얼굴이 캐릭터 레퍼런스와 다른 사람이다",
    "fix_en": "Adjust Euiyong's facial features to match the specific identity in the character reference. Preserve his frown, gaze direction, lighting, clothing, hand, envelope, and the background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "하단 누런 서류 봉투가 열어 검사하는 상태가 아니라 닫혀 있다",
    "fix_en": "Redraw the top edge of the yellow envelope so it appears held open for inspection. Preserve the printed text, Euiyong's hand, face, expression, clothing, and the background.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "머리 위 레퍼런스 남색 비니가 없고 짧은 머리가 그대로 드러나 있다",
    "fix_en": "Place the navy beanie from the character reference on Euiyong's head. Preserve his face, expression, lighting, jacket, hoodie, hand, envelope, and the background.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Render the '수사기록' text as physical ink absorbed into the paper, matching the envelope's lighting, texture, and curvature. Preserve Euiyong's face, expression, hair, clothing, hand, the envelope's size and placement, and the bookshelf background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1875,
      "verdict_ko": "봉투의 기울기와 원근감에 맞게 '수사기록' 텍스트가 자연스럽게 밀착되어 있어 지시사항의 질감 사실주의를 훌륭히 충족합니다."
     },
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "봉투는 기울어져 있으나 텍스트가 수평으로 합성된 것처럼 떠 있어 표면의 방향을 따르지 않는 점이 아쉽습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.875
     },
     "adjusted": {
      "A": 1.571,
      "B": 1.875
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1875,
      "verdict_ko": "봉투의 기울기와 원근감에 맞게 '수사기록' 텍스트가 자연스럽게 밀착되어 있어 지시사항의 질감 사실주의를 훌륭히 충족합니다."
     },
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "봉투는 기울어져 있으나 텍스트가 수평으로 합성된 것처럼 떠 있어 표면의 방향을 따르지 않는 점이 아쉽습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1875,
      "verdict_ko": "프롬프트의 지시사항인 인물의 외형, 시선, 조명 및 '수사기록' 텍스트를 매우 정확하게 구현함 (두 후보는 시각적으로 거의 동일함)."
     },
     {
      "label": "B",
      "score": 2000,
      "verdict_ko": "프롬프트의 지시사항인 인물의 외형, 시선, 조명 및 '수사기록' 텍스트를 매우 정확하게 구현함 (두 후보는 시각적으로 거의 동일함)."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.875,
      "B": 2.0
     },
     "adjusted": {
      "A": 1.875,
      "B": 2.0
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1875,
      "verdict_ko": "프롬프트의 지시사항인 인물의 외형, 시선, 조명 및 '수사기록' 텍스트를 매우 정확하게 구현함 (두 후보는 시각적으로 거의 동일함)."
     },
     {
      "label": "A",
      "score": 2000,
      "verdict_ko": "프롬프트의 지시사항인 인물의 외형, 시선, 조명 및 '수사기록' 텍스트를 매우 정확하게 구현함 (두 후보는 시각적으로 거의 동일함)."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 3571,
     "B": 3750
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S15sh4"
  }
 },
 "S15sh5::cine": {
  "applied": true,
  "fingerprint": "bb3fb89e06861bf77cddc14d7a1ea37e8b5c09bc7139a6d7db0099b6d03dcdf5",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S15sh5_sel.png",
  "source_sha256": "8d0e6fa0f33bf329d8419b2a19f53148daf95fa3fcb6a0b5af343a03384bda1d",
  "file": "S15sh5_cine.png",
  "latency_ms": 10826
 },
 "S16sh3::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 샷은 소형차 내부의 운전석을 배경으로 하며, 인물이 운전석에 앉아 창밖으로 손을 내미는 상황이므로 차량 내부의 좌석 배치와 방향성이 정확하게 표현되어야 합니다.",
  "input_fingerprint": "afc9a37872690887"
 },
 "S16sh3::signage": {
  "fp": "32075a7dc95d552a",
  "inscriptions": []
 },
 "confinedfp::c958aa790ec7": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/confinedfp_base_c958aa790ec7.png",
  "place_text": "Inside the compact car at the driver's seat, beside the lowered window on the deserted alley.",
  "input_fingerprint": "8612ead4c3377d87"
 },
 "S16sh3::confined_fp": {
  "reads": {
   "controls": "Window control is located on the top-left of the diagram, positioned on the driver's side interior door.",
   "mirrors": "A side mirror is located on the middle-left exterior of the car, pointing left (outward) from the driver's side.",
   "camera": "Positioned at the bottom-right outside the vehicle (rear passenger side), pointing diagonally toward the top-left (driver's seat).",
   "occupants": "지국현 (과거) occupies the middle-left driver's seat; 김선영 stands outside the vehicle on the right, adjacent to the front passenger seat."
  },
  "mismatches": [
   "The text dictates that 김선영 appears on the middle-left of the frame and 지국현 on the middle-right, which would require the camera to be forward of 김선영 pointing backward. The diagram contradicts this by placing the camera behind her to the bottom-right, pointing diagonally forward (top-left), which visually reverses their positions, placing 김선영 on the right and 지국현 on the left of the frame."
  ],
  "scene_description_en": "The camera aims diagonally forward and left across the cabin from outside the rear passenger side. In the near right foreground, 김선영 stands beside the open front passenger window. Through the opening, the empty front passenger seat occupies the center-right midground, bordered by the center console directly in the middle and the empty rear seat in the near left foreground. On the midground left, 지국현 occupies the driver's seat. In the far left background, a window control panel sits on the driver's interior door. Anchoring the far left exterior edge, the driver's side mirror faces outward to the left, reflecting only the off-screen street and capturing none of the interior.",
  "fixed": true,
  "input_fingerprint": "8898a12891804e8c"
 },
 "era_assess::630e751cfa07c193": {
  "subjects": [
   {
    "subject_native": "2001~2017년 한국의 경차 및 소형차 내부",
    "search_terms_native": [
     "마티즈 내부",
     "모닝 운전석",
     "2000년대 경차 실내",
     "구형 소형차 대시보드"
    ],
    "language_lock_native": "검색 시 다른 언어를 혼용하지 말고 오직 한국어로만 검색을 수행하십시오.",
    "reason_ko": "2000년대에서 2010년대 초반 한국의 대중적인 경차(모닝, 마티즈 등) 실내 대시보드 구조, 스티어링 휠 디자인, 저가형 플라스틱 질감은 일반적인 현대식 차량이나 서구권 차량과 확연히 달라 고증이 필수적입니다."
   }
  ]
 },
 "era_ref::138cf336788818a3": {
  "subject": "2001~2017년 한국의 경차 및 소형차 내부",
  "terms": [
   "마티즈 내부",
   "모닝 운전석",
   "2000년대 경차 실내",
   "구형 소형차 대시보드"
  ],
  "queries": [
   [
    "2000년대 한국 경차 실내 마티즈 모닝 운전석 대시보드",
    "구형 소형차 대시보드 마티즈 내부 모닝 운전석"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://static.wixstatic.com/media/01189d_19e9d2db4f414fcb903581c3510a5a7f~mv2.jpg/v1/fill/w_960%2Ch_764%2Cal_c%2Cq_85%2Cenc_avif%2Cquality_auto/01189d_19e9d2db4f414fcb903581c3510a5a7f~mv2.jpg",
  "picked_reason_ko": "2000년대 중후반 한국 대우 마티즈의 실내를 여러 시점에서 선명하게 보여 주어, 경차 특유의 좁은 비례와 대시보드·좌석·도어·계기판의 재료 및 구성을 가장 충실히 파악할 수 있다.",
  "sha256": "bbaee13987f957145aa382040cb362e349befe50a0c6ac3bd6392354c313db90",
  "file": "eraref_138cf336788818a3.png"
 },
 "S16sh3": {
  "input_fingerprint": "5c4f89d7072b7b4c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): late night.\n\nSHOT TEXT (authoritative, Korean): 열린 차창 밖으로 손을 들어 올린 채 미소 짓는 지국현의 얼굴\n\nLOCATION (lock): Inside the compact car at the driver's seat, beside the lowered window on the deserted alley. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 김선영's rear three-quarter shoulder height just outside the passenger side, the camera advances obliquely through the open window toward 지국현 across the cabin. 지국현's smiling face sits in the middle-right with his raised hand gesturing toward 김선영, while her purple padded shoulder and hesitant posture remain at the near left edge as the physical recipient of the invitation.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 김선영 in the middle-left of the frame, foreground, looks toward 지국현 inside the car; 지국현 (과거) in the middle-right of the frame, midground, reaches for 김선영 outside the passenger window.\n- KEY BACKGROUND ELEMENTS: 열린 조수석 차창 (Open) — The passenger-side opening faces the camera, allowing an unobstructed oblique view across the cabin; used as The open window provides the through-opening frame around 지국현; 빨간 마티즈 (Moving slowly through the empty alley) — Its passenger side faces the camera, with the driver positioned across the cabin; used as Establishes the vehicle and gives a physical border to the window exchange; 김선영의 보라색 패딩 (Worn by 김선영) — The rear shoulder and side of the jacket face camera at the left edge; used as Places 김선영 as the hesitant foreground recipient without making her frontal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Late-night illumination from the stated streetlight is kept low-key and naturalistic, with subdued color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong is still wearing the purple padded jacket established when she left home as she approaches Ji Gukhyeon's car.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (과거) (Korean 남성, 20대 중반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera aims diagonally forward and left across the cabin from outside the rear passenger side. In the near right foreground, 김선영 stands beside the open front passenger window. Through the opening, the empty front passenger seat occupies the center-right midground, bordered by the center console directly in the middle and the empty rear seat in the near left foreground. On the midground left, 지국현 occupies the driver's seat. In the far left background, a window control panel sits on the driver's interior door. Anchoring the far left exterior edge, the driver's side mirror faces outward to the left, reflecting only the off-screen street and capturing none of the interior.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): late night.\n\nSHOT TEXT (authoritative, Korean): 열린 차창 밖으로 손을 들어 올린 채 미소 짓는 지국현의 얼굴\n\nLOCATION (lock): Inside the compact car at the driver's seat, beside the lowered window on the deserted alley. The shot takes place here.\n",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera aims diagonally forward and left across the cabin from outside the rear passenger side. In the near right foreground, 김선영 stands beside the open front passenger window. Through the opening, the empty front passenger seat occupies the center-right midground, bordered by the center console directly in the middle and the empty rear seat in the near left foreground. On the midground left, 지국현 occupies the driver's seat. In the far left background, a window control panel sits on the driver's interior door. Anchoring the far left exterior edge, the driver's side mirror faces outward to the left, reflecting only the off-screen street and capturing none of the interior.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): late night.\n\nSHOT TEXT (authoritative, Korean): 열린 차창 밖으로 손을 들어 올린 채 미소 짓는 지국현의 얼굴\n\nLOCATION (lock): Inside the compact car at the driver's seat, beside the lowered window on the deserted alley. The shot takes place here.\n\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S16sh3_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2001~2017년 한국의 경차 및 소형차 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_138cf336788818a3.png"
    },
    {
     "label": "지국현 (과거)",
     "path": "<bytes:1122770>"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S16sh3_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2001~2017년 한국의 경차 및 소형차 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_138cf336788818a3.png"
    },
    {
     "label": "지국현 (과거)",
     "path": "<bytes:1122770>"
    }
   ]
  },
  "initial_roll_all_fail": true,
  "gq": {
   "route": "combined",
   "gap": 0.5,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.0,
    "B": 1.167
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트에 명시되지 않은 발명된 인물(파란 후드티 남성) 추가",
     "[openrouter:x-ai/grok-4.6] 샷에 없는 남성 인물을 조수석 밖에 추가함(invented extra body)"
    ],
    "B": [
     "[gemini-pro] 물리적으로 불가능한 구조(창틀에 붙은 스티어링 휠 및 엉망인 대시보드)",
     "[gemini-pro] 스테이징과 다른 위치에 인물 배치(지국현이 차량 외부 조수석 방향에 위치)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1000,
   "B": 1167
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1000,
    "verdict_ko": "요구된 김선영(보라색 패딩) 대신 전혀 다른 남성을 배치하고, 창문 너머의 교감 대신 조수석 문을 활짝 열어버려 핵심 연출을 완전히 실패했습니다.  ★위반: [gemini-pro] 프롬프트에 명시되지 않은 발명된 인물(파란 후드티 남성) 추가 / [openrouter:x-ai/grok-4.6] 샷에 없는 남성 인물을 조수석 밖에 추가함(invented extra body)"
   },
   {
    "label": "B",
    "score": 1167,
    "verdict_ko": "지국현이 운전석이 아닌 외부에 서 있으며, 스티어링 휠이 창틀에 붙어있는 등 물리적으로 불가능한 차량 내부 구조를 렌더링하여 사용할 수 없습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 구조(창틀에 붙은 스티어링 휠 및 엉망인 대시보드) / [gemini-pro] 스테이징과 다른 위치에 인물 배치(지국현이 차량 외부 조수석 방향에 위치)"
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S16sh3_confinedfp.png"
   },
   {
    "label": "PERIOD REFERENCE — 2001~2017년 한국의 경차 및 소형차 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_138cf336788818a3.png"
   },
   {
    "label": "지국현 (과거)",
    "path": "<bytes:1122770>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 중앙, 운전석에 앉아 있어야 할 지국현이 차량 외부에 서서 안을 들여다보고 있습니다.",
     "fix_en": "Erase the man standing outside the car, replacing him with the dark street background. Maintain the car interior, seats, steering wheel, and lighting exactly as they are.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 좌측, 스티어링 휠이 대시보드와 연결되지 않고 창문 밖 허공과 문틀에 기형적으로 붙어 있습니다.",
     "fix_en": "Remove the steering wheel attached to the left window frame, revealing the dark street view behind it. Maintain the man outside, the car interior, seats, and lighting exactly as they are.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 우측, 운전석 스티어링 휠 뒤에 있어야 할 계기판이 차량 중앙 센터 콘솔 위쪽에 잘못 생성되어 있습니다.",
     "fix_en": "Replace the instrument panel on the center console with standard car radio and air conditioning controls. Maintain the man outside, the seats, the steering wheel, and lighting exactly as they are.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 좌측 끝, 거리만 비춰야 하는 사이드미러에 차량의 창틀과 내부 구조가 반사되어 있습니다.",
     "fix_en": "Change the reflection in the side mirror to show only a dark asphalt street, removing the window frame reflection. Maintain the man outside, the car interior, mirror housing, and lighting exactly as they are.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 중앙, 운전석에 앉아 있어야 할 지국현이 차량 외부에 서서 안을 들여다보고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 좌측, 스티어링 휠이 대시보드와 연결되지 않고 창문 밖 허공과 문틀에 기형적으로 붙어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 우측, 운전석 스티어링 휠 뒤에 있어야 할 계기판이 차량 중앙 센터 콘솔 위쪽에 잘못 생성되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 좌측 끝, 거리만 비춰야 하는 사이드미러에 차량의 창틀과 내부 구조가 반사되어 있습니다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 좌측 끝, 거리만 비춰야 하는 사이드미러에 차량의 창틀과 내부 구조가 반사되어 있습니다.",
    "fix_en": "Change the reflection in the side mirror to show only a dark asphalt street, removing the window frame reflection. Maintain the man outside, the car interior, mirror housing, and lighting exactly as they are.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Erase the man standing outside the car, replacing him with the dark street background. Maintain the car interior, seats, steering wheel, and lighting exactly as they are.\n- Remove the steering wheel attached to the left window frame, revealing the dark street view behind it. Maintain the man outside, the car interior, seats, and lighting exactly as they are.\n- Replace the instrument panel on the center console with standard car radio and air conditioning controls. Maintain the man outside, the seats, the steering wheel, and lighting exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1333,
      "verdict_ko": "차량 내부 디테일을 레퍼런스와 완벽에 가깝게 구현하고 지국현을 운전석에 올바르게 배치했으나, 샷 텍스트의 핵심인 '창밖으로 손을 올리고 미소 짓는' 동작을 전혀 수행하지 않았으며 김선영이 누락되었습니다."
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "미소와 손을 올린 동작은 묘사했으나 지국현이 운전석이 아닌 차량 외부에 서 있어 레이아웃을 정면으로 위반한 심각한 오류가 있으며, 역시 김선영이 누락되었습니다.  ★위반: [gemini-pro] 지국현이 운전석에 탑승하지 않고 차량 외부에 위치함 (레이아웃 배치 위반)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.25,
      "B": 1.333
     },
     "violations": {
      "A": [
       "[gemini-pro] 지국현이 운전석에 탑승하지 않고 차량 외부에 위치함 (레이아웃 배치 위반)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1333,
      "verdict_ko": "차량 내부 디테일을 레퍼런스와 완벽에 가깝게 구현하고 지국현을 운전석에 올바르게 배치했으나, 샷 텍스트의 핵심인 '창밖으로 손을 올리고 미소 짓는' 동작을 전혀 수행하지 않았으며 김선영이 누락되었습니다."
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "미소와 손을 올린 동작은 묘사했으나 지국현이 운전석이 아닌 차량 외부에 서 있어 레이아웃을 정면으로 위반한 심각한 오류가 있으며, 역시 김선영이 누락되었습니다.  ★위반: [gemini-pro] 지국현이 운전석에 탑승하지 않고 차량 외부에 위치함 (레이아웃 배치 위반)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "차량 내부 디테일은 훌륭하나, 샷 텍스트의 행동(미소, 창밖으로 손)을 누락했고 운전자의 머리 방향이 물리적으로 불가능하게 비틀려 있는 치명적인 구조적 오류가 있습니다.  ★위반: [gemini-pro] Physically impossible anatomy: 운전자의 양팔은 앞을 향해 뻗어 스티어링 휠을 잡고 있으나, 우측 후방에 있는 카메라를 향해 고개를 돌렸음에도 불구하고 왼쪽 얼굴이 보이는 것은 물리적으로 불가능한 목과 머리의 방향임."
     },
     {
      "label": "B",
      "score": 500,
      "verdict_ko": "지국현이 운전석이 아닌 차량 외부로 밀려났으며, 스티어링 휠과 대시보드가 분리되는 등 차량 내부 기하학이 완전히 붕괴된 치명적 오류가 있습니다.  ★위반: [gemini-pro] Physically impossible built space: 스티어링 휠이 대시보드와 연결되지 않은 채 차량 측면이나 허공에 분리되어 위치함. / [gemini-pro] A person seated in a place the staging does not put them: 지국현이 운전석을 차지해야 한다는 레이아웃 지시를 무시하고 차량 외부 지면에 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.429,
      "B": 1.0
     },
     "adjusted": {
      "A": 1.179,
      "B": 0.5
     },
     "violations": {
      "A": [
       "[gemini-pro] Physically impossible anatomy: 운전자의 양팔은 앞을 향해 뻗어 스티어링 휠을 잡고 있으나, 우측 후방에 있는 카메라를 향해 고개를 돌렸음에도 불구하고 왼쪽 얼굴이 보이는 것은 물리적으로 불가능한 목과 머리의 방향임."
      ],
      "B": [
       "[gemini-pro] Physically impossible built space: 스티어링 휠이 대시보드와 연결되지 않은 채 차량 측면이나 허공에 분리되어 위치함.",
       "[gemini-pro] A person seated in a place the staging does not put them: 지국현이 운전석을 차지해야 한다는 레이아웃 지시를 무시하고 차량 외부 지면에 서 있음."
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.571,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "차량 내부 디테일은 훌륭하나, 샷 텍스트의 행동(미소, 창밖으로 손)을 누락했고 운전자의 머리 방향이 물리적으로 불가능하게 비틀려 있는 치명적인 구조적 오류가 있습니다.  ★위반: [gemini-pro] Physically impossible anatomy: 운전자의 양팔은 앞을 향해 뻗어 스티어링 휠을 잡고 있으나, 우측 후방에 있는 카메라를 향해 고개를 돌렸음에도 불구하고 왼쪽 얼굴이 보이는 것은 물리적으로 불가능한 목과 머리의 방향임."
     },
     {
      "label": "A",
      "score": 500,
      "verdict_ko": "지국현이 운전석이 아닌 차량 외부로 밀려났으며, 스티어링 휠과 대시보드가 분리되는 등 차량 내부 기하학이 완전히 붕괴된 치명적 오류가 있습니다.  ★위반: [gemini-pro] Physically impossible built space: 스티어링 휠이 대시보드와 연결되지 않은 채 차량 측면이나 허공에 분리되어 위치함. / [gemini-pro] A person seated in a place the staging does not put them: 지국현이 운전석을 차지해야 한다는 레이아웃 지시를 무시하고 차량 외부 지면에 서 있음."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1750,
     "B": 2512
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "confined_fp": {
   "base_key": "confinedfp::c958aa790ec7",
   "apt_reason": "이 샷은 소형차 내부의 운전석을 배경으로 하며, 인물이 운전석에 앉아 창밖으로 손을 내미는 상황이므로 차량 내부의 좌석 배치와 방향성이 정확하게 표현되어야 합니다.",
   "fixed": true,
   "mismatches": [
    "The text dictates that 김선영 appears on the middle-left of the frame and 지국현 on the middle-right, which would require the camera to be forward of 김선영 pointing backward. The diagram contradicts this by placing the camera behind her to the bottom-right, pointing diagonally forward (top-left), which visually reverses their positions, placing 김선영 on the right and 지국현 on the left of the frame."
   ],
   "era_research": {
    "subject": "2001~2017년 한국의 경차 및 소형차 내부",
    "queries": [
     [
      "2000년대 한국 경차 실내 마티즈 모닝 운전석 대시보드",
      "구형 소형차 대시보드 마티즈 내부 모닝 운전석"
     ]
    ],
    "picked_url": "https://static.wixstatic.com/media/01189d_19e9d2db4f414fcb903581c3510a5a7f~mv2.jpg/v1/fill/w_960%2Ch_764%2Cal_c%2Cq_85%2Cenc_avif%2Cquality_auto/01189d_19e9d2db4f414fcb903581c3510a5a7f~mv2.jpg",
    "sha256": "bbaee13987f957145aa382040cb362e349befe50a0c6ac3bd6392354c313db90",
    "file": "eraref_138cf336788818a3.png"
   }
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S16sh3::cine": {
  "applied": true,
  "fingerprint": "59fcbee630f5d5b31657a2048bb8f488902afec09dcae20405ea8bd0324d18c1",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S16sh3_sel.png",
  "source_sha256": "624a58bf9b2e8a284f1415bbd4a3641d287b05e8ad87e9390f5ab98a72fc9625",
  "file": "S16sh3_cine.png",
  "latency_ms": 11245
 },
 "S16sh9::confined_fp_apt": {
  "applies": true,
  "reason_ko": "차량 내부 조수석이라는 방향성과 위치 지정이 뚜렷한 공간에서 촬영되므로, 좌석 및 대시보드, 창문 등 차량 내부의 정확한 공간 배치를 위해 평면도 보조가 필요합니다.",
  "input_fingerprint": "093792795395cf28"
 },
 "S16sh9::signage": {
  "fp": "09c6fe77204983ba",
  "inscriptions": []
 },
 "confinedfp::7f322c8f2baf": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/confinedfp_base_7f322c8f2baf.png",
  "place_text": "Inside the stopped car at the front passenger seat on the riverside road, enclosed by the adjacent seat, dashboard, window, and controls.",
  "input_fingerprint": "48320b8b9ce67df4"
 },
 "S16sh9::confined_fp": {
  "reads": {
   "controls": "A Window Control is located on the right side of the Front Passenger Seat.",
   "mirrors": "No mirrors are depicted in the diagram.",
   "camera": "The camera is positioned directly behind the Front Passenger Seat, pointing straight forward toward the dashboard.",
   "occupants": "Kim Seon-yeong is in the Front Passenger Seat, facing forward toward the dashboard."
  },
  "mismatches": [
   "The text places the camera near the center console to capture a tight three-quarter facial frame of Kim Seon-yeong. The diagram places the camera directly behind the passenger seat, pointing straight forward, which would capture the back of her head and the back of the seat, not her face."
  ],
  "scene_description_en": "The camera is positioned in the back row, looking directly forward at the back of the front passenger seat. In the center foreground, Kim Seon-yeong sits in this seat facing away from the camera, meaning only the back of her head and shoulders are visible. To the right of the seat, a window control unit appears on screen. Further right, the riverside window flanks the edge of the frame. In the background directly ahead, the passenger side of the dashboard spans across the view. No mirrors are present in the scene to reflect the occupant or surroundings.",
  "fixed": true,
  "input_fingerprint": "3e1906ec5e07ffad"
 },
 "era_assess::bfdfc03bdd3a9076": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 초반 한국 국산차 내부 대시보드 및 조수석",
    "search_terms_native": [
     "2000년대 국산차 내부",
     "NF소나타 실내 조수석",
     "아반떼 HD 대시보드",
     "SM5 실내 인테리어"
    ],
    "language_lock_native": "이 검색은 오직 한국어로만 수행되어야 하며 다른 언어의 단어를 혼합하거나 번역해서는 안 됩니다.",
    "reason_ko": "2000년대부터 2010년대 초반 사이 한국에서 흔히 볼 수 있었던 현대, 기아, 르노삼성 자동차 특유의 대시보드 레이아웃, 조수석 형태, 센터페시아 재질은 글로벌 표준이나 최신 차량 인테리어와 확연히 다릅니다."
   }
  ]
 },
 "era_ref::e89b15ab00eb5e8e": {
  "subject": "2000년대~2010년대 초반 한국 국산차 내부 대시보드 및 조수석",
  "terms": [
   "2000년대 국산차 내부",
   "NF소나타 실내 조수석",
   "아반떼 HD 대시보드",
   "SM5 실내 인테리어"
  ],
  "queries": [
   [
    "2000년대 국산차 실내 대시보드 조수석 엔에프 쏘나타 아반떼 에이치디 에스엠파이브",
    "2010년대 초반 국산 중형차 준중형차 실내 인테리어 대시보드 조수석"
   ],
   [
    "엔에프 쏘나타 실내 조수석 대시보드 사진",
    "아반떼 에이치디 실내 대시보드 사진",
    "뉴 에스엠파이브 실내 조수석 대시보드 사진"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://cnp1753-prod.s3.amazonaws.com/cars/6Xnd6kDaaG6jVu/nWPjoEQ6GHY58bLRlhT0.jpg",
  "picked_reason_ko": "3번은 해당 연대의 한국 국산차 실내를 넓고 선명하게 담아 대시보드 전체 구성과 센터페시아, 조수석 및 좌석 재질까지 가장 잘 읽힌다.",
  "sha256": "3fc806b7ba03145c0e9c77f9b477f41a7985a86d48e6d1b3ccb5eb7fb5c038c8",
  "file": "eraref_e89b15ab00eb5e8e.png"
 },
 "S16sh9": {
  "input_fingerprint": "19bbf131fce27d21",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): late night.\n\nSHOT TEXT (authoritative, Korean): 멈춰 선 차 안, 동공이 크게 확장된 채 숨을 죽이고 굳어버린 김선영의 창백한 얼굴\n\nLOCATION (lock): Inside the stopped car at the front passenger seat on the riverside road, enclosed by the adjacent seat, dashboard, window, and controls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Near the center console and slightly above 김선영's eyeline, the inward move ends in a tight three-quarter facial frame, isolating her against only narrow fragments of the passenger seat and side window. Her shoulders lock into the seat, her pupils are widened, and her arrested breath reads in the stillness of her mouth as she stares away from the camera toward what she has realized.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 김선영 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 조수석과 차내 (The car has stopped) — The passenger seat surrounds her shoulders while the center-console side remains closest to camera; used as A narrow contextual border that confirms 김선영 remains inside the stopped car; 조수석 옆 차창 (Beside the stopped car at the riverside road) — The window borders the far side of her face, with the late-night riverside beyond kept indistinct; used as Provides a restrained outer edge without competing with her frozen expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained low-level late-night illumination leaves her explicitly pale face legible while maintaining somber neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong remains in the purple padded jacket inside the stopped car; no clothing change has yet occurred.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김선영 (Korean 소녀, 17세, 부드러운 타원형 얼굴, 검은색 긴 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned in the back row, looking directly forward at the back of the front passenger seat. In the center foreground, Kim Seon-yeong sits in this seat facing away from the camera, meaning only the back of her head and shoulders are visible. To the right of the seat, a window control unit appears on screen. Further right, the riverside window flanks the edge of the frame. In the background directly ahead, the passenger side of the dashboard spans across the view. No mirrors are present in the scene to reflect the occupant or surroundings.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): late night.\n\nSHOT TEXT (authoritative, Korean): 멈춰 선 차 안, 동공이 크게 확장된 채 숨을 죽이고 굳어버린 김선영의 창백한 얼굴\n\nLOCATION (lock): Inside the stopped car at the front passenger seat on the riverside road, enclosed by the adjacent seat, dashboard, window, and controls. The shot takes place here.\n",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned in the back row, looking directly forward at the back of the front passenger seat. In the center foreground, Kim Seon-yeong sits in this seat facing away from the camera, meaning only the back of her head and shoulders are visible. To the right of the seat, a window control unit appears on screen. Further right, the riverside window flanks the edge of the frame. In the background directly ahead, the passenger side of the dashboard spans across the view. No mirrors are present in the scene to reflect the occupant or surroundings.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): late night.\n\nSHOT TEXT (authoritative, Korean): 멈춰 선 차 안, 동공이 크게 확장된 채 숨을 죽이고 굳어버린 김선영의 창백한 얼굴\n\nLOCATION (lock): Inside the stopped car at the front passenger seat on the riverside road, enclosed by the adjacent seat, dashboard, window, and controls. The shot takes place here.\n\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S16sh9_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대~2010년대 초반 한국 국산차 내부 대시보드 및 조수석: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_e89b15ab00eb5e8e.png"
    },
    {
     "label": "김선영",
     "path": "<bytes:870689>"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S16sh9_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대~2010년대 초반 한국 국산차 내부 대시보드 및 조수석: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_e89b15ab00eb5e8e.png"
    },
    {
     "label": "김선영",
     "path": "<bytes:870689>"
    }
   ]
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "김선영은 오른쪽 어깨 너머로 뒤쪽(카메라 및 뒷좌석 방향)을 돌아보고 있음.",
    "built_space": "운전대 1개, 대시보드, 센터 콘솔이 보임. 카메라는 지시된 센터 콘솔 근처가 아닌 뒷좌석 공간에 위치해 앞을 향하고 있음.",
    "entities": "김선영(젊은 여성, 보라색 패딩, 검은색 비니)이 조수석에 위치하며 지시사항과 일치함.",
    "hard_violations": [],
    "physics": "김선영의 하중은 조수석 시트에 의해 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "김선영은 오른쪽 어깨 너머로 뒤쪽(카메라 및 뒷좌석 방향)을 응시하고 있음.",
    "built_space": "운전대 1개와 대시보드가 보이며, 카메라는 지시를 어기고 뒷좌석 위치에서 앞을 바라보게 배치됨.",
    "entities": "김선영(보라색 패딩, 검은색 비니 착용)의 인물 설정은 일치함.",
    "hard_violations": [],
    "physics": "김선영은 조수석에 앉아 체중을 지탱하고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 5,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "지정된 클로즈업 스케일과 센터 콘솔 카메라 위치를 무시하고 뒷좌석에서의 와이드 샷을 연출하여 프레이밍 지시를 크게 어겼으나, 차량 내부 레퍼런스 재현도가 B보다 낫습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "클로즈업 프레이밍 및 카메라 위치 지시를 완전히 무시한 와이드 샷이며, 차량 레퍼런스(운전대 로고 등)의 디테일 재현도도 A에 비해 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S16sh9_confinedfp.png"
   },
   {
    "label": "PERIOD REFERENCE — 2000년대~2010년대 초반 한국 국산차 내부 대시보드 및 조수석: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_e89b15ab00eb5e8e.png"
   },
   {
    "label": "김선영",
    "path": "<bytes:870689>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "플로어플랜과 달리 김선영이 카메라를 향해 얼굴을 돌리고 있어 창백한 얼굴이 정면으로 노출된다.",
     "fix_en": "Redraw the character to face entirely forward, looking away from the camera, showing only the back of her black beanie and purple puffer jacket with no face visible. Maintain the character's position, clothing, the car interior set, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "장면에 거울이 없어야 한다는 명시적 지시에도 불구하고 룸미러와 우측 사이드미러가 포함되어 있습니다.",
     "fix_en": "Remove the rear-view mirror from the windshield and the side mirror from the right window, filling those areas with the dark night landscape and distant city lights. Maintain the character, her position, clothing, the car interior set, the lighting, and the framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "카메라가 뒷좌석에서 조수석 등받이 뒤만 보도록 고정된 구도가 아니라 실내 전체가 넓게 잡힌다.",
     "fix_en": "Darken the driver's seat and steering wheel on the left side with deep shadows to minimize their presence, focusing attention entirely on the passenger seat. Maintain the character, her position, clothing, the car interior set, the lighting, and the framing.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "평면도와 레이아웃 지시와 달리, 인물이 카메라를 등지지 않고 뒤돌아 얼굴을 완전히 노출하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "장면에 거울이 없어야 한다는 명시적 지시에도 불구하고 룸미러와 우측 사이드미러가 포함되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "인물이 카메라 렌즈를 직접 응시하고 있어 관찰되지 않는 상황이라는 연출에 어긋납니다.",
     "severity": "major"
    },
    {
     "issue_ko": "플로어플랜과 달리 김선영이 카메라를 향해 얼굴을 돌리고 있어 창백한 얼굴이 정면으로 노출된다.",
     "severity": "critical"
    },
    {
     "issue_ko": "장면 레이아웃에 없다고 한 룸미러가 전면 유리 상단에 보인다.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 뒷좌석에서 조수석 등받이 뒤만 보도록 고정된 구도가 아니라 실내 전체가 넓게 잡힌다.",
     "severity": "major"
    },
    {
     "issue_ko": "플로어플랜에 없는 운전석과 운전대가 화면 왼쪽에 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "장면에 거울이 없어야 한다는 명시적 지시에도 불구하고 룸미러와 우측 사이드미러가 포함되어 있습니다.",
    "fix_en": "Remove the rear-view mirror from the windshield and the side mirror from the right window, filling those areas with the dark night landscape and distant city lights. Maintain the character, her position, clothing, the car interior set, the lighting, and the framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "카메라가 뒷좌석에서 조수석 등받이 뒤만 보도록 고정된 구도가 아니라 실내 전체가 넓게 잡힌다.",
    "fix_en": "Darken the driver's seat and steering wheel on the left side with deep shadows to minimize their presence, focusing attention entirely on the passenger seat. Maintain the character, her position, clothing, the car interior set, the lighting, and the framing.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the character to face entirely forward, looking away from the camera, showing only the back of her black beanie and purple puffer jacket with no face visible. Maintain the character's position, clothing, the car interior set, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "샷 텍스트의 표정 묘사를 따르느라 평면도의 '뒷모습만 보여야 한다'는 지침을 어기고 카메라를 마주보았으나, 조수석을 중앙에 둔 지정된 카메라 구도, 심야 강변도로의 배경, 인물의 정체성 및 지정된 감정 표현을 매우 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "지시대로 인물의 뒷모습을 배치했으나, 레퍼런스 이미지의 주간 조명과 구도를 그대로 복사하여 심야(late night) 시간대 Lock을 위반했고 카메라가 조수석 뒤가 아닌 차량 중앙에 위치하는 등 핵심 지시사항을 무시했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "김선영이 뒤를 돌아 카메라를 정면으로 응시하며 눈을 크게 뜨고 있음.",
      "built_space": "야간의 자동차 내부. 카메라가 조수석 뒤에 위치하여 조수석이 중앙 전경을 차지함. 왼쪽에는 운전석과 스티어링 휠, 정면 배경에는 조수석 측 대시보드가 보임. 우측 문에는 창문 제어기가 있고 창밖으로 어두운 강변 도로가 보임.",
      "entities": "김선영(보라색 패딩, 검은 비니 착용)의 외모가 레퍼런스와 일치하며 창백한 얼굴과 확장된 동공이 잘 묘사됨. 대시보드는 2000년대 국산차 레퍼런스의 우드 트림과 디자인을 야간 조명에 맞춰 반영함.",
      "hard_violations": [],
      "physics": "조수석에 앉은 채로 상체를 뒤로 비틀어 자세를 취하고 있으며, 시트가 몸을 자연스럽게 지탱함."
     },
     {
      "label": "B",
      "direction": "인물이 카메라를 등지고 앞유리 쪽을 향해 전방을 주시함.",
      "built_space": "주간의 자동차 내부. 카메라가 뒷좌석 중앙에 위치하여 스티어링 휠, 센터 콘솔, 내비게이션 등 대시보드 전체가 정면에 보이며 조수석은 우측으로 치우쳐 있음.",
      "entities": "김선영의 뒷모습(보라색 패딩, 검은 비니). 차량 내부 대시보드는 레퍼런스 이미지를 완전히 복사하여 붙여넣은 상태임.",
      "hard_violations": [],
      "physics": "조수석에 앉아 있으며 시트가 몸을 지지함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "샷 텍스트의 표정 묘사를 따르느라 평면도의 '뒷모습만 보여야 한다'는 지침을 어기고 카메라를 마주보았으나, 조수석을 중앙에 둔 지정된 카메라 구도, 심야 강변도로의 배경, 인물의 정체성 및 지정된 감정 표현을 매우 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "지시대로 인물의 뒷모습을 배치했으나, 레퍼런스 이미지의 주간 조명과 구도를 그대로 복사하여 심야(late night) 시간대 Lock을 위반했고 카메라가 조수석 뒤가 아닌 차량 중앙에 위치하는 등 핵심 지시사항을 무시했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "김선영이 뒤를 돌아 카메라를 정면으로 응시하며 눈을 크게 뜨고 있음.",
      "built_space": "야간의 자동차 내부. 카메라가 조수석 뒤에 위치하여 조수석이 중앙 전경을 차지함. 왼쪽에는 운전석과 스티어링 휠, 정면 배경에는 조수석 측 대시보드가 보임. 우측 문에는 창문 제어기가 있고 창밖으로 어두운 강변 도로가 보임.",
      "entities": "김선영(보라색 패딩, 검은 비니 착용)의 외모가 레퍼런스와 일치하며 창백한 얼굴과 확장된 동공이 잘 묘사됨. 대시보드는 2000년대 국산차 레퍼런스의 우드 트림과 디자인을 야간 조명에 맞춰 반영함.",
      "hard_violations": [],
      "physics": "조수석에 앉은 채로 상체를 뒤로 비틀어 자세를 취하고 있으며, 시트가 몸을 자연스럽게 지탱함."
     },
     {
      "label": "B",
      "direction": "인물이 카메라를 등지고 앞유리 쪽을 향해 전방을 주시함.",
      "built_space": "주간의 자동차 내부. 카메라가 뒷좌석 중앙에 위치하여 스티어링 휠, 센터 콘솔, 내비게이션 등 대시보드 전체가 정면에 보이며 조수석은 우측으로 치우쳐 있음.",
      "entities": "김선영의 뒷모습(보라색 패딩, 검은 비니). 차량 내부 대시보드는 레퍼런스 이미지를 완전히 복사하여 붙여넣은 상태임.",
      "hard_violations": [],
      "physics": "조수석에 앉아 있으며 시트가 몸을 지지함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "고정 조건인 심야 강변 도로 배경을 완벽히 지켰으며, 숏 텍스트가 요구한 인물의 창백하고 굳은 표정을 훌륭하게 포착했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "고정 조건인 심야 시간대를 위반하여 밝은 낮으로 묘사했으며, 숏 텍스트의 핵심인 표정을 전혀 보여주지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물이 차량 정면 유리창을 향해 똑바로 앉아 있음.",
      "built_space": "차량 내부. 스티어링 휠 1개, 조수석에 인물 착석. 시점은 뒷좌석.",
      "entities": "보라색 패딩과 비니를 착용. 뒷모습만 보여 얼굴과 표정 확인 불가.",
      "hard_violations": [
       "시간대 고정 조건(late night)을 위반하여 밝은 환경으로 묘사됨."
      ],
      "physics": "조수석 시트에 앉아 안정적으로 지지됨."
     },
     {
      "label": "B",
      "direction": "인물이 왼쪽 어깨 너머로 뒤를 돌아 카메라를 응시함.",
      "built_space": "차량 내부 뒷좌석 시점. 스티어링 휠 1개, 조수석 착석. 우측 창밖으로 강변 야경 보임.",
      "entities": "레퍼런스 인물과 일치하는 외모 및 의상. 지시문에 맞는 창백한 얼굴과 확장된 동공 묘사됨.",
      "hard_violations": [],
      "physics": "조수석 시트에 앉은 채 상체를 뒤로 튼 상태로 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "고정 조건인 심야 강변 도로 배경을 완벽히 지켰으며, 숏 텍스트가 요구한 인물의 창백하고 굳은 표정을 훌륭하게 포착했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "고정 조건인 심야 시간대를 위반하여 밝은 낮으로 묘사했으며, 숏 텍스트의 핵심인 표정을 전혀 보여주지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물이 차량 정면 유리창을 향해 똑바로 앉아 있음.",
      "built_space": "차량 내부. 스티어링 휠 1개, 조수석에 인물 착석. 시점은 뒷좌석.",
      "entities": "보라색 패딩과 비니를 착용. 뒷모습만 보여 얼굴과 표정 확인 불가.",
      "hard_violations": [
       "시간대 고정 조건(late night)을 위반하여 밝은 환경으로 묘사됨."
      ],
      "physics": "조수석 시트에 앉아 안정적으로 지지됨."
     },
     {
      "label": "A",
      "direction": "인물이 왼쪽 어깨 너머로 뒤를 돌아 카메라를 응시함.",
      "built_space": "차량 내부 뒷좌석 시점. 스티어링 휠 1개, 조수석 착석. 우측 창밖으로 강변 야경 보임.",
      "entities": "레퍼런스 인물과 일치하는 외모 및 의상. 지시문에 맞는 창백한 얼굴과 확장된 동공 묘사됨.",
      "hard_violations": [],
      "physics": "조수석 시트에 앉은 채 상체를 뒤로 튼 상태로 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "confined_fp": {
   "base_key": "confinedfp::7f322c8f2baf",
   "apt_reason": "차량 내부 조수석이라는 방향성과 위치 지정이 뚜렷한 공간에서 촬영되므로, 좌석 및 대시보드, 창문 등 차량 내부의 정확한 공간 배치를 위해 평면도 보조가 필요합니다.",
   "fixed": true,
   "mismatches": [
    "The text places the camera near the center console to capture a tight three-quarter facial frame of Kim Seon-yeong. The diagram places the camera directly behind the passenger seat, pointing straight forward, which would capture the back of her head and the back of the seat, not her face."
   ],
   "era_research": {
    "subject": "2000년대~2010년대 초반 한국 국산차 내부 대시보드 및 조수석",
    "queries": [
     [
      "2000년대 국산차 실내 대시보드 조수석 엔에프 쏘나타 아반떼 에이치디 에스엠파이브",
      "2010년대 초반 국산 중형차 준중형차 실내 인테리어 대시보드 조수석"
     ],
     [
      "엔에프 쏘나타 실내 조수석 대시보드 사진",
      "아반떼 에이치디 실내 대시보드 사진",
      "뉴 에스엠파이브 실내 조수석 대시보드 사진"
     ]
    ],
    "picked_url": "https://cnp1753-prod.s3.amazonaws.com/cars/6Xnd6kDaaG6jVu/nWPjoEQ6GHY58bLRlhT0.jpg",
    "sha256": "3fc806b7ba03145c0e9c77f9b477f41a7985a86d48e6d1b3ccb5eb7fb5c038c8",
    "file": "eraref_e89b15ab00eb5e8e.png"
   }
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S16sh9::cine": {
  "applied": true,
  "fingerprint": "63d1ddc8673e15fe1a3b04f833767ed876ef44dad745fc0af0cb2263e0ebf127",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S16sh9_sel.png",
  "source_sha256": "3ebb66278bd53574c8bc656e842409b0fb0da53671cdf646518672627ab64cdc",
  "file": "S16sh9_cine.png",
  "latency_ms": 10378
 },
 "S17sh2::signage": {
  "fp": "fd17a61e3fadc6ac",
  "inscriptions": [
   {
    "surface_native": "브리핑 화이트보드",
    "text_native": "사건 개요",
    "reason_ko": "수사과장실 브리핑 구역의 현장감을 살리기 위해 사건 자료가 제시되는 배경의 화이트보드에 적합한 한글 표기가 필요합니다."
   }
  ]
 },
 "S17sh2": {
  "input_fingerprint": "fe953627be22d9df",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철을 향해 시선을 고정한 채 미간을 찌푸린 전택수의 얼굴\n\nLOCATION (lock): In the briefing area of the investigation chief's office, with case materials being presented indoors. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind 강주철's shoulder and slightly above 전택수's eyeline, the static camera holds a close three-quarter view with 전택수 placed off-center and restrained look space toward 강주철. 전택수 leans subtly into the briefing with his brow tightened and eyes fixed on 강주철, whose soft shoulder and partial profile remain at the near edge to preserve the relationship axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 강주철 in the middle-left of the frame, foreground; 전택수 in the middle-right of the frame, midground, looks toward 강주철 delivering the briefing.\n- KEY BACKGROUND ELEMENTS: 브리핑 모니터 (Displaying case materials) — Part of the active screen face is visible beyond the conversational axis, carrying the case-summary presentation; used as A partial background cue that prepares the logical cut from 전택수's reaction to the case montage; 수사과장실 책상 (Used for the briefing) — Its edge recedes behind 강주철's foreground shoulder toward 전택수; used as Keeps the procedural setting legible behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast emphasizes the furrow without theatrical shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 브리핑 화이트보드: \"사건 개요\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철을 향해 시선을 고정한 채 미간을 찌푸린 전택수의 얼굴\n\nLOCATION (lock): In the briefing area of the investigation chief's office, with case materials being presented indoors. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind 강주철's shoulder and slightly above 전택수's eyeline, the static camera holds a close three-quarter view with 전택수 placed off-center and restrained look space toward 강주철. 전택수 leans subtly into the briefing with his brow tightened and eyes fixed on 강주철, whose soft shoulder and partial profile remain at the near edge to preserve the relationship axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 강주철 in the middle-left of the frame, foreground; 전택수 in the middle-right of the frame, midground, looks toward 강주철 delivering the briefing.\n- KEY BACKGROUND ELEMENTS: 브리핑 모니터 (Displaying case materials) — Part of the active screen face is visible beyond the conversational axis, carrying the case-summary presentation; used as A partial background cue that prepares the logical cut from 전택수's reaction to the case montage; 수사과장실 책상 (Used for the briefing) — Its edge recedes behind 강주철's foreground shoulder toward 전택수; used as Keeps the procedural setting legible behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast emphasizes the furrow without theatrical shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 브리핑 화이트보드: \"사건 개요\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 강주철을 향해 시선을 고정한 채 미간을 찌푸린 전택수의 얼굴\n\nLOCATION (lock): In the briefing area of the investigation chief's office, with case materials being presented indoors. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind 강주철's shoulder and slightly above 전택수's eyeline, the static camera holds a close three-quarter view with 전택수 placed off-center and restrained look space toward 강주철. 전택수 leans subtly into the briefing with his brow tightened and eyes fixed on 강주철, whose soft shoulder and partial profile remain at the near edge to preserve the relationship axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 강주철 in the middle-left of the frame, foreground; 전택수 in the middle-right of the frame, midground, looks toward 강주철 delivering the briefing.\n- KEY BACKGROUND ELEMENTS: 브리핑 모니터 (Displaying case materials) — Part of the active screen face is visible beyond the conversational axis, carrying the case-summary presentation; used as A partial background cue that prepares the logical cut from 전택수's reaction to the case montage; 수사과장실 책상 (Used for the briefing) — Its edge recedes behind 강주철's foreground shoulder toward 전택수; used as Keeps the procedural setting legible behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime office illumination with restrained neutral-to-cool color and moderate contrast emphasizes the furrow without theatrical shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 브리핑 화이트보드: \"사건 개요\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "요청된 브리핑 모니터와 '사건 개요'가 적힌 화이트보드를 모두 배경에 정확히 배치했으며, 미간을 찌푸린 인물의 표정과 구도를 훌륭하게 구현함."
   },
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "인물의 외형과 오버더숄더 구도는 양호하나, 화이트보드 없이 대형 모니터 화면에 지정된 텍스트를 배치하여 세부 배경 지시를 온전히 따르지 못함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S14sh7_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "지시사항에는 '사건 개요' 텍스트를 '브리핑 화이트보드'에 물리적인 글씨로 렌더링하도록 지정되어 있으나, 실제 이미지에서는 배경 우측의 '브리핑 모니터' 디지털 화면에 해당 텍스트가 표시되어 있습니다.",
     "fix_en": "Convert the digital screen in the background right into a physical whiteboard, replacing the digital interface with hand-drawn case materials and leaving the text as dry-erase marker writing. Preserve Jeon Taek-soo's appearance, the foreground character, and the lighting.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "전경 좌측에 위치한 인물(강주철)이 레퍼런스 이미지의 배경 우측에 앉아있던 인물과 동일한 의상(회색 정장 재킷과 흰색 셔츠)을 착용하고 있어, 레퍼런스의 다른 인물 의상을 가져오지 말라는 지시를 위반했습니다.",
     "fix_en": "Change the clothing of the foreground character on the left to a dark brown casual jacket, ensuring it does not resemble the gray suit. Preserve Jeon Taek-soo's exact appearance, pose, the background, and lighting.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "왼쪽 전경의 강주철이 부분 옆모습이 아니라 뒷머리와 어깨만 보인다",
     "fix_en": "Adjust the head angle of the foreground character on the left to reveal a partial profile, showing the edge of the cheek and nose instead of just the back of the head. Preserve Jeon Taek-soo's appearance, the background layout, and the overall lighting.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "강주철 어깨 뒤로 전택수 쪽으로 이어져야 할 수사과장실 책상 가장자리가 보이지 않는다",
     "fix_en": "Insert the dark wooden edge of the briefing desk extending from the lower left behind the foreground shoulder toward the midground right. Preserve Jeon Taek-soo's appearance, pose, the foreground character, and the background elements.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "지시사항에는 '사건 개요' 텍스트를 '브리핑 화이트보드'에 물리적인 글씨로 렌더링하도록 지정되어 있으나, 실제 이미지에서는 배경 우측의 '브리핑 모니터' 디지털 화면에 해당 텍스트가 표시되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전경 좌측에 위치한 인물(강주철)이 레퍼런스 이미지의 배경 우측에 앉아있던 인물과 동일한 의상(회색 정장 재킷과 흰색 셔츠)을 착용하고 있어, 레퍼런스의 다른 인물 의상을 가져오지 말라는 지시를 위반했습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "왼쪽 전경의 강주철이 부분 옆모습이 아니라 뒷머리와 어깨만 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "강주철 어깨 뒤로 전택수 쪽으로 이어져야 할 수사과장실 책상 가장자리가 보이지 않는다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "지시사항에는 '사건 개요' 텍스트를 '브리핑 화이트보드'에 물리적인 글씨로 렌더링하도록 지정되어 있으나, 실제 이미지에서는 배경 우측의 '브리핑 모니터' 디지털 화면에 해당 텍스트가 표시되어 있습니다.",
    "fix_en": "Convert the digital screen in the background right into a physical whiteboard, replacing the digital interface with hand-drawn case materials and leaving the text as dry-erase marker writing. Preserve Jeon Taek-soo's appearance, the foreground character, and the lighting.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "전경 좌측에 위치한 인물(강주철)이 레퍼런스 이미지의 배경 우측에 앉아있던 인물과 동일한 의상(회색 정장 재킷과 흰색 셔츠)을 착용하고 있어, 레퍼런스의 다른 인물 의상을 가져오지 말라는 지시를 위반했습니다.",
    "fix_en": "Change the clothing of the foreground character on the left to a dark brown casual jacket, ensuring it does not resemble the gray suit. Preserve Jeon Taek-soo's exact appearance, pose, the background, and lighting.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "왼쪽 전경의 강주철이 부분 옆모습이 아니라 뒷머리와 어깨만 보인다",
    "fix_en": "Adjust the head angle of the foreground character on the left to reveal a partial profile, showing the edge of the cheek and nose instead of just the back of the head. Preserve Jeon Taek-soo's appearance, the background layout, and the overall lighting.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "강주철 어깨 뒤로 전택수 쪽으로 이어져야 할 수사과장실 책상 가장자리가 보이지 않는다",
    "fix_en": "Insert the dark wooden edge of the briefing desk extending from the lower left behind the foreground shoulder toward the midground right. Preserve Jeon Taek-soo's appearance, pose, the foreground character, and the background elements.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S14sh7"
  }
 },
 "S17sh2::cine": {
  "applied": true,
  "fingerprint": "8ceb5a0b2279673ebd16f8da54f00b6d0f9a9cadbf11f2c481714392f8901f1d",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S17sh2_sel.png",
  "source_sha256": "997135d20ce0bc0b1a303283d563feb40735f4f2309e1d5ea2d8dd4318c3b814",
  "file": "S17sh2_cine.png",
  "latency_ms": 9180
 },
 "S18sh3::signage": {
  "fp": "321c5106b3fe065f",
  "inscriptions": [
   {
    "surface_native": "편의점 간판",
    "text_native": "24시 편의점",
    "reason_ko": "전택수가 응시하는 건너편 편의점 건물의 정체성을 시각적으로 전달하기 위한 간판 표기."
   },
   {
    "surface_native": "정육점 아크릴 간판",
    "text_native": "제일정육점",
    "reason_ko": "주인공이 서 있는 골목 모퉁이가 과거 정육점이었던 장소임을 나타내는 오래되고 빛바랜 간판."
   }
  ]
 },
 "S18sh3": {
  "input_fingerprint": "9cde67378c96cf87",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 골목 한가운데 멈춰 선 채 건너편 편의점 쪽으로 고개를 돌린 전택수\n\nLOCATION (lock): At a fork in an old residential alley, outside the former butcher-shop site and facing the convenience store across the way. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height on 전택수's rear three-quarter side, the lateral track settles as he comes to a complete halt near the center of the forked alley. His weight finishes over the front foot while his torso remains aligned with the route ahead and only his head turns across the frame toward the convenience-store side, with the opposing alley and storefront direction held in the right background.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward convenience store across the alley; convenience store across the alley in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 주택가 골목 갈림길 (Located between the old chicken shop and the opposite alley); used as The principal spatial junction around which 전택수 stops and redirects his attention; 옛날 통닭집 (Occupying the former butcher-shop location) — The shop frontage faces the forked alley behind 전택수; used as Marks the location where the witness last saw the victim and anchors the near side of the frame; 건너편 편의점 (Occupying the former arcade location) — Its frontage faces back toward the intersection from the opposite alley; used as Provides the visible destination of 전택수's head turn across the alley.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime exterior illumination uses restrained neutral-to-cool color and moderate contrast without added atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편의점 간판: \"24시 편의점\"\n- 정육점 아크릴 간판: \"제일정육점\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 골목 한가운데 멈춰 선 채 건너편 편의점 쪽으로 고개를 돌린 전택수\n\nLOCATION (lock): At a fork in an old residential alley, outside the former butcher-shop site and facing the convenience store across the way. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height on 전택수's rear three-quarter side, the lateral track settles as he comes to a complete halt near the center of the forked alley. His weight finishes over the front foot while his torso remains aligned with the route ahead and only his head turns across the frame toward the convenience-store side, with the opposing alley and storefront direction held in the right background.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward convenience store across the alley; convenience store across the alley in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 주택가 골목 갈림길 (Located between the old chicken shop and the opposite alley); used as The principal spatial junction around which 전택수 stops and redirects his attention; 옛날 통닭집 (Occupying the former butcher-shop location) — The shop frontage faces the forked alley behind 전택수; used as Marks the location where the witness last saw the victim and anchors the near side of the frame; 건너편 편의점 (Occupying the former arcade location) — Its frontage faces back toward the intersection from the opposite alley; used as Provides the visible destination of 전택수's head turn across the alley.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime exterior illumination uses restrained neutral-to-cool color and moderate contrast without added atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편의점 간판: \"24시 편의점\"\n- 정육점 아크릴 간판: \"제일정육점\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 골목 한가운데 멈춰 선 채 건너편 편의점 쪽으로 고개를 돌린 전택수\n\nLOCATION (lock): At a fork in an old residential alley, outside the former butcher-shop site and facing the convenience store across the way. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height on 전택수's rear three-quarter side, the lateral track settles as he comes to a complete halt near the center of the forked alley. His weight finishes over the front foot while his torso remains aligned with the route ahead and only his head turns across the frame toward the convenience-store side, with the opposing alley and storefront direction held in the right background.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward convenience store across the alley; convenience store across the alley in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 주택가 골목 갈림길 (Located between the old chicken shop and the opposite alley); used as The principal spatial junction around which 전택수 stops and redirects his attention; 옛날 통닭집 (Occupying the former butcher-shop location) — The shop frontage faces the forked alley behind 전택수; used as Marks the location where the witness last saw the victim and anchors the near side of the frame; 건너편 편의점 (Occupying the former arcade location) — Its frontage faces back toward the intersection from the opposite alley; used as Provides the visible destination of 전택수's head turn across the alley.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime exterior illumination uses restrained neutral-to-cool color and moderate contrast without added atmospheric effects.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편의점 간판: \"24시 편의점\"\n- 정육점 아크릴 간판: \"제일정육점\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선이 우측 배경의 편의점을 향함.",
    "built_space": "좌측에 정육점, 우측에 편의점이 위치한 골목 갈림길.",
    "entities": "캐릭터 레퍼런스(50대 남성, 남색 재킷, 모자)와 간판 텍스트(제일정육점, 24시 편의점)가 모두 일치함.",
    "hard_violations": [],
    "physics": "지면에 두 발을 딛고 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "시선이 우측 배경의 편의점을 향함.",
    "built_space": "좌측에 정육점, 우측에 편의점이 위치한 골목 갈림길.",
    "entities": "간판 텍스트는 일치하나, 인물이 지시된 전택수가 아닌 정장 차림의 젊은 남성임.",
    "hard_violations": [
     "프롬프트가 명시적으로 금지한 이전 샷 스틸의 인물(젊은 남성, 정장)을 등장시킴"
    ],
    "physics": "지면에 서 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "명시된 캐릭터 레퍼런스, 공간 배치, 간판 텍스트를 모두 정확하게 구현함."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "배경은 훌륭하게 묘사했으나 프롬프트에서 금지한 이전 샷 스틸의 인물을 그대로 가져온 치명적인 오류가 있음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S16sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "장소와 조명 분위기를 유지해야 하는 이전 컷(PREVIOUS SHOT) 레퍼런스 이미지(야간, 도로)와 전혀 다르게 주간 골목길로 렌더링됨.",
     "fix_en": "Change the lighting to a nighttime atmosphere and replace the daytime alley background with a road environment to match the previous shot still, keeping the character's size and placement in the frame exactly as they are.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "인물이 완전히 멈춰 섰다는 프롬프트 지시와 달리, 뒷발 뒤꿈치가 들려 있어 걷고 있는 자세임.",
     "fix_en": "Plant the character's back foot flat on the ground and bring his legs into a stationary standing stance, preserving his clothing, the camera angle, and the surrounding alley environment exactly as they are.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "몸통을 진행 방향에 두고 고개만 돌린 것이 아니라 상체 전체가 우측 편의점 쪽으로 향해 있다",
     "fix_en": "Rotate the character's torso so his chest and shoulders face forward down the alley, leaving only his head turned toward the right, without altering his jacket, pants, or the background storefronts.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "장소와 조명 분위기를 유지해야 하는 이전 컷(PREVIOUS SHOT) 레퍼런스 이미지(야간, 도로)와 전혀 다르게 주간 골목길로 렌더링됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "인물이 완전히 멈춰 섰다는 프롬프트 지시와 달리, 뒷발 뒤꿈치가 들려 있어 걷고 있는 자세임.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷 스틸에 고정된 야간 도로·차량 장소가 아니라 전혀 다른 주간 주택가 골목으로 나왔다",
     "severity": "critical"
    },
    {
     "issue_ko": "골목 한가운데 완전히 멈춘 것이 아니라 중좌측 인물이 뒷꿈치가 들린 보행 보폭으로 걷고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "몸통을 진행 방향에 두고 고개만 돌린 것이 아니라 상체 전체가 우측 편의점 쪽으로 향해 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "인물이 완전히 멈춰 섰다는 프롬프트 지시와 달리, 뒷발 뒤꿈치가 들려 있어 걷고 있는 자세임.",
    "fix_en": "Plant the character's back foot flat on the ground and bring his legs into a stationary standing stance, preserving his clothing, the camera angle, and the surrounding alley environment exactly as they are.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "몸통을 진행 방향에 두고 고개만 돌린 것이 아니라 상체 전체가 우측 편의점 쪽으로 향해 있다",
    "fix_en": "Rotate the character's torso so his chest and shoulders face forward down the alley, leaving only his head turned toward the right, without altering his jacket, pants, or the background storefronts.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the lighting to a nighttime atmosphere and replace the daytime alley background with a road environment to match the previous shot still, keeping the character's size and placement in the frame exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "명시된 낮 시간대와 주요 배경(골목 갈림길, 정육점, 편의점), 지정된 간판 텍스트를 모두 정확하게 구현하여 텍스트 프롬프트의 요구사항을 훌륭하게 충족함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "이전 샷 레퍼런스의 야간 도로 배경 분위기는 따랐으나, 프롬프트가 최우선으로 요구한 낮 시간대와 상점 배경, 구도를 완전히 누락함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 오른쪽 어깨 너머에 위치한 편의점을 향하고 있음.",
      "built_space": "골목 갈림길. 좌측에 정육점 건물, 우측에 편의점이 위치함.",
      "entities": "전택수(레퍼런스와 외모, 복장 일치), 정육점 간판(\"제일정육점\" 일치), 편의점 간판(\"24시 편의점\" 일치).",
      "hard_violations": [],
      "physics": "오른발을 앞으로 내딛고 포장된 지면을 딛고 서 있음."
     },
     {
      "label": "B",
      "direction": "시선은 오른쪽 어깨 너머 도로의 우측을 향하고 있음.",
      "built_space": "가로등이 켜진 텅 빈 왕복 4차선 도로. 골목이나 상점이 없음.",
      "entities": "전택수(레퍼런스와 외모, 복장 일치). 요구된 상점 및 간판 피사체 없음.",
      "hard_violations": [],
      "physics": "아스팔트 도로 위에 두 발로 서서 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "명시된 낮 시간대와 주요 배경(골목 갈림길, 정육점, 편의점), 지정된 간판 텍스트를 모두 정확하게 구현하여 텍스트 프롬프트의 요구사항을 훌륭하게 충족함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "이전 샷 레퍼런스의 야간 도로 배경 분위기는 따랐으나, 프롬프트가 최우선으로 요구한 낮 시간대와 상점 배경, 구도를 완전히 누락함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 오른쪽 어깨 너머에 위치한 편의점을 향하고 있음.",
      "built_space": "골목 갈림길. 좌측에 정육점 건물, 우측에 편의점이 위치함.",
      "entities": "전택수(레퍼런스와 외모, 복장 일치), 정육점 간판(\"제일정육점\" 일치), 편의점 간판(\"24시 편의점\" 일치).",
      "hard_violations": [],
      "physics": "오른발을 앞으로 내딛고 포장된 지면을 딛고 서 있음."
     },
     {
      "label": "B",
      "direction": "시선은 오른쪽 어깨 너머 도로의 우측을 향하고 있음.",
      "built_space": "가로등이 켜진 텅 빈 왕복 4차선 도로. 골목이나 상점이 없음.",
      "entities": "전택수(레퍼런스와 외모, 복장 일치). 요구된 상점 및 간판 피사체 없음.",
      "hard_violations": [],
      "physics": "아스팔트 도로 위에 두 발로 서서 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "명시된 낮 시간대, 골목길 배경, 인물의 방향, 지정된 간판 텍스트를 모두 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 낮 시간대와 주요 장소(정육점, 편의점)를 모두 누락하고 야간 도로를 배경으로 생성하여 핵심 지시를 어겼습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "인물의 시선과 고개가 우측 배경에 위치한 편의점을 정확히 향하고 있음.",
      "built_space": "골목 갈림길을 중심으로 좌측에 정육점, 우측 배경에 편의점이 프레임 지시에 맞게 배치됨.",
      "entities": "인물(전택수)의 외양과 복장이 참조와 일치함. '제일정육점', '24시 편의점' 간판 텍스트가 정확히 구현됨.",
      "hard_violations": [],
      "physics": "두 발이 지면에 닿아 있으며, 앞발에 체중이 실린 멈춰 선 자세가 지지됨."
     },
     {
      "label": "A",
      "direction": "고개를 오른쪽으로 돌렸으나 텍스트가 요구한 목표물(편의점)이 존재하지 않음.",
      "built_space": "넓은 포장도로와 가로등만 존재하며, 지시된 주택가 골목길 구조가 없음.",
      "entities": "인물 외양은 일치하나, 요구된 건물 및 간판이 전혀 없음.",
      "hard_violations": [
       "명시된 시간대(낮) 위반 및 야간 렌더링",
       "지시된 장소(골목길, 정육점, 편의점) 누락"
      ],
      "physics": "발이 도로 면에 닿아 체중을 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "명시된 낮 시간대, 골목길 배경, 인물의 방향, 지정된 간판 텍스트를 모두 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 낮 시간대와 주요 장소(정육점, 편의점)를 모두 누락하고 야간 도로를 배경으로 생성하여 핵심 지시를 어겼습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선과 고개가 우측 배경에 위치한 편의점을 정확히 향하고 있음.",
      "built_space": "골목 갈림길을 중심으로 좌측에 정육점, 우측 배경에 편의점이 프레임 지시에 맞게 배치됨.",
      "entities": "인물(전택수)의 외양과 복장이 참조와 일치함. '제일정육점', '24시 편의점' 간판 텍스트가 정확히 구현됨.",
      "hard_violations": [],
      "physics": "두 발이 지면에 닿아 있으며, 앞발에 체중이 실린 멈춰 선 자세가 지지됨."
     },
     {
      "label": "B",
      "direction": "고개를 오른쪽으로 돌렸으나 텍스트가 요구한 목표물(편의점)이 존재하지 않음.",
      "built_space": "넓은 포장도로와 가로등만 존재하며, 지시된 주택가 골목길 구조가 없음.",
      "entities": "인물 외양은 일치하나, 요구된 건물 및 간판이 전혀 없음.",
      "hard_violations": [
       "명시된 시간대(낮) 위반 및 야간 렌더링",
       "지시된 장소(골목길, 정육점, 편의점) 누락"
      ],
      "physics": "발이 도로 면에 닿아 체중을 지지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 7
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S16sh3"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S18sh3::cine": {
  "applied": true,
  "fingerprint": "9f54c85f14d7d64e43d50f5e776fb8796b962bbaa18cef2c3f000651ea5eae12",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S18sh3_sel.png",
  "source_sha256": "049c24da6e841a704715f99d580f5d462fe3d43dca1c94006274247b69d5c59a",
  "file": "S18sh3_cine.png",
  "latency_ms": 10546
 },
 "S18sh4::signage": {
  "fp": "172efd2d11c9a1b8",
  "inscriptions": [
   {
    "surface_native": "편의점 간판",
    "text_native": "24시 편의점",
    "reason_ko": "골목길 모퉁이에 있는 편의점의 외경을 자연스럽고 사실적으로 나타내기 위해 한글로 된 간판 표기가 필요합니다."
   }
  ]
 },
 "S18sh4": {
  "input_fingerprint": "93f6e71f021cae2c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 편의점 쪽을 손가락으로 가리킨 채 입을 벌린 서의용의 측면\n\nLOCATION (lock): In the same residential alley junction, with a clear exterior view toward the convenience store occupying the former arcade site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height beside 전택수, hold a tight medium side profile of 서의용, perpendicular to his pointing direction; his open mouth sits in the upper-left area while his extended hand leads across the lower frame toward the convenience store and branching alleys on the right. Keep the camera static and let his gaze remain absorbed along the route he is explaining rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, points to convenience store across the junction; convenience store across the junction in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 편의점 (visible across the opposite alley) — Its street-facing side is seen beyond 서의용's extended hand; used as Visible endpoint of the pointing gesture across the junction; 갈림길과 좌우 골목 (branching beside the convenience-store direction); used as Provides the spatial split described by 서의용, with the pointing arm visually dividing the routes; 옛날 통닭집 (present at the junction) — The side facing the junction remains partially visible behind 서의용; used as Anchors the immediate location behind the explanation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight is rendered with restrained neutral-to-cool color, moderate contrast, and a naturalistic documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same old residential alley intersection, storefronts, and daylight. Exclude the standing senior officer and retain only the younger officer pointing across the alley.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편의점 간판: \"24시 편의점\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 편의점 쪽을 손가락으로 가리킨 채 입을 벌린 서의용의 측면\n\nLOCATION (lock): In the same residential alley junction, with a clear exterior view toward the convenience store occupying the former arcade site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height beside 전택수, hold a tight medium side profile of 서의용, perpendicular to his pointing direction; his open mouth sits in the upper-left area while his extended hand leads across the lower frame toward the convenience store and branching alleys on the right. Keep the camera static and let his gaze remain absorbed along the route he is explaining rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, points to convenience store across the junction; convenience store across the junction in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 편의점 (visible across the opposite alley) — Its street-facing side is seen beyond 서의용's extended hand; used as Visible endpoint of the pointing gesture across the junction; 갈림길과 좌우 골목 (branching beside the convenience-store direction); used as Provides the spatial split described by 서의용, with the pointing arm visually dividing the routes; 옛날 통닭집 (present at the junction) — The side facing the junction remains partially visible behind 서의용; used as Anchors the immediate location behind the explanation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight is rendered with restrained neutral-to-cool color, moderate contrast, and a naturalistic documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same old residential alley intersection, storefronts, and daylight. Exclude the standing senior officer and retain only the younger officer pointing across the alley.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편의점 간판: \"24시 편의점\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 편의점 쪽을 손가락으로 가리킨 채 입을 벌린 서의용의 측면\n\nLOCATION (lock): In the same residential alley junction, with a clear exterior view toward the convenience store occupying the former arcade site. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height beside 전택수, hold a tight medium side profile of 서의용, perpendicular to his pointing direction; his open mouth sits in the upper-left area while his extended hand leads across the lower frame toward the convenience store and branching alleys on the right. Keep the camera static and let his gaze remain absorbed along the route he is explaining rather than toward the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, points to convenience store across the junction; convenience store across the junction in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 편의점 (visible across the opposite alley) — Its street-facing side is seen beyond 서의용's extended hand; used as Visible endpoint of the pointing gesture across the junction; 갈림길과 좌우 골목 (branching beside the convenience-store direction); used as Provides the spatial split described by 서의용, with the pointing arm visually dividing the routes; 옛날 통닭집 (present at the junction) — The side facing the junction remains partially visible behind 서의용; used as Anchors the immediate location behind the explanation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight is rendered with restrained neutral-to-cool color, moderate contrast, and a naturalistic documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same old residential alley intersection, storefronts, and daylight. Exclude the standing senior officer and retain only the younger officer pointing across the alley.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편의점 간판: \"24시 편의점\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "서의용이 우측 편의점을 향해 왼손으로 가리키며 시선을 두고 있음.",
    "built_space": "이전 샷의 골목 교차로 구조가 유지되었으며, 좌측 정육점과 우측 편의점이 올바르게 배치됨.",
    "entities": "캐릭터의 복장과 좌측 허리 파우치 위치가 레퍼런스와 일치하며 '24시 편의점' 텍스트가 정확함. 단, 프롬프트에 명시된 옛날 통닭집은 누락됨.",
    "hard_violations": [],
    "physics": "지면에 안정적으로 서서 뻗은 팔과 몸의 무게 중심이 자연스럽게 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "서의용이 우측 편의점을 향해 왼손으로 가리키며 시선을 두고 있음.",
    "built_space": "교차로 구조가 구현되었으나, 좌측 정육점 건물에 기존에 없던 형태의 간판이 추가됨.",
    "entities": "캐릭터의 권총 홀스터가 우측에서 좌측으로 반전되어 묘사되었으며, 프롬프트에서 금지한 추가 텍스트('옛날통닭')가 등장함.",
    "hard_violations": [
     "프롬프트 텍스트('옛날 통닭집')가 이미지 내에 금지된 텍스트('옛날통닭')로 유출됨"
    ],
    "physics": "지면에 서 있으나 가리키는 왼손의 손가락 길이와 관절 형태가 비정상적으로 묘사됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "레퍼런스의 배경 요소와 캐릭터 디테일을 완벽히 유지하며 자연스러운 자세와 간판 텍스트를 구현함."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "캐릭터의 홀스터 위치가 반전되었고, 금지된 텍스트가 배경에 노출되었으며 손가락 묘사가 어색함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S18sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "오른손 손등이 카메라를 향하고 있음에도 엄지손가락이 위쪽에 달려 있는 치명적인 해부학적 오류가 있습니다.",
     "fix_en": "Redraw the pointing hand to correctly depict a right hand: since the back of the hand is facing the camera, place the thumb on the bottom edge. Preserve the man's identity, his face, his clothing, the set, the light, and the framing exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "뻗은 손이 화면 하단(lower frame)을 가로지르라는 지시와 달리 화면 중앙 상단에 위치해 있습니다.",
     "fix_en": "Lower the extended arm so it leads across the lower frame area.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "오른손 손등이 카메라를 향하고 있음에도 엄지손가락이 위쪽에 달려 있는 치명적인 해부학적 오류가 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "뻗은 손이 화면 하단(lower frame)을 가로지르라는 지시와 달리 화면 중앙 상단에 위치해 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "서의용이 요구된 측면 프로필이 아니라 얼굴이 카메라 쪽으로 열린 3/4 각도로 보인다.",
     "severity": "major"
    },
    {
     "issue_ko": "가리키는 손이 프레임 하단을 가로지르지 않고 화면 한가운데 높이에 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "벌린 입이 프레임 좌상단이 아니라 중상단 왼쪽에 있다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "뻗은 손이 화면 하단(lower frame)을 가로지르라는 지시와 달리 화면 중앙 상단에 위치해 있습니다.",
    "fix_en": "Lower the extended arm so it leads across the lower frame area.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the pointing hand to correctly depict a right hand: since the back of the hand is facing the camera, place the thumb on the bottom edge. Preserve the man's identity, his face, his clothing, the set, the light, and the framing exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 구도와 액션을 정확히 구현했으며, 캐릭터 레퍼런스의 의상 디테일(가죽 재킷)을 임의 변형 없이 충실히 유지하여 더 우수함."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "손 모양은 다소 자연스러우나, 레퍼런스에 없는 지퍼 디테일(재킷 옆면 및 소매)을 임의로 추가하여 의상 일관성을 훼손함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "서의용은 화면 오른쪽 배경에 위치한 편의점('24시 편의점')을 향해 손가락으로 가리키고 있으며, 시선도 가리키는 방향을 향함.",
      "built_space": "야외 골목 교차로 배경. 인물 뒤 왼쪽에는 '제일정육점' 간판이 있는 건물이, 가리키는 방향인 오른쪽 배경에는 '24시 편의점'이 올바르게 위치함.",
      "entities": "서의용(비니, 가죽 재킷, 후드티, 배지 착용한 40대 남성)의 외형과 의상이 레퍼런스와 일치함. 편의점 간판의 '24시 편의점' 텍스트가 정확히 구현됨.",
      "hard_violations": [],
      "physics": "인물은 두 발로 땅에 서 있으며, 오른팔을 들어 가리키는 자세를 스스로 자연스럽게 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "서의용은 오른쪽 배경의 편의점을 향해 손가락으로 가리키고 있으며, 시선은 그 방향을 정확히 향함.",
      "built_space": "야외 골목 교차로 배경. 왼쪽 '제일정육점' 건물과 오른쪽 '24시 편의점' 건물이 지시된 대로 배치됨.",
      "entities": "서의용의 인상착의와 '24시 편의점' 간판 텍스트는 일치하나, 가죽 재킷의 옆면과 소매에 레퍼런스에 없는 지퍼가 추가됨.",
      "hard_violations": [],
      "physics": "인물은 땅에 서 있으며, 오른팔을 뻗어 가리키는 동작을 스스로 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 구도와 액션을 정확히 구현했으며, 캐릭터 레퍼런스의 의상 디테일(가죽 재킷)을 임의 변형 없이 충실히 유지하여 더 우수함."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "손 모양은 다소 자연스러우나, 레퍼런스에 없는 지퍼 디테일(재킷 옆면 및 소매)을 임의로 추가하여 의상 일관성을 훼손함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서의용은 화면 오른쪽 배경에 위치한 편의점('24시 편의점')을 향해 손가락으로 가리키고 있으며, 시선도 가리키는 방향을 향함.",
      "built_space": "야외 골목 교차로 배경. 인물 뒤 왼쪽에는 '제일정육점' 간판이 있는 건물이, 가리키는 방향인 오른쪽 배경에는 '24시 편의점'이 올바르게 위치함.",
      "entities": "서의용(비니, 가죽 재킷, 후드티, 배지 착용한 40대 남성)의 외형과 의상이 레퍼런스와 일치함. 편의점 간판의 '24시 편의점' 텍스트가 정확히 구현됨.",
      "hard_violations": [],
      "physics": "인물은 두 발로 땅에 서 있으며, 오른팔을 들어 가리키는 자세를 스스로 자연스럽게 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "서의용은 오른쪽 배경의 편의점을 향해 손가락으로 가리키고 있으며, 시선은 그 방향을 정확히 향함.",
      "built_space": "야외 골목 교차로 배경. 왼쪽 '제일정육점' 건물과 오른쪽 '24시 편의점' 건물이 지시된 대로 배치됨.",
      "entities": "서의용의 인상착의와 '24시 편의점' 간판 텍스트는 일치하나, 가죽 재킷의 옆면과 소매에 레퍼런스에 없는 지퍼가 추가됨.",
      "hard_violations": [],
      "physics": "인물은 땅에 서 있으며, 오른팔을 뻗어 가리키는 동작을 스스로 지탱하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1875,
      "verdict_ko": "캐릭터 레퍼런스와 얼굴 형태가 잘 일치하며, 가리키는 손가락의 묘사도 매우 자연스럽게 수정되어 프롬프트를 훌륭하게 충족합니다."
     },
     {
      "label": "A",
      "score": 1800,
      "verdict_ko": "전반적인 구도와 배경은 훌륭하나, 가리키는 오른손의 엄지손가락이 어색하게 솟아 있어 형태가 다소 부자연스럽습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.8,
      "B": 1.875
     },
     "adjusted": {
      "A": 1.8,
      "B": 1.875
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1875,
      "verdict_ko": "캐릭터 레퍼런스와 얼굴 형태가 잘 일치하며, 가리키는 손가락의 묘사도 매우 자연스럽게 수정되어 프롬프트를 훌륭하게 충족합니다."
     },
     {
      "label": "B",
      "score": 1800,
      "verdict_ko": "전반적인 구도와 배경은 훌륭하나, 가리키는 오른손의 엄지손가락이 어색하게 솟아 있어 형태가 다소 부자연스럽습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1882,
     "B": 1805
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S18sh3"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S18sh4::cine": {
  "applied": true,
  "fingerprint": "e0a27f42d2dcc8ec80912ce74c4e0d1af49072ce3438367cef1ea738db849c16",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S18sh4_sel.png",
  "source_sha256": "771ede261ae3b611fae09e66683a02cc77df41c92b91efe7538d64da8acb28af",
  "file": "S18sh4_cine.png",
  "latency_ms": 11651
 },
 "S19sh2::confined_fp_apt": {
  "applies": true,
  "reason_ko": "승합차 운전석이라는 통제된 내부 공간을 배경으로 하며, 운전자가 룸미러를 통해 뒷좌석의 인물을 바라보는 시선과 위치 관계가 시각적으로 정확히 성립해야 하는 장면이기 때문입니다.",
  "input_fingerprint": "7ae937e9be459b20"
 },
 "S19sh2::signage": {
  "fp": "5aae016226b265a9",
  "inscriptions": []
 },
 "confinedfp::e3c2eda228bd": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/confinedfp_base_e3c2eda228bd.png",
  "place_text": "Inside the moving van's driver compartment, framed by the rearview mirror, front seat, dashboard, and windshield.",
  "input_fingerprint": "43813def7b193672"
 },
 "S19sh2::confined_fp": {
  "reads": {
   "controls": "A steering wheel is attached by a line to the outer left side of the front left seat.",
   "mirrors": "A rearview mirror is positioned at the top center of the windshield, with its reflective surface facing back toward the interior cabin.",
   "camera": "No camera symbol is present on the diagram.",
   "occupants": "서의용 occupies the front left seat, facing forward. 전택수 occupies the right side of the rear seat, facing forward."
  },
  "mismatches": [
   "The text provides specific camera placement and a close-up framing instruction, but the diagram completely omits a camera.",
   "The diagram places the steering wheel attached to the exterior left side of the driver's seat, rather than in front of the driver between the seat and dashboard."
  ],
  "scene_description_en": "The camera looks forward through the vehicle cabin from a position in the rear. In the near right foreground, 전택수 sits in the rear seat facing away toward the windshield, positioned next to a rear side window on the far right edge. In the midground left, 서의용 sits in the front seat facing forward, with a steering wheel attached to his outer left side. The front passenger seat in the midground right remains empty. In the far background, a dashboard and windshield span the width of the frame. Mounted at the top center of the windshield is a rearview mirror, its surface facing backward to reflect the occupants.",
  "fixed": true,
  "input_fingerprint": "5e3db88fd32622ea"
 },
 "era_assess::e06a9a8865c8fb0c": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 이삿짐 트럭(탑차) 운전석 내부",
    "search_terms_native": [
     "포터2 운전석 내부",
     "봉고3 대시보드",
     "마이티 트럭 내부"
    ],
    "language_lock_native": "검색어는 오직 한국어로만 구성되어야 하며 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 한국의 대표적인 트럭인 포터나 봉고, 마이티의 대시보드, 운전대 및 인테리어 구조는 서구형 밴이나 일반 승용차와 크게 달라 고증이 필수적입니다."
   }
  ]
 },
 "era_ref::b8e36c40d24f884b": {
  "subject": "2000년대 한국 이삿짐 트럭(탑차) 운전석 내부",
  "terms": [
   "포터2 운전석 내부",
   "봉고3 대시보드",
   "마이티 트럭 내부"
  ],
  "queries": [
   [
    "2000년대 포터2 운전석 내부 봉고3 대시보드 마이티 트럭 내부",
    "한국 이삿짐 탑차 운전석 내부 포터2 봉고3 마이티"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://www.orbisauto.com/images/cars/36865eb7-b41f-448b-a920-de1e909da806.png",
  "picked_reason_ko": "4번은 2000년대 한국에서 흔히 쓰인 현대 소형 상용트럭 계열의 운전석을 일상적인 사용 상태로 보여 주며, 대시보드·좌석·변속기·수납부의 형태와 재질이 가장 명확하게 읽힌다.",
  "sha256": "63bce373e81175acf7a7f72638b949711d24a90d688b6c55556517c659483d87",
  "file": "eraref_b8e36c40d24f884b.png"
 },
 "S19sh2": {
  "input_fingerprint": "be31378b8524a229",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전석 룸미러를 통해 뒷좌석의 전택수를 곁눈질로 바라보는 서의용의 눈빛 클로즈업\n\nLOCATION (lock): Inside the moving van's driver compartment, framed by the rearview mirror, front seat, dashboard, and windshield. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Continue the dolly inward from just behind the passenger side of the driver's seat until the rearview mirror and 서의용's reflected eyes form a close composition, viewed obliquely rather than on his frontal axis. His eyes occupy the upper central area as they cut toward 전택수 in the rear seat, while 전택수 remains a soft profile near the side window behind the mirror line, absorbed in the passing view.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 서의용 reflected eyes in the upper-center of the frame, foreground, looks toward 전택수 in the rear seat; 전택수 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 룸미러 (positioned above the front seats) — The mirror reflects 서의용's eyes and aligns their glance toward the rear-seat area; used as Primary reflective frame isolating 서의용's sideways glance; 뒷좌석 옆 차창 (rural scenery passing beyond it) — The rural scenery is visible beyond 전택수's profile; used as Keeps 전택수 connected to the rural landscape he is watching; 운전석과 조수석 (occupied driving interior) — Their rear and side edges face the camera and narrow the view toward the mirror; used as Creates the close interior obstruction around the mirror composition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light in the vehicle is kept subdued and natural, with neutral-to-cool rendering and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera looks forward through the vehicle cabin from a position in the rear. In the near right foreground, 전택수 sits in the rear seat facing away toward the windshield, positioned next to a rear side window on the far right edge. In the midground left, 서의용 sits in the front seat facing forward, with a steering wheel attached to his outer left side. The front passenger seat in the midground right remains empty. In the far background, a dashboard and windshield span the width of the frame. Mounted at the top center of the windshield is a rearview mirror, its surface facing backward to reflect the occupants.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전석 룸미러를 통해 뒷좌석의 전택수를 곁눈질로 바라보는 서의용의 눈빛 클로즈업\n\nLOCATION (lock): Inside the moving van's driver compartment, framed by the rearview mirror, front seat, dashboard, and windshield. The shot takes place here.\n",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera looks forward through the vehicle cabin from a position in the rear. In the near right foreground, 전택수 sits in the rear seat facing away toward the windshield, positioned next to a rear side window on the far right edge. In the midground left, 서의용 sits in the front seat facing forward, with a steering wheel attached to his outer left side. The front passenger seat in the midground right remains empty. In the far background, a dashboard and windshield span the width of the frame. Mounted at the top center of the windshield is a rearview mirror, its surface facing backward to reflect the occupants.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 운전석 룸미러를 통해 뒷좌석의 전택수를 곁눈질로 바라보는 서의용의 눈빛 클로즈업\n\nLOCATION (lock): Inside the moving van's driver compartment, framed by the rearview mirror, front seat, dashboard, and windshield. The shot takes place here.\n\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S19sh2_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대 한국 이삿짐 트럭(탑차) 운전석 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_b8e36c40d24f884b.png"
    },
    {
     "label": "서의용",
     "path": "<bytes:929743>"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S19sh2_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대 한국 이삿짐 트럭(탑차) 운전석 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_b8e36c40d24f884b.png"
    },
    {
     "label": "서의용",
     "path": "<bytes:929743>"
    }
   ]
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "거울 속 서의용은 뒷좌석의 전택수를 곁눈질하고 있으며, 전택수는 측면 창밖을 바라봄.",
    "built_space": "차량 내부에서 정면 앞유리를 향하는 시점이며 좌측 하단에 계기판이 있으나, 정작 있어야 할 운전석의 실제 운전자가 화면에서 누락됨.",
    "entities": "거울에 비친 서의용(비니, 가죽재킷)과 전택수의 외형은 프롬프트 지시와 대체로 일치함.",
    "hard_violations": [
     "물리적 운전자의 누락 (거울엔 보이나 실제 운전석은 비어 있음)",
     "카메라 시점과 실제 차량 구조가 양립할 수 없는 불가능한 스테이징"
    ],
    "physics": "실제 물리적 공간에는 운전자가 없는데 룸미러 속에만 운전자가 존재하는 불가능한 반사 및 형태를 보임."
   },
   {
    "label": "B",
    "direction": "전경의 서의용이 몸과 고개를 돌려 카메라 측(뒷좌석 방향)을 직접 응시함.",
    "built_space": "카메라가 조수석 뒤편에서 운전석을 바라보는 구도이며, 상단에 룸미러가 위치함.",
    "entities": "서의용의 복장(비니, 가죽재킷)은 일치하나, 거울 속에 비친 인물은 밝은 갈색 옷을 입은 완전히 다른 사람임.",
    "hard_violations": [
     "거울 속에 전혀 다른 인물이 비침 (발명된 인물 및 불가능한 반사)",
     "실제 인물은 뒤를 돌아보나 거울 속 인물은 정면을 향하는 물리적 모순"
    ],
    "physics": "전경의 서의용과 룸미러 속 인물의 자세, 시선, 복장이 완전히 엇갈리는 물리적으로 불가능한 거울 반사임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 3,
   "B": 2
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "룸미러 속에 서의용과 전택수의 구도를 담으려 한 점은 긍정적이나, 차량 앞유리를 향한 실제 공간에 운전자가 전혀 존재하지 않는 치명적인 물리적/공간적 모순이 발생하여 실격입니다."
   },
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "거울을 통한 눈빛 클로즈업이라는 프롬프트의 핵심 시점을 무시했으며, 룸미러 속에 전혀 다른 옷을 입고 다른 곳을 보는 인물이 비치는 치명적인 반사 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S19sh2_confinedfp.png"
   },
   {
    "label": "PERIOD REFERENCE — 2000년대 한국 이삿짐 트럭(탑차) 운전석 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_b8e36c40d24f884b.png"
   },
   {
    "label": "서의용",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "평면도상 후방이어야 할 카메라 위치가 대시보드 위로 잘못 설정되어, 차량 내부에 배치되어야 할 인물들이 화면에 없음.",
     "fix_en": "Maintain the current camera position. Darken the dashboard slightly to reduce its foreground prominence. Preserve the rearview mirror, reflections, and outside background.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "앞(앞유리)을 보고 앉은 뒷좌석 인물이 룸미러에는 뒷통수만 보이는 광학적으로 불가능한 반사가 나타남.",
     "fix_en": "Change the reflection of the passenger on the right side of the rearview mirror to show his face and front torso instead of the back of his head. Preserve the driver's reflection, the mirror frame, and the dashboard.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "앞유리 상단에 부착된 룸미러에 스티어링 휠 전체가 반사되는 것은 광학적으로 불가능함.",
     "fix_en": "Remove the steering wheel reflected in the lower left of the rearview mirror, replacing it with the continuation of the driver's dark jacket. Preserve the driver's face, the mirror frame, and the dashboard.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "서의용의 눈빛 클로즈업을 지시했으나, 룸미러 속에는 상반신 전체가 포함된 미디엄 숏으로 연출됨.",
     "fix_en": "Enlarge the driver's face within the rearview mirror reflection to focus more on his eyes. Preserve the mirror boundaries and the dashboard.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "평면도상 후방이어야 할 카메라 위치가 대시보드 위로 잘못 설정되어, 차량 내부에 배치되어야 할 인물들이 화면에 없음.",
     "severity": "critical"
    },
    {
     "issue_ko": "앞(앞유리)을 보고 앉은 뒷좌석 인물이 룸미러에는 뒷통수만 보이는 광학적으로 불가능한 반사가 나타남.",
     "severity": "critical"
    },
    {
     "issue_ko": "앞유리 상단에 부착된 룸미러에 스티어링 휠 전체가 반사되는 것은 광학적으로 불가능함.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용의 눈빛 클로즈업을 지시했으나, 룸미러 속에는 상반신 전체가 포함된 미디엄 숏으로 연출됨.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 후방에서 전방을 보는 배치가 아니라 하단 대시보드 쪽에서 룸미러를 정면으로 담았다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 중앙 룸미러에 운전석 서의용 얼굴 전체와 뒷좌석 인물이 나란히 들어와 광학적으로 불가능한 반사다.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용 눈빛 클로즈업이 아니라 룸미러 전체와 대시보드가 넓게 보이는 샷이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "서의용의 눈빛 클로즈업을 지시했으나, 룸미러 속에는 상반신 전체가 포함된 미디엄 숏으로 연출됨.",
    "fix_en": "Enlarge the driver's face within the rearview mirror reflection to focus more on his eyes. Preserve the mirror boundaries and the dashboard.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Maintain the current camera position. Darken the dashboard slightly to reduce its foreground prominence. Preserve the rearview mirror, reflections, and outside background.\n- Change the reflection of the passenger on the right side of the rearview mirror to show his face and front torso instead of the back of his head. Preserve the driver's reflection, the mirror frame, and the dashboard.\n- Remove the steering wheel reflected in the lower left of the rearview mirror, replacing it with the continuation of the driver's dark jacket. Preserve the driver's face, the mirror frame, and the dashboard.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "룸미러를 통한 눈빛 클로즈업이라는 샷 텍스트의 구도를 정확히 구현했으며, 캐릭터의 외형과 시선 처리도 레퍼런스와 일치합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 클로즈업 구도를 무시한 채 배경 레퍼런스만 그대로 복사했으며, 스티어링 휠의 구조가 무너진 심각한 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "거울 속 서의용의 시선이 옆(뒷좌석의 인물 방향)을 향하고 있음.",
      "built_space": "카메라가 대시보드와 룸미러를 클로즈업으로 잡고 있으며, 거울을 통해 운전석과 뒷좌석의 공간이 반사되어 보임.",
      "entities": "서의용은 비니, 후드, 가죽 재킷 등 레퍼런스의 복장과 외형을 정확히 갖추고 있음.",
      "hard_violations": [],
      "physics": "모든 인물과 사물이 물리적으로 올바르게 위치하고 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "거울 속 인물들의 시선이 특별한 목적지 없이 정면을 향함.",
      "built_space": "차량 내부 전체를 보여주는 와이드 샷이며, 프롬프트가 요구한 뒷좌석 공간이 존재하지 않음.",
      "entities": "거울 속에 아주 작게 묘사된 인물들의 디테일이나 신원을 파악하기 어려움.",
      "hard_violations": [
       "물리적으로 불가능하게 일그러지고 반파된 형태의 스티어링 휠",
       "플로어 플랜에서 요구한 뒷좌석 공간 및 해당 위치의 인물 누락"
      ],
      "physics": "스티어링 휠의 절반이 사라진 채 기형적으로 공중에 떠 있는 듯한 오류가 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "룸미러를 통한 눈빛 클로즈업이라는 샷 텍스트의 구도를 정확히 구현했으며, 캐릭터의 외형과 시선 처리도 레퍼런스와 일치합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 클로즈업 구도를 무시한 채 배경 레퍼런스만 그대로 복사했으며, 스티어링 휠의 구조가 무너진 심각한 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "거울 속 서의용의 시선이 옆(뒷좌석의 인물 방향)을 향하고 있음.",
      "built_space": "카메라가 대시보드와 룸미러를 클로즈업으로 잡고 있으며, 거울을 통해 운전석과 뒷좌석의 공간이 반사되어 보임.",
      "entities": "서의용은 비니, 후드, 가죽 재킷 등 레퍼런스의 복장과 외형을 정확히 갖추고 있음.",
      "hard_violations": [],
      "physics": "모든 인물과 사물이 물리적으로 올바르게 위치하고 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "거울 속 인물들의 시선이 특별한 목적지 없이 정면을 향함.",
      "built_space": "차량 내부 전체를 보여주는 와이드 샷이며, 프롬프트가 요구한 뒷좌석 공간이 존재하지 않음.",
      "entities": "거울 속에 아주 작게 묘사된 인물들의 디테일이나 신원을 파악하기 어려움.",
      "hard_violations": [
       "물리적으로 불가능하게 일그러지고 반파된 형태의 스티어링 휠",
       "플로어 플랜에서 요구한 뒷좌석 공간 및 해당 위치의 인물 누락"
      ],
      "physics": "스티어링 휠의 절반이 사라진 채 기형적으로 공중에 떠 있는 듯한 오류가 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 거울 속 시선은 표현되었으나, 카메라가 뒷좌석이 아닌 대시보드에 위치하여 플로어 플랜의 카메라 위치 조건을 심각하게 위반함."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "실제 좌석에는 아무도 없는데 룸미러에만 인물들이 비치는 물리적으로 불가능한 연출이 발생하여 완전히 실패함."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "룸미러 안의 서의용이 우측 후방을 향해 곁눈질로 시선을 던짐.",
      "built_space": "하단에 대시보드와 계기판이 보여 카메라가 차량 앞쪽에 있음. 거울 속에는 서의용이 앞좌석, 전택수가 뒷좌석에 배치됨.",
      "entities": "서의용은 비니와 가죽 재킷을 입고 있으며 신원이 일치함. 뒷좌석의 전택수도 확인됨.",
      "hard_violations": [
       "명시된 카메라 위치 위반 (카메라가 뒷좌석이 아닌 대시보드 위에 위치함)"
      ],
      "physics": "카메라가 대시보드 위에서 룸미러를 촬영하고 있으며 반사된 뷰 자체는 성립하나 시점 위치가 틀림."
     },
     {
      "label": "A",
      "direction": "좌석에 인물이 없어 시선을 확인할 수 없음. 거울 속 인물들은 정면을 바라봄.",
      "built_space": "카메라는 뒷좌석 쪽에 위치해 앞을 향하며 대시보드와 빈 앞좌석이 보임.",
      "entities": "서의용과 전택수가 실제 좌석에 존재하지 않고 거울 속에만 비침.",
      "hard_violations": [
       "물리적으로 불가능한 반사 (좌석에 없는 인물이 거울에 비침)",
       "잘못된 인물 배치 (거울 속에서 두 인물이 나란히 앞좌석 쪽에 있음)"
      ],
      "physics": "실제 공간에 없는 인물들이 룸미러에 반사되어 나타나는 유령 반사 현상."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "요구된 거울 속 시선은 표현되었으나, 카메라가 뒷좌석이 아닌 대시보드에 위치하여 플로어 플랜의 카메라 위치 조건을 심각하게 위반함."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "실제 좌석에는 아무도 없는데 룸미러에만 인물들이 비치는 물리적으로 불가능한 연출이 발생하여 완전히 실패함."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "룸미러 안의 서의용이 우측 후방을 향해 곁눈질로 시선을 던짐.",
      "built_space": "하단에 대시보드와 계기판이 보여 카메라가 차량 앞쪽에 있음. 거울 속에는 서의용이 앞좌석, 전택수가 뒷좌석에 배치됨.",
      "entities": "서의용은 비니와 가죽 재킷을 입고 있으며 신원이 일치함. 뒷좌석의 전택수도 확인됨.",
      "hard_violations": [
       "명시된 카메라 위치 위반 (카메라가 뒷좌석이 아닌 대시보드 위에 위치함)"
      ],
      "physics": "카메라가 대시보드 위에서 룸미러를 촬영하고 있으며 반사된 뷰 자체는 성립하나 시점 위치가 틀림."
     },
     {
      "label": "B",
      "direction": "좌석에 인물이 없어 시선을 확인할 수 없음. 거울 속 인물들은 정면을 바라봄.",
      "built_space": "카메라는 뒷좌석 쪽에 위치해 앞을 향하며 대시보드와 빈 앞좌석이 보임.",
      "entities": "서의용과 전택수가 실제 좌석에 존재하지 않고 거울 속에만 비침.",
      "hard_violations": [
       "물리적으로 불가능한 반사 (좌석에 없는 인물이 거울에 비침)",
       "잘못된 인물 배치 (거울 속에서 두 인물이 나란히 앞좌석 쪽에 있음)"
      ],
      "physics": "실제 공간에 없는 인물들이 룸미러에 반사되어 나타나는 유령 반사 현상."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "confined_fp": {
   "base_key": "confinedfp::e3c2eda228bd",
   "apt_reason": "승합차 운전석이라는 통제된 내부 공간을 배경으로 하며, 운전자가 룸미러를 통해 뒷좌석의 인물을 바라보는 시선과 위치 관계가 시각적으로 정확히 성립해야 하는 장면이기 때문입니다.",
   "fixed": true,
   "mismatches": [
    "The text provides specific camera placement and a close-up framing instruction, but the diagram completely omits a camera.",
    "The diagram places the steering wheel attached to the exterior left side of the driver's seat, rather than in front of the driver between the seat and dashboard."
   ],
   "era_research": {
    "subject": "2000년대 한국 이삿짐 트럭(탑차) 운전석 내부",
    "queries": [
     [
      "2000년대 포터2 운전석 내부 봉고3 대시보드 마이티 트럭 내부",
      "한국 이삿짐 탑차 운전석 내부 포터2 봉고3 마이티"
     ]
    ],
    "picked_url": "https://www.orbisauto.com/images/cars/36865eb7-b41f-448b-a920-de1e909da806.png",
    "sha256": "63bce373e81175acf7a7f72638b949711d24a90d688b6c55556517c659483d87",
    "file": "eraref_b8e36c40d24f884b.png"
   }
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S19sh2::cine": {
  "applied": true,
  "fingerprint": "a85fa75a99599205e29b5fa76941fc29b0de2bb789898835f2dec1f6f5c4de5e",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S19sh2_sel.png",
  "source_sha256": "e23a70b20d7101cd2382c50e6991beb353a1234ffe21f672d768e9b4d24e58e0",
  "file": "S19sh2_cine.png",
  "latency_ms": 9794
 },
 "S20sh1::signage": {
  "fp": "ed1fbfa642cf39c7",
  "inscriptions": []
 },
 "S20sh1": {
  "input_fingerprint": "4cb070dc972124b6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 드들강 변 강둑에 선 채 유유히 흐르는 강물을 무표정하게 내려다보는 전택수의 상체\n\nLOCATION (lock): On the open riverbank embankment at the victim's discovery site, overlooking the flowing river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height on 전택수's downstream three-quarter side, hold a static medium view across his upper body toward the river below his eyeline. Place him slightly off-center with his shoulders held still and head lowered, allowing the slowly flowing water to remain the destination of his blank, absorbed gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward river below the bank; river below the bank in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 드들강 물줄기 (flowing slowly); used as The visual endpoint of 전택수's lowered gaze and the procedural focus of the location; 강둑과 강변 (overlooking the river); used as Establishes 전택수's elevated relation to the water before he lowers himself in later shots.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight is treated with a subdued neutral-to-cool palette, moderate contrast, and natural tonal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 드들강 변 강둑에 선 채 유유히 흐르는 강물을 무표정하게 내려다보는 전택수의 상체\n\nLOCATION (lock): On the open riverbank embankment at the victim's discovery site, overlooking the flowing river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height on 전택수's downstream three-quarter side, hold a static medium view across his upper body toward the river below his eyeline. Place him slightly off-center with his shoulders held still and head lowered, allowing the slowly flowing water to remain the destination of his blank, absorbed gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward river below the bank; river below the bank in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 드들강 물줄기 (flowing slowly); used as The visual endpoint of 전택수's lowered gaze and the procedural focus of the location; 강둑과 강변 (overlooking the river); used as Establishes 전택수's elevated relation to the water before he lowers himself in later shots.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight is treated with a subdued neutral-to-cool palette, moderate contrast, and natural tonal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 드들강 변 강둑에 선 채 유유히 흐르는 강물을 무표정하게 내려다보는 전택수의 상체\n\nLOCATION (lock): On the open riverbank embankment at the victim's discovery site, overlooking the flowing river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height on 전택수's downstream three-quarter side, hold a static medium view across his upper body toward the river below his eyeline. Place him slightly off-center with his shoulders held still and head lowered, allowing the slowly flowing water to remain the destination of his blank, absorbed gaze.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward river below the bank; river below the bank in the lower-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 드들강 물줄기 (flowing slowly); used as The visual endpoint of 전택수's lowered gaze and the procedural focus of the location; 강둑과 강변 (overlooking the river); used as Establishes 전택수's elevated relation to the water before he lowers himself in later shots.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight is treated with a subdued neutral-to-cool palette, moderate contrast, and natural tonal detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 우측 하단의 수면을 향해 아래로 고정되어 있음.",
    "built_space": "강둑의 경사면이 인물 뒤에 배경으로 위치하며, 강물이 인물의 주위를 둘러싸고 있음.",
    "entities": "전택수(레퍼런스와 일치하는 외모, 경찰 재킷, 수사 모자 착용)가 존재함.",
    "hard_violations": [
     "인물이 강둑 위에서 물을 내려다보아야 하나(elevated relation), 강물 속에 직접 서 있어 공간 연출을 위반함."
    ],
    "physics": "인물의 하반신이 수면에 잠겨 있어 물 바닥에 지지하고 서 있음."
   },
   {
    "label": "B",
    "direction": "시선은 프레임 우측 하단의 강물을 향해 조용히 아래로 향해 있음.",
    "built_space": "전경에 강둑의 구조물이 아웃포커싱되어 있으며, 우측 하단으로 강물이 흐름.",
    "entities": "전택수(레퍼런스와 일치하는 외모, 정확한 의상 및 모자)가 존재함.",
    "hard_violations": [],
    "physics": "전경의 강둑 구조물 뒤에 서서 고정된 자세로 지지하고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "인물이 강둑 위에 서서 수면을 내려다보는 구도를 정확히 구현하였고, 지시된 프레이밍과 캐릭터 레퍼런스를 충실히 반영했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "인물이 강둑 위가 아닌 물 속에 서 있어 '강둑에 선 채'라는 핵심 공간 연출 지시를 정면으로 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S5sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "인물의 재킷 오른쪽 소매 패치와 왼쪽 가슴 부분에 알아볼 수 없는 뭉개진 글자와 선이 들어갔습니다.",
     "fix_en": "Remove the garbled white text from the right chest and right sleeve patch, replacing it with plain navy fabric matching the jacket. Preserve the man's face, posture, cap, the rest of his clothing, the concrete embankment, and the river water.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 숏 레퍼런스에서 강물 위에 떠 있던 시신이 고정된 프랍으로 유지되지 않고 화면에서 누락되었습니다.",
     "fix_en": "Render the floating body from the previous shot in the river background. Preserve the man's pose, his clothing, the concrete bank, and the lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "잠긴 이전 장면의 강둑과 달리 계단식 콘크리트 호안이 배경을 이룬다.",
     "fix_en": "Change the stepped concrete embankment to a continuous, flat sloped surface matching the previous shot. Preserve the man, his attire, and the water.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "인물의 재킷 오른쪽 소매 패치와 왼쪽 가슴 부분에 알아볼 수 없는 뭉개진 글자와 선이 들어갔습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 숏 레퍼런스에서 강물 위에 떠 있던 시신이 고정된 프랍으로 유지되지 않고 화면에서 누락되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "잠긴 이전 장면의 강둑과 달리 계단식 콘크리트 호안이 배경을 이룬다.",
     "severity": "major"
    },
    {
     "issue_ko": "점퍼 가슴과 소매 패치의 글자가 깨져 있다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 숏 레퍼런스에서 강물 위에 떠 있던 시신이 고정된 프랍으로 유지되지 않고 화면에서 누락되었습니다.",
    "fix_en": "Render the floating body from the previous shot in the river background. Preserve the man's pose, his clothing, the concrete bank, and the lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "잠긴 이전 장면의 강둑과 달리 계단식 콘크리트 호안이 배경을 이룬다.",
    "fix_en": "Change the stepped concrete embankment to a continuous, flat sloped surface matching the previous shot. Preserve the man, his attire, and the water.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the garbled white text from the right chest and right sleeve patch, replacing it with plain navy fabric matching the jacket. Preserve the man's face, posture, cap, the rest of his clothing, the concrete embankment, and the river water.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레퍼런스에 존재하는 재킷의 패치와 수사관 식별 문구를 유지하여 의상 구현도가 더 높음."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "문구를 지우는 과정에서 레퍼런스의 필수적인 재킷 패치와 디테일까지 모두 삭제하여 의상 고증을 훼손함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선이 우측 하단의 강물을 정확히 향함.",
      "built_space": "콘크리트 강둑 구조. 카메라와 인물의 위치 관계가 적절함.",
      "entities": "전택수의 외모가 레퍼런스와 일치하며, 재킷의 패치 및 디테일이 보존됨.",
      "hard_violations": [],
      "physics": "강둑 경사면에 두 발로 서서 체중을 안정적으로 지탱함."
     },
     {
      "label": "B",
      "direction": "시선이 우측 하단의 강물을 정확히 향함.",
      "built_space": "콘크리트 강둑 구조. 카메라와 인물의 위치 관계가 적절함.",
      "entities": "전택수의 외모는 일치하나, 재킷이 패치가 없는 민무늬로 변형됨.",
      "hard_violations": [],
      "physics": "강둑 경사면에 두 발로 서서 체중을 안정적으로 지탱함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레퍼런스에 존재하는 재킷의 패치와 수사관 식별 문구를 유지하여 의상 구현도가 더 높음."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "문구를 지우는 과정에서 레퍼런스의 필수적인 재킷 패치와 디테일까지 모두 삭제하여 의상 고증을 훼손함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 우측 하단의 강물을 정확히 향함.",
      "built_space": "콘크리트 강둑 구조. 카메라와 인물의 위치 관계가 적절함.",
      "entities": "전택수의 외모가 레퍼런스와 일치하며, 재킷의 패치 및 디테일이 보존됨.",
      "hard_violations": [],
      "physics": "강둑 경사면에 두 발로 서서 체중을 안정적으로 지탱함."
     },
     {
      "label": "B",
      "direction": "시선이 우측 하단의 강물을 정확히 향함.",
      "built_space": "콘크리트 강둑 구조. 카메라와 인물의 위치 관계가 적절함.",
      "entities": "전택수의 외모는 일치하나, 재킷이 패치가 없는 민무늬로 변형됨.",
      "hard_violations": [],
      "physics": "강둑 경사면에 두 발로 서서 체중을 안정적으로 지탱함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "참조 이미지에 있는 재킷의 로고와 패치가 생략되었으나, B 후보의 깨진 텍스트와 의미 없는 문구를 깔끔하게 제거하여 전체적인 시각적 사실감이 더 높습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "재킷의 디테일을 추가하려 했으나, 참조에 없는 'SALICS'라는 영문과 해독할 수 없는 깨진 한글 패치를 생성하여 지시사항(invent none)을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선이 화면 우측 하단에 있는 흐르는 강물을 정확히 향하고 있음.",
      "built_space": "참조 이미지와 동일한 형태의 콘크리트 강둑이 구현되었으며, 인물이 둑 위에 적절히 위치함.",
      "entities": "전택수의 얼굴, 체형, 모자, 체크 셔츠가 참조와 잘 일치함. 단, 재킷의 텍스트와 패치가 생략됨. 이전 샷의 물속 시신은 지시대로 제거됨.",
      "hard_violations": [],
      "physics": "화면 밖의 발로 강둑 지면을 디디고 서 있는 자세가 물리적으로 자연스러움."
     },
     {
      "label": "B",
      "direction": "전택수의 시선이 화면 우측 하단에 있는 흐르는 강물을 정확히 향하고 있음.",
      "built_space": "참조 이미지와 동일한 형태의 콘크리트 강둑이 구현되었으며, 인물이 둑 위에 적절히 위치함.",
      "entities": "전택수의 기본 인상과 복장은 일치하나, 재킷에 원본에 없는 'SALICS' 및 깨진 글자 패치가 추가됨. 이전 샷의 시신은 지시대로 제거됨.",
      "hard_violations": [],
      "physics": "화면 밖의 발로 강둑 지면을 디디고 서 있는 자세가 물리적으로 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "참조 이미지에 있는 재킷의 로고와 패치가 생략되었으나, B 후보의 깨진 텍스트와 의미 없는 문구를 깔끔하게 제거하여 전체적인 시각적 사실감이 더 높습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "재킷의 디테일을 추가하려 했으나, 참조에 없는 'SALICS'라는 영문과 해독할 수 없는 깨진 한글 패치를 생성하여 지시사항(invent none)을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수의 시선이 화면 우측 하단에 있는 흐르는 강물을 정확히 향하고 있음.",
      "built_space": "참조 이미지와 동일한 형태의 콘크리트 강둑이 구현되었으며, 인물이 둑 위에 적절히 위치함.",
      "entities": "전택수의 얼굴, 체형, 모자, 체크 셔츠가 참조와 잘 일치함. 단, 재킷의 텍스트와 패치가 생략됨. 이전 샷의 물속 시신은 지시대로 제거됨.",
      "hard_violations": [],
      "physics": "화면 밖의 발로 강둑 지면을 디디고 서 있는 자세가 물리적으로 자연스러움."
     },
     {
      "label": "A",
      "direction": "전택수의 시선이 화면 우측 하단에 있는 흐르는 강물을 정확히 향하고 있음.",
      "built_space": "참조 이미지와 동일한 형태의 콘크리트 강둑이 구현되었으며, 인물이 둑 위에 적절히 위치함.",
      "entities": "전택수의 기본 인상과 복장은 일치하나, 재킷에 원본에 없는 'SALICS' 및 깨진 글자 패치가 추가됨. 이전 샷의 시신은 지시대로 제거됨.",
      "hard_violations": [],
      "physics": "화면 밖의 발로 강둑 지면을 디디고 서 있는 자세가 물리적으로 자연스러움."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 12
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S5sh2"
  }
 },
 "S20sh1::cine": {
  "applied": true,
  "fingerprint": "2173fd4b9fb9e019f994d6e2e0728098bbb5e24dcbecc4d3c737a5c71a905d99",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S20sh1_sel.png",
  "source_sha256": "1b2edf13fe2354fb467031f9ac7cfe515834669a524b88484ce511be9fd6fa77",
  "file": "S20sh1_cine.png",
  "latency_ms": 11931
 },
 "S20sh5::signage": {
  "fp": "3dd0e1c455cd2ac3",
  "inscriptions": []
 },
 "S20sh5": {
  "input_fingerprint": "b444f4d87eb15c6c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물에 손을 담근 채 씁쓸한 눈빛으로 수면을 응시하는 전택수의 얼굴\n\nLOCATION (lock): At the water's edge of the riverbank, close enough to reach into the cold river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly at water-surface height slightly forward of 전택수's side profile, looking gently upward without entering his frontal axis. His bitter face occupies most of the upper-left area while his lowered eyes lead to the water at lower right; the arm descending out of frame keeps his immersed hand legible at the edge of the composition.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward water surface beside his immersed hand; water surface beside his immersed hand in the lower-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 강물 수면 (flowing around his immersed hand); used as Receives 전택수's hand and remains the focus of his downward stare; 강변 가장자리 (beside his crouched position); used as Provides a restrained edge behind his lowered posture without competing with his face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains subdued and neutral-to-cool, preserving natural skin and water tones without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same riverbank, flowing water, muted daylight, and cold riverside look. Exclude the earlier upright pose and frame the man crouched with his hand in the water.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물에 손을 담근 채 씁쓸한 눈빛으로 수면을 응시하는 전택수의 얼굴\n\nLOCATION (lock): At the water's edge of the riverbank, close enough to reach into the cold river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly at water-surface height slightly forward of 전택수's side profile, looking gently upward without entering his frontal axis. His bitter face occupies most of the upper-left area while his lowered eyes lead to the water at lower right; the arm descending out of frame keeps his immersed hand legible at the edge of the composition.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward water surface beside his immersed hand; water surface beside his immersed hand in the lower-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 강물 수면 (flowing around his immersed hand); used as Receives 전택수's hand and remains the focus of his downward stare; 강변 가장자리 (beside his crouched position); used as Provides a restrained edge behind his lowered posture without competing with his face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains subdued and neutral-to-cool, preserving natural skin and water tones without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same riverbank, flowing water, muted daylight, and cold riverside look. Exclude the earlier upright pose and frame the man crouched with his hand in the water.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 물에 손을 담근 채 씁쓸한 눈빛으로 수면을 응시하는 전택수의 얼굴\n\nLOCATION (lock): At the water's edge of the riverbank, close enough to reach into the cold river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly at water-surface height slightly forward of 전택수's side profile, looking gently upward without entering his frontal axis. His bitter face occupies most of the upper-left area while his lowered eyes lead to the water at lower right; the arm descending out of frame keeps his immersed hand legible at the edge of the composition.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward water surface beside his immersed hand; water surface beside his immersed hand in the lower-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 강물 수면 (flowing around his immersed hand); used as Receives 전택수's hand and remains the focus of his downward stare; 강변 가장자리 (beside his crouched position); used as Provides a restrained edge behind his lowered posture without competing with his face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daylight remains subdued and neutral-to-cool, preserving natural skin and water tones without dramatic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same riverbank, flowing water, muted daylight, and cold riverside look. Exclude the earlier upright pose and frame the man crouched with his hand in the water.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "남성의 시선은 강물에 담긴 자신의 손과 수면을 향해 아래로 내려다보고 있습니다.",
    "built_space": "야외 강변 환경으로, 콘크리트 제방과 수면이 위치하며 별도의 인공 구조물 내부는 없습니다.",
    "entities": "남성은 지정된 연령대와 외모를 갖추고 있으며, 레퍼런스와 동일한 모자, 재킷, 셔츠를 착용하고 있습니다.",
    "hard_violations": [],
    "physics": "남성은 다리로 체중을 지탱하며 웅크려 앉아 있고, 한 손은 수면에 닿아 지지되고 있어 물리적으로 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "남성의 시선은 물속에 잠긴 손 부근의 수면을 향해 정확히 고정되어 있습니다.",
    "built_space": "야외 강변으로 제방 가장자리에 위치하며 공간적 모순이 없습니다.",
    "entities": "지정된 캐릭터 전택수의 얼굴, 체형, 복장(모자, 남색 재킷 등)이 레퍼런스와 정확히 일치합니다.",
    "hard_violations": [],
    "physics": "웅크린 자세로 다리와 허리를 통해 안정적으로 지탱되고 있으며, 손이 물속에 잠겨 있는 상태가 자연스럽게 표현되었습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 4,
   "B": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "프레이밍이 클로즈업이 아닌 미디엄 샷으로 너무 넓게 잡혔으며, 카메라가 수면 높이보다 높아 요청된 로우 앵글 구도를 살리지 못했습니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "요청된 클로즈업 척도에 더 가깝게 프레이밍되었으며, 카메라를 낮춰 인물의 얼굴을 좌측 상단에, 물에 담근 손을 우측 하단에 배치한 텍스트의 지시를 충실히 구현했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S20sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "레퍼런스에서 인물의 좌측 손목에 있던 시계가 생성된 이미지에서는 우측 손목에 채워져 있습니다.",
     "fix_en": "Remove the silver watch from the right wrist, replacing it with bare skin and the edge of the navy jacket sleeve. Preserve the man's pose, face, clothing, the water, and the framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "레퍼런스에 있는 재킷 우측 소매의 방패 모양 패치(로고 및 텍스트)가 누락되었습니다.",
     "fix_en": "Add the dark shield-shaped patch with white text to the upper right sleeve of the navy jacket. Preserve the man's face, pose, the rest of the clothing, the lighting, and the riverbank set.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "수면 높이에서 약간 올려다보며 얼굴이 좌측 상단 대부분을 차지해야 하는 클로즈업 샷 지시와 달리, 카메라가 위에서 내려다보는 구도의 미디엄 샷으로 렌더링되었습니다.",
     "fix_en": "Crop and scale the frame to a close-up focused on the man's face in the upper-left and his right hand in the lower-right, eliminating the visible legs. Preserve the man's features, clothing, lighting, and the water surface.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "물가 가장자리에서 손을 뻗는 설정(reach into the cold river)과 달리, 하반신이 물속에 깊게 잠긴 상태로 웅크리고 있습니다.",
     "fix_en": "Redraw the water around the man's legs as dry khaki fabric resting on the concrete bank, placing him at the water's edge. Preserve the man's upper body pose, face, right hand in the water, lighting, and the river background.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "레퍼런스에서 인물의 좌측 손목에 있던 시계가 생성된 이미지에서는 우측 손목에 채워져 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "레퍼런스에 있는 재킷 우측 소매의 방패 모양 패치(로고 및 텍스트)가 누락되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "수면 높이에서 약간 올려다보며 얼굴이 좌측 상단 대부분을 차지해야 하는 클로즈업 샷 지시와 달리, 카메라가 위에서 내려다보는 구도의 미디엄 샷으로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "물가 가장자리에서 손을 뻗는 설정(reach into the cold river)과 달리, 하반신이 물속에 깊게 잠긴 상태로 웅크리고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 강변 가장자리에서 손을 뻗은 게 아니라 강물 속에 무릎과 재킷 자락까지 잠긴 채 웅크리고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "시선이 오른아래 수면이 아니라 카메라를 향해 정면에 가깝게 향해 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신과 무릎이 모두 보이는 미디엄 샷이며, 얼굴이 상단 왼쪽을 대부분 차지하지 않는다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "레퍼런스에서 인물의 좌측 손목에 있던 시계가 생성된 이미지에서는 우측 손목에 채워져 있습니다.",
    "fix_en": "Remove the silver watch from the right wrist, replacing it with bare skin and the edge of the navy jacket sleeve. Preserve the man's pose, face, clothing, the water, and the framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "레퍼런스에 있는 재킷 우측 소매의 방패 모양 패치(로고 및 텍스트)가 누락되었습니다.",
    "fix_en": "Add the dark shield-shaped patch with white text to the upper right sleeve of the navy jacket. Preserve the man's face, pose, the rest of the clothing, the lighting, and the riverbank set.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "수면 높이에서 약간 올려다보며 얼굴이 좌측 상단 대부분을 차지해야 하는 클로즈업 샷 지시와 달리, 카메라가 위에서 내려다보는 구도의 미디엄 샷으로 렌더링되었습니다.",
    "fix_en": "Crop and scale the frame to a close-up focused on the man's face in the upper-left and his right hand in the lower-right, eliminating the visible legs. Preserve the man's features, clothing, lighting, and the water surface.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "물가 가장자리에서 손을 뻗는 설정(reach into the cold river)과 달리, 하반신이 물속에 깊게 잠긴 상태로 웅크리고 있습니다.",
    "fix_en": "Redraw the water around the man's legs as dry khaki fabric resting on the concrete bank, placing him at the water's edge. Preserve the man's upper body pose, face, right hand in the water, lighting, and the river background.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh1"
  }
 },
 "S20sh5::cine": {
  "applied": true,
  "fingerprint": "e1abc6dc34a6d76d92e1031abb1aeff94a51c529a2734d998c7d5fb5373d3826",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S20sh5_sel.png",
  "source_sha256": "68eb609d91fea5a98a75bdb2e07a47c6184ea1f65bdd02247938090265e4d401",
  "file": "S20sh5_cine.png",
  "latency_ms": 11295
 },
 "S21sh3::signage": {
  "fp": "5be4e0191a87dbeb",
  "inscriptions": [
   {
    "surface_native": "서류철 표지",
    "text_native": "수사기록",
    "reason_ko": "수사과장실 책상 위에 놓인 사건 파일임을 명확히 보여주기 위해 표지에 한국어로 '수사기록'이라는 표기가 필요합니다."
   }
  ]
 },
 "S21sh3": {
  "input_fingerprint": "e527dcf3a0833c89",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상 위 서류들 사이에 놓인 강변 현장 사진과 김선영의 얼굴 사진 위로 뻗은 전택수의 손\n\nLOCATION (lock): At the document-covered desk inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close above the desk at a steep high three-quarter angle, crane into a tight evidence composition just behind 전택수's reaching arm, excluding every live face. His hand crosses the documents and aligns the riverbank recovery photograph beside 김선영's school-uniform portrait, with both photographs fully readable as separate evidence items rather than enlarged graphic imagery.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 전택수's reaching arm and hand in the lower-left of the frame, foreground, reaches for paired evidence photographs; paired evidence photographs in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 강변 현장 사진 (placed among the case documents) — Its printed front faces the camera and shows the riverbank recovery scene at evidentiary distance, without emphasizing graphic detail; used as Primary evidence surface beneath 전택수's approaching hand; 김선영의 사진 (being aligned beside the scene photograph) — Its printed front faces the camera and shows 김선영 wearing a school uniform; used as Pairs the victim's identity with the recovery scene; 사건 서류들 (spread across the desk) — Their printed faces and overlapping edges are visible around the photographs, while specific text remains secondary; used as Creates layered procedural context around the two photographs; 책상 (supporting the documents and photographs) — Its top surface faces the steeply angled camera; used as Defines the boundary of the overhead evidence composition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office illumination is rendered as restrained neutral ambient light with moderate contrast and documentary clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Kim Seon-young’s nude body floats horizontally face-down on the river surface beside the bank, with her head and torso directed downward and her limbs extended in the prone orientation. A flesh-colored stocking hangs from the end of one ankle.\n\nIMMOBILE BODIES OBEY GRAVITY: a dead or unconscious person exerts no muscular effort — every body part rests fully on its support and slumps with gravity; never lifted, raised or posed as presenting. A gripped object stays in a fallen hand, not held up. Unspecified parts resolve to the most gravity-compliant, supported position.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu is handling the riverside crime-scene photograph and Seonyeong's school-uniform portrait, preparing to place them together.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상 위 서류들 사이에 놓인 강변 현장 사진과 김선영의 얼굴 사진 위로 뻗은 전택수의 손\n\nLOCATION (lock): At the document-covered desk inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close above the desk at a steep high three-quarter angle, crane into a tight evidence composition just behind 전택수's reaching arm, excluding every live face. His hand crosses the documents and aligns the riverbank recovery photograph beside 김선영's school-uniform portrait, with both photographs fully readable as separate evidence items rather than enlarged graphic imagery.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 전택수's reaching arm and hand in the lower-left of the frame, foreground, reaches for paired evidence photographs; paired evidence photographs in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 강변 현장 사진 (placed among the case documents) — Its printed front faces the camera and shows the riverbank recovery scene at evidentiary distance, without emphasizing graphic detail; used as Primary evidence surface beneath 전택수's approaching hand; 김선영의 사진 (being aligned beside the scene photograph) — Its printed front faces the camera and shows 김선영 wearing a school uniform; used as Pairs the victim's identity with the recovery scene; 사건 서류들 (spread across the desk) — Their printed faces and overlapping edges are visible around the photographs, while specific text remains secondary; used as Creates layered procedural context around the two photographs; 책상 (supporting the documents and photographs) — Its top surface faces the steeply angled camera; used as Defines the boundary of the overhead evidence composition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office illumination is rendered as restrained neutral ambient light with moderate contrast and documentary clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Kim Seon-young’s nude body floats horizontally face-down on the river surface beside the bank, with her head and torso directed downward and her limbs extended in the prone orientation. A flesh-colored stocking hangs from the end of one ankle.\n\nIMMOBILE BODIES OBEY GRAVITY: a dead or unconscious person exerts no muscular effort — every body part rests fully on its support and slumps with gravity; never lifted, raised or posed as presenting. A gripped object stays in a fallen hand, not held up. Unspecified parts resolve to the most gravity-compliant, supported position.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu is handling the riverside crime-scene photograph and Seonyeong's school-uniform portrait, preparing to place them together.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상 위 서류들 사이에 놓인 강변 현장 사진과 김선영의 얼굴 사진 위로 뻗은 전택수의 손\n\nLOCATION (lock): At the document-covered desk inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close above the desk at a steep high three-quarter angle, crane into a tight evidence composition just behind 전택수's reaching arm, excluding every live face. His hand crosses the documents and aligns the riverbank recovery photograph beside 김선영's school-uniform portrait, with both photographs fully readable as separate evidence items rather than enlarged graphic imagery.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 전택수's reaching arm and hand in the lower-left of the frame, foreground, reaches for paired evidence photographs; paired evidence photographs in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 강변 현장 사진 (placed among the case documents) — Its printed front faces the camera and shows the riverbank recovery scene at evidentiary distance, without emphasizing graphic detail; used as Primary evidence surface beneath 전택수's approaching hand; 김선영의 사진 (being aligned beside the scene photograph) — Its printed front faces the camera and shows 김선영 wearing a school uniform; used as Pairs the victim's identity with the recovery scene; 사건 서류들 (spread across the desk) — Their printed faces and overlapping edges are visible around the photographs, while specific text remains secondary; used as Creates layered procedural context around the two photographs; 책상 (supporting the documents and photographs) — Its top surface faces the steeply angled camera; used as Defines the boundary of the overhead evidence composition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office illumination is rendered as restrained neutral ambient light with moderate contrast and documentary clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nIMMOBILE CHARACTER POSE — CANONICAL (identical wherever this character appears in ANY panel; on any conflict THIS POSE WINS): Kim Seon-young’s nude body floats horizontally face-down on the river surface beside the bank, with her head and torso directed downward and her limbs extended in the prone orientation. A flesh-colored stocking hangs from the end of one ankle.\n\nIMMOBILE BODIES OBEY GRAVITY: a dead or unconscious person exerts no muscular effort — every body part rests fully on its support and slumps with gravity; never lifted, raised or posed as presenting. A gripped object stays in a fallen hand, not held up. Unspecified parts resolve to the most gravity-compliant, supported position.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu is handling the riverside crime-scene photograph and Seonyeong's school-uniform portrait, preparing to place them together.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.4,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.15,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gemini-pro] '수사기록' 텍스트가 회전된 서류철 표지의 각도 및 원근을 따르지 않고 화면에 완벽한 수평으로 삽입되어 있어, '평면 스티커나 그래픽 오버레이처럼 보이면 안 된다(Nothing may look like a flat sticker... pasted onto the frame)'는 재질 및 렌더링 지침을 명백히 위반함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "A": 1150,
   "B": 1667
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1150,
    "verdict_ko": "오른쪽 서류철의 '수사기록' 텍스트가 표지의 기울기를 무시하고 화면에 수평으로 삽입되어 그래픽 오버레이 금지 지침을 위반했으며, 인물 사진의 날짜 디테일도 누락되었습니다.  ★위반: [gemini-pro] '수사기록' 텍스트가 회전된 서류철 표지의 각도 및 원근을 따르지 않고 화면에 완벽한 수평으로 삽입되어 있어, '평면 스티커나 그래픽 오버레이처럼 보이면 안 된다(Nothing may look like a flat sticker... pasted onto the frame)'는 재질 및 렌더링 지침을 명백히 위반함."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "참조 이미지의 사진 디테일(인물 사진 우측 하단의 날짜 포함)과 지정된 텍스트를 완벽하게 재현했으며, 물리적인 표면 질감과 앵글까지 프롬프트의 요구를 훌륭하게 충족했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S17sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "PROP REFERENCE — 김선영 얼굴 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1066360>"
   },
   {
    "label": "PROP REFERENCE — 드들강 사건 현장 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1164634>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "좌측 인물이 레퍼런스에 고정된 모자를 착용하지 않았으며 재킷과 셔츠의 디테일이 다름.",
     "fix_en": "Add the navy blue cap to the head on the left. Keep the hands, photographs, desk, and documents exactly as they are.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "화면 왼쪽에 전택수의 머리와 옆얼굴이 보여 살아있는 얼굴을 모두 제외하라는 프레이밍을 위반한다.",
     "fix_en": "Remove the head and face on the left, filling the space with desk surface and documents. Maintain the hands, photographs, and the rest of the desk setting.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "강변 현장 사진과 얼굴 사진이 책상 서류 사이에 놓여 손 아래에 있지 않고 양손에 들려 있다.",
     "fix_en": "Place the two photographs flat on the desk and reposition the hands to rest near them. Keep the head, desk, documents, and lighting unchanged.",
     "severity": "critical",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "사진들이 책상 위에 놓여 있지 않고 양손으로 공중에 들려 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "좌측 인물이 레퍼런스에 고정된 모자를 착용하지 않았으며 재킷과 셔츠의 디테일이 다름.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 왼쪽에 전택수의 머리와 옆얼굴이 보여 살아있는 얼굴을 모두 제외하라는 프레이밍을 위반한다.",
     "severity": "critical"
    },
    {
     "issue_ko": "강변 현장 사진과 얼굴 사진이 책상 서류 사이에 놓여 손 아래에 있지 않고 양손에 들려 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "한 손이 서류 위로 뻗어 두 사진을 맞추는 것이 아니라 왼손과 오른손이 사진을 각각 쥐고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 뒷머리에 이전 샷에서 쓰던 수사 모자가 없다.",
     "severity": "major"
    },
    {
     "issue_ko": "책상 위 급각 인서트 클로즈업이 아니라 머리까지 포함한 오버숄더 구도이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "좌측 인물이 레퍼런스에 고정된 모자를 착용하지 않았으며 재킷과 셔츠의 디테일이 다름.",
    "fix_en": "Add the navy blue cap to the head on the left. Keep the hands, photographs, desk, and documents exactly as they are.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the head and face on the left, filling the space with desk surface and documents. Maintain the hands, photographs, and the rest of the desk setting.\n- Place the two photographs flat on the desk and reposition the hands to rest near them. Keep the head, desk, documents, and lighting unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "사진을 책상 위에 두지 않고 양손으로 들고 있으며 샷 텍스트의 타이트한 프레이밍 지시보다 넓게 잡혔으나, 치명적인 구조적 오류가 없어 최종 선택되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "화면 분할(콜라주)과 신체 중복이라는 치명적 위반(Hard Violation)이 발생했으며, 얼굴을 배제하라는 카메라 지시도 완전히 무시하여 탈락입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 손에 든 두 장의 사진을 향함.",
      "built_space": "서류와 파일이 여러 겹으로 흩어진 책상 표면 하나가 보임.",
      "entities": "전택수의 뒷머리와 어깨, 강변 현장 사진, 날짜가 적힌 김선영 사진, '수사기록' 텍스트가 적힌 서류철.",
      "hard_violations": [],
      "physics": "양손이 사진의 모서리를 잡고 공중에 들고 있음."
     },
     {
      "label": "B",
      "direction": "왼쪽 인물은 카메라 렌즈를 정면으로 응시하며, 오른쪽 패널의 손은 책상을 향해 놓여 있음.",
      "built_space": "좌측 인물 초상과 우측 스탠드가 켜진 책상 위 공간이 분할된 콜라주 화면.",
      "entities": "전택수의 정면 상반신, 별도로 등장하는 양손, 강변 현장 사진, 김선영 사진.",
      "hard_violations": [
       "collage/panels",
       "duplicated or extra bodies"
      ],
      "physics": "우측 패널의 손과 사진들은 책상 표면에 의해 온전히 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "사진을 책상 위에 두지 않고 양손으로 들고 있으며 샷 텍스트의 타이트한 프레이밍 지시보다 넓게 잡혔으나, 치명적인 구조적 오류가 없어 최종 선택되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "화면 분할(콜라주)과 신체 중복이라는 치명적 위반(Hard Violation)이 발생했으며, 얼굴을 배제하라는 카메라 지시도 완전히 무시하여 탈락입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 손에 든 두 장의 사진을 향함.",
      "built_space": "서류와 파일이 여러 겹으로 흩어진 책상 표면 하나가 보임.",
      "entities": "전택수의 뒷머리와 어깨, 강변 현장 사진, 날짜가 적힌 김선영 사진, '수사기록' 텍스트가 적힌 서류철.",
      "hard_violations": [],
      "physics": "양손이 사진의 모서리를 잡고 공중에 들고 있음."
     },
     {
      "label": "B",
      "direction": "왼쪽 인물은 카메라 렌즈를 정면으로 응시하며, 오른쪽 패널의 손은 책상을 향해 놓여 있음.",
      "built_space": "좌측 인물 초상과 우측 스탠드가 켜진 책상 위 공간이 분할된 콜라주 화면.",
      "entities": "전택수의 정면 상반신, 별도로 등장하는 양손, 강변 현장 사진, 김선영 사진.",
      "hard_violations": [
       "collage/panels",
       "duplicated or extra bodies"
      ],
      "physics": "우측 패널의 손과 사진들은 책상 표면에 의해 온전히 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "화면이 분할된 콜라주 형태이며, 살아있는 얼굴을 배제하라는 카메라 프레이밍 지시를 정면으로 위반하여 인물의 얼굴을 노출시켰습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "얼굴을 배제하고 팔 뒤에서 내려다보는 지정된 하이 앵글 인서트 샷을 정확히 구현했으며, 증거 사진들과 '수사기록' 텍스트 배치가 훌륭합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽 인물은 카메라 렌즈를 정면으로 응시하고 있으며, 오른쪽 화면의 손은 책상 위의 사진 가장자리에 놓여 있음.",
      "built_space": "화면이 좌우로 분할된 콜라주 구도이며, 오른쪽 절반에 책상 표면과 서류들이 배치됨.",
      "entities": "전택수의 얼굴이 노출됨(지시 위반). 강변 현장 사진과 김선영의 교복 사진은 참조 이미지와 일치함. '수사기록' 텍스트가 보이지 않음.",
      "hard_violations": [
       "콜라주/패널(분할 화면) 사용",
       "살아있는 얼굴 배제 지시 위반(전택수의 정면 얼굴 노출)"
      ],
      "physics": "손은 책상과 사진 위에 자연스럽게 얹혀 있으나 인물의 시점과 손의 위치가 물리적으로 연결되지 않는 구도임."
     },
     {
      "label": "B",
      "direction": "카메라는 인물의 어깨 뒤에서 책상을 내려다보고 있으며, 양손은 두 장의 증거 사진을 마주 보게 쥐고 있음.",
      "built_space": "인물의 팔과 어깨 너머로 책상 위 서류와 텍스트가 배치된 단일 공간.",
      "entities": "지시대로 라이브 페이스가 완전히 배제됨(뒷모습만 등장). 강변 현장 사진과 김선영 사진이 지시된 대로 명확히 보임. 서류철 표지에 '수사기록' 텍스트가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "양손이 각각 사진의 가장자리를 잡고 서류들 위에서 지탱하고 있으며 중력에 맞게 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "화면이 분할된 콜라주 형태이며, 살아있는 얼굴을 배제하라는 카메라 프레이밍 지시를 정면으로 위반하여 인물의 얼굴을 노출시켰습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "얼굴을 배제하고 팔 뒤에서 내려다보는 지정된 하이 앵글 인서트 샷을 정확히 구현했으며, 증거 사진들과 '수사기록' 텍스트 배치가 훌륭합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "왼쪽 인물은 카메라 렌즈를 정면으로 응시하고 있으며, 오른쪽 화면의 손은 책상 위의 사진 가장자리에 놓여 있음.",
      "built_space": "화면이 좌우로 분할된 콜라주 구도이며, 오른쪽 절반에 책상 표면과 서류들이 배치됨.",
      "entities": "전택수의 얼굴이 노출됨(지시 위반). 강변 현장 사진과 김선영의 교복 사진은 참조 이미지와 일치함. '수사기록' 텍스트가 보이지 않음.",
      "hard_violations": [
       "콜라주/패널(분할 화면) 사용",
       "살아있는 얼굴 배제 지시 위반(전택수의 정면 얼굴 노출)"
      ],
      "physics": "손은 책상과 사진 위에 자연스럽게 얹혀 있으나 인물의 시점과 손의 위치가 물리적으로 연결되지 않는 구도임."
     },
     {
      "label": "A",
      "direction": "카메라는 인물의 어깨 뒤에서 책상을 내려다보고 있으며, 양손은 두 장의 증거 사진을 마주 보게 쥐고 있음.",
      "built_space": "인물의 팔과 어깨 너머로 책상 위 서류와 텍스트가 배치된 단일 공간.",
      "entities": "지시대로 라이브 페이스가 완전히 배제됨(뒷모습만 등장). 강변 현장 사진과 김선영 사진이 지시된 대로 명확히 보임. 서류철 표지에 '수사기록' 텍스트가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "양손이 각각 사진의 가장자리를 잡고 서류들 위에서 지탱하고 있으며 중력에 맞게 자연스러움."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S17sh2"
  }
 },
 "S21sh3::cine": {
  "applied": true,
  "fingerprint": "b78f1cb6d723172e3476b0949d415f7be1b09eed411b5517d5d8bd1d1575da74",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S21sh3_sel.png",
  "source_sha256": "3ee3d05cfb7aa947f5c31d4dd7579eb6aff11fde026df4c609dcd14b3428c01a",
  "file": "S21sh3_cine.png",
  "latency_ms": 12607
 },
 "S21sh5::signage": {
  "fp": "d2b4db9ae807049f",
  "inscriptions": [
   {
    "surface_native": "수사과장 명패",
    "text_native": "수사과장",
    "reason_ko": "수사과장실 내부라는 공간적 배경을 사실적으로 나타내기 위해 사무실 내에 위치한 직책 명패가 필요합니다."
   }
  ]
 },
 "S21sh5": {
  "input_fingerprint": "885bbeb469e1ff17",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용과 강주철 쪽을 향해 고개를 든 채 단호한 표정으로 입술을 뗀 전택수\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the push near the desk edge slightly below 전택수's seated eye line, holding him in a tight three-quarter profile as he lifts his head from the paired photographs and parts his lips. Keep him off-center toward 서의용 and 강주철 in his eyeline, using their soft shoulder edges in the opposite foreground to make his firm address spatially clear.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 나란히 놓인 두 사진 (aligned on the desk) — The printed fronts remain partially visible below 전택수, one showing the riverbank recovery scene and the other 김선영 in school uniform; used as Keeps the evidentiary basis of 전택수's decision at the lower frame edge; 책상 가장자리 (between the seated participants) — Its near edge runs across the lower portion of the frame; used as Separates 전택수 from the two men he is addressing and supports the low camera position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office light remains neutral and naturalistic, with restrained contrast concentrating attention on 전택수's expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nude riverside crime-scene photograph and Seonyeong's school-uniform portrait remain placed side by side on the desk.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사과장 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용과 강주철 쪽을 향해 고개를 든 채 단호한 표정으로 입술을 뗀 전택수\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the push near the desk edge slightly below 전택수's seated eye line, holding him in a tight three-quarter profile as he lifts his head from the paired photographs and parts his lips. Keep him off-center toward 서의용 and 강주철 in his eyeline, using their soft shoulder edges in the opposite foreground to make his firm address spatially clear.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 나란히 놓인 두 사진 (aligned on the desk) — The printed fronts remain partially visible below 전택수, one showing the riverbank recovery scene and the other 김선영 in school uniform; used as Keeps the evidentiary basis of 전택수's decision at the lower frame edge; 책상 가장자리 (between the seated participants) — Its near edge runs across the lower portion of the frame; used as Separates 전택수 from the two men he is addressing and supports the low camera position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office light remains neutral and naturalistic, with restrained contrast concentrating attention on 전택수's expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nude riverside crime-scene photograph and Seonyeong's school-uniform portrait remain placed side by side on the desk.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사과장 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용과 강주철 쪽을 향해 고개를 든 채 단호한 표정으로 입술을 뗀 전택수\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the push near the desk edge slightly below 전택수's seated eye line, holding him in a tight three-quarter profile as he lifts his head from the paired photographs and parts his lips. Keep him off-center toward 서의용 and 강주철 in his eyeline, using their soft shoulder edges in the opposite foreground to make his firm address spatially clear.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 나란히 놓인 두 사진 (aligned on the desk) — The printed fronts remain partially visible below 전택수, one showing the riverbank recovery scene and the other 김선영 in school uniform; used as Keeps the evidentiary basis of 전택수's decision at the lower frame edge; 책상 가장자리 (between the seated participants) — Its near edge runs across the lower portion of the frame; used as Separates 전택수 from the two men he is addressing and supports the low camera position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office light remains neutral and naturalistic, with restrained contrast concentrating attention on 전택수's expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The nude riverside crime-scene photograph and Seonyeong's school-uniform portrait remain placed side by side on the desk.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사과장 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수가 좌측 전경의 두 남성에게 시선을 고정하고 입을 벌리고 있음.",
    "built_space": "책상을 사이에 두고 마주 앉은 구조. 책상 위에는 명패와 두 장의 사진이 놓여 있음.",
    "entities": "전택수의 외모와 복장이 참조와 일치함. '수사과장' 명패와 지정된 두 사진(현장, 여학생)이 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "인물들은 의자에 안정적으로 앉아 있고, 사진은 책상 위에 평평하게 놓여 있음(다만 방향이 카메라를 향함)."
   },
   {
    "label": "B",
    "direction": "전택수가 화면 좌우 전경에 배치된 두 남성 쪽을 바라보며 입을 엶.",
    "built_space": "책상을 사이에 둔 배치이나, 전경의 인물들이 좌우 양쪽에 있어 전택수가 화면 중앙에 위치함.",
    "entities": "전택수의 인상 및 복장, '수사과장' 명패, 두 장의 사진이 지정된 대로 등장함.",
    "hard_violations": [
     "우측 여학생 사진을 지지하는 요소가 전혀 없이 얇은 모서리로 책상에 수직으로 서 있음 (nothing supports it)."
    ],
    "physics": "우측 사진이 아무런 물리적 지지 없이 책상 위에 서 있어 불가능한 상태임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 오프센터 프레이밍과 3/4 측면 각도를 잘 구현했으며 물리적 오류가 없으나, 책상 위 사진들이 사용자가 아닌 카메라를 향해 놓인 점이 아쉬움."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "우측 사진이 지지대 없이 세워져 있는 물리적 오류(Hard violation)가 있으며, 인물이 중앙에 배치되어 오프센터 프레이밍 지시를 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S21sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 컷(PREVIOUS SHOT STILL)에서는 전택수가 모자를 쓰지 않은 상태(맨머리)이나, 본 컷에서는 모자를 쓰고 있어 복장 및 헤어스타일 연속성 유지(LOCKED) 지침을 위반함.",
     "fix_en": "Remove the cap from Jeon Taek-su's head and replace it with his short, neatly trimmed black hair as seen in the previous shot. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "책상 위에 놓인 두 장의 사진 속 내용(여학생의 얼굴, 강변 현장의 요원 배치 및 통제선 구도)이 이전 컷의 사진과 다르게 변형됨.",
     "fix_en": "Redraw the two photographs on the desk to match their exact contents from the previous shot still—the left photo showing three hazmat-suited figures on a riverbank with yellow tape, and the right showing the schoolgirl in uniform with a red and green plaid bow. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "이전 컷에서 전택수 앞 책상에 놓여 있던 '수사기록' 서류철 및 기타 문서들이 모두 누락됨.",
     "fix_en": "Add the missing '수사기록' file folder and other document piles to the desk in front of Jeon Taek-su.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "클로즈업·타이트한 3/4 프로필이 아니라 책상과 상대가 다 보이는 미디엄 샷이다",
     "fix_en": "Move the camera closer for a tighter close-up framing on Jeon Taek-su.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이전 스틸의 서류 더미가 사라지고 책상이 거의 비어 배경 소파·화분이 다르다",
     "fix_en": "Restore the document piles on the desk and match the background sofa and plant exactly to the previous still.",
     "severity": "major",
     "observation_index": 6
    },
    {
     "issue_ko": "두 사진이 하단 가장자리에 부분적으로 비치지 않고 책상 위에 온전히 드러나 있다",
     "fix_en": "Reposition the two photographs lower on the desk so they are partially cut off by the lower frame edge.",
     "severity": "major",
     "observation_index": 7
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이전 컷(PREVIOUS SHOT STILL)에서는 전택수가 모자를 쓰지 않은 상태(맨머리)이나, 본 컷에서는 모자를 쓰고 있어 복장 및 헤어스타일 연속성 유지(LOCKED) 지침을 위반함.",
     "severity": "critical"
    },
    {
     "issue_ko": "책상 위에 놓인 두 장의 사진 속 내용(여학생의 얼굴, 강변 현장의 요원 배치 및 통제선 구도)이 이전 컷의 사진과 다르게 변형됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 컷에서 전택수 앞 책상에 놓여 있던 '수사기록' 서류철 및 기타 문서들이 모두 누락됨.",
     "severity": "major"
    },
    {
     "issue_ko": "샷 텍스트에 없는 두 남성의 머리와 어깨가 왼쪽 전경에 크게 들어와 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "클로즈업·타이트한 3/4 프로필이 아니라 책상과 상대가 다 보이는 미디엄 샷이다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 이전 스틸에 없던 수사 모자와 경찰 패치 재킷을 입고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "이전 스틸의 서류 더미가 사라지고 책상이 거의 비어 배경 소파·화분이 다르다",
     "severity": "major"
    },
    {
     "issue_ko": "두 사진이 하단 가장자리에 부분적으로 비치지 않고 책상 위에 온전히 드러나 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 컷에서 전택수 앞 책상에 놓여 있던 '수사기록' 서류철 및 기타 문서들이 모두 누락됨.",
    "fix_en": "Add the missing '수사기록' file folder and other document piles to the desk in front of Jeon Taek-su.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "클로즈업·타이트한 3/4 프로필이 아니라 책상과 상대가 다 보이는 미디엄 샷이다",
    "fix_en": "Move the camera closer for a tighter close-up framing on Jeon Taek-su.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "이전 스틸의 서류 더미가 사라지고 책상이 거의 비어 배경 소파·화분이 다르다",
    "fix_en": "Restore the document piles on the desk and match the background sofa and plant exactly to the previous still.",
    "severity": "major",
    "observation_index": 6
   },
   {
    "issue_ko": "두 사진이 하단 가장자리에 부분적으로 비치지 않고 책상 위에 온전히 드러나 있다",
    "fix_en": "Reposition the two photographs lower on the desk so they are partially cut off by the lower frame edge.",
    "severity": "major",
    "observation_index": 7
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the cap from Jeon Taek-su's head and replace it with his short, neatly trimmed black hair as seen in the previous shot. Preserve the people present and their positions, their clothing, the set, the light, and the framing.\n- Redraw the two photographs on the desk to match their exact contents from the previous shot still—the left photo showing three hazmat-suited figures on a riverbank with yellow tape, and the right showing the schoolgirl in uniform with a red and green plaid bow. Preserve the people present and their positions, their clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "이전 샷의 헤어스타일(모자 미착용)을 정확히 유지하여 연속성 조건을 완벽히 충족했으며, 요구된 카메라 구도와 소품 배치도 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "이전 샷 스틸과 동일한 의상 및 헤어스타일을 유지해야 함에도 캐릭터 레퍼런스의 모자를 착용하여 치명적인 연속성 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 좌측 전경에 있는 두 남성을 향해 있음.",
      "built_space": "사무실 내부 책상 앞. 책상 모서리가 화면 하단을 가로지르며 인물 사이의 공간을 분리함.",
      "entities": "전택수(모자 착용, 연속성 위반), 앞모습이 보이지 않는 두 남성의 어깨, 현장 및 여학생 사진 2장, '수사과장' 명패가 존재함.",
      "hard_violations": [],
      "physics": "인물들은 의자에 앉아 무게중심을 지지하고 있으며, 책상 위의 사진과 명패는 표면에 안정적으로 놓여 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 화면 좌측 전경에 있는 두 남성을 향해 있음.",
      "built_space": "사무실 내부 책상 앞. 책상 모서리가 화면 하단을 가로지르며 인물 사이의 공간을 분리함.",
      "entities": "전택수(모자 미착용, 이전 샷과 일치하는 짧은 머리), 앞모습이 보이지 않는 두 남성의 어깨, 현장 및 여학생 사진 2장, '수사과장' 명패가 존재함.",
      "hard_violations": [],
      "physics": "인물들은 의자에 앉아 무게중심을 지지하고 있으며, 책상 위의 사진과 명패는 표면에 안정적으로 놓여 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "이전 샷의 헤어스타일(모자 미착용)을 정확히 유지하여 연속성 조건을 완벽히 충족했으며, 요구된 카메라 구도와 소품 배치도 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "이전 샷 스틸과 동일한 의상 및 헤어스타일을 유지해야 함에도 캐릭터 레퍼런스의 모자를 착용하여 치명적인 연속성 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 좌측 전경에 있는 두 남성을 향해 있음.",
      "built_space": "사무실 내부 책상 앞. 책상 모서리가 화면 하단을 가로지르며 인물 사이의 공간을 분리함.",
      "entities": "전택수(모자 착용, 연속성 위반), 앞모습이 보이지 않는 두 남성의 어깨, 현장 및 여학생 사진 2장, '수사과장' 명패가 존재함.",
      "hard_violations": [],
      "physics": "인물들은 의자에 앉아 무게중심을 지지하고 있으며, 책상 위의 사진과 명패는 표면에 안정적으로 놓여 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 화면 좌측 전경에 있는 두 남성을 향해 있음.",
      "built_space": "사무실 내부 책상 앞. 책상 모서리가 화면 하단을 가로지르며 인물 사이의 공간을 분리함.",
      "entities": "전택수(모자 미착용, 이전 샷과 일치하는 짧은 머리), 앞모습이 보이지 않는 두 남성의 어깨, 현장 및 여학생 사진 2장, '수사과장' 명패가 존재함.",
      "hard_violations": [],
      "physics": "인물들은 의자에 앉아 무게중심을 지지하고 있으며, 책상 위의 사진과 명패는 표면에 안정적으로 놓여 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "이전 샷과의 연속성을 유지하라는 지시에 따라 모자를 쓰지 않은 모습을 정확히 반영했으며, 인물 배치와 시선 처리도 훌륭합니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "구도와 텍스트는 우수하나, 이전 샷의 연속성 지시를 무시하고 캐릭터 레퍼런스의 모자를 씌운 점이 결정적인 감점 요인입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 앞쪽 전경에 위치한 두 남자의 어깨 너머로 그들을 정확히 향하고 있으며, 입술을 떼고 말하려는 듯한 단호한 표정을 짓고 있음.",
      "built_space": "사무실 책상을 사이에 두고 전택수와 두 남자가 마주앉은 구조이며, 카메라는 전택수의 눈높이보다 살짝 아래에서 타이트하게 인물을 포착하고 있음.",
      "entities": "전택수의 얼굴과 복장은 레퍼런스와 일치하며, 이전 샷의 연속성에 맞춰 모자를 쓰지 않은 머리를 묘사함. 책상 위 강가 현장 사진과 교복 입은 소녀 사진, 띄어쓰기가 약간 들어간 '수사 과장' 명패가 존재함.",
      "hard_violations": [],
      "physics": "사진과 명패는 책상 표면에 올바르게 놓여 있으며, 인물의 자세와 손 위치 등 물리적 지지 상태가 자연스러움."
     },
     {
      "label": "B",
      "direction": "전택수가 고개를 들고 전경의 두 남자를 응시하고 있으며, 입이 살짝 벌어진 채로 시선이 목표를 향함.",
      "built_space": "책상과 인물의 배치, 카메라 앵글 및 포커스가 프롬프트가 요구한 대화 상황의 공간적 구도에 잘 부합함.",
      "entities": "얼굴, 복장, 책상 위 두 장의 사진과 '수사과장' 명패(띄어쓰기 없음)는 정확하나, 이전 샷에서 모자가 없었던 전택수에게 레퍼런스의 모자를 씌워 연속성 규칙을 위반함.",
      "hard_violations": [],
      "physics": "사물들이 책상 위에 안정적으로 지지되어 있고, 인물의 착석 자세에 물리적 오류가 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 샷과의 연속성을 유지하라는 지시에 따라 모자를 쓰지 않은 모습을 정확히 반영했으며, 인물 배치와 시선 처리도 훌륭합니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "구도와 텍스트는 우수하나, 이전 샷의 연속성 지시를 무시하고 캐릭터 레퍼런스의 모자를 씌운 점이 결정적인 감점 요인입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수의 시선은 화면 앞쪽 전경에 위치한 두 남자의 어깨 너머로 그들을 정확히 향하고 있으며, 입술을 떼고 말하려는 듯한 단호한 표정을 짓고 있음.",
      "built_space": "사무실 책상을 사이에 두고 전택수와 두 남자가 마주앉은 구조이며, 카메라는 전택수의 눈높이보다 살짝 아래에서 타이트하게 인물을 포착하고 있음.",
      "entities": "전택수의 얼굴과 복장은 레퍼런스와 일치하며, 이전 샷의 연속성에 맞춰 모자를 쓰지 않은 머리를 묘사함. 책상 위 강가 현장 사진과 교복 입은 소녀 사진, 띄어쓰기가 약간 들어간 '수사 과장' 명패가 존재함.",
      "hard_violations": [],
      "physics": "사진과 명패는 책상 표면에 올바르게 놓여 있으며, 인물의 자세와 손 위치 등 물리적 지지 상태가 자연스러움."
     },
     {
      "label": "A",
      "direction": "전택수가 고개를 들고 전경의 두 남자를 응시하고 있으며, 입이 살짝 벌어진 채로 시선이 목표를 향함.",
      "built_space": "책상과 인물의 배치, 카메라 앵글 및 포커스가 프롬프트가 요구한 대화 상황의 공간적 구도에 잘 부합함.",
      "entities": "얼굴, 복장, 책상 위 두 장의 사진과 '수사과장' 명패(띄어쓰기 없음)는 정확하나, 이전 샷에서 모자가 없었던 전택수에게 레퍼런스의 모자를 씌워 연속성 규칙을 위반함.",
      "hard_violations": [],
      "physics": "사물들이 책상 위에 안정적으로 지지되어 있고, 인물의 착석 자세에 물리적 오류가 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh3"
  }
 },
 "S21sh5::cine": {
  "applied": true,
  "fingerprint": "14b9e9a4d68bea95886c2b92155e4ce433d03b88afe461c44c2c3045d12c15d9",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S21sh5_sel.png",
  "source_sha256": "3c69385902a8be46f552e7c95ca0feb06f73d9893e2d362186a14355257a8d1d",
  "file": "S21sh5_cine.png",
  "latency_ms": 12109
 },
 "S22sh7::signage": {
  "fp": "6d54b4a9770f4b13",
  "inscriptions": [
   {
    "surface_native": "사무실 벽면의 부서 안내판",
    "text_native": "강력1반",
    "reason_ko": "한국 경찰서 강력반 사무실 내부라는 공간적 배경을 명확하고 사실적으로 나타내기 위해 벽면에 부서 표지판이 필요합니다."
   }
  ]
 },
 "S22sh7": {
  "input_fingerprint": "908014dffcb1b0d5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 얼떨떨한 표정으로 나상혁의 손을 맞잡은 서의용의 어색한 얼굴\n\nLOCATION (lock): Beside the detective's workstation in the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Continue the dolly close behind 나상혁's extended arm and just outside his shoulder, framing 서의용's bewildered face at upper center with their clasped hands retained along the lower edge. 서의용 leans into the involuntary handshake while staring awkwardly at 나상혁, whose shoulder and partial profile provide a restrained foreground frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 나상혁 in the middle-left of the frame, foreground, reaches for 서의용's clasped hand; 서의용 in the middle-center of the frame, midground, reaches for 나상혁's clasped hand.\n- KEY BACKGROUND ELEMENTS: 서의용의 책상 (recently used by 나상혁) — Its working side is behind 나상혁 and its outer edge recedes into the frame; used as Keeps the displaced-workspace context visible behind the uneasy introduction; 컴퓨터 화면 (present on the desk) — The display face is only partially visible at an oblique angle; its specific contents are not emphasized; used as Provides subdued office context without drawing focus from the handshake.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office illumination is kept neutral-to-cool and naturalistic, with moderate contrast and no heightened dramatic source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면의 부서 안내판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 얼떨떨한 표정으로 나상혁의 손을 맞잡은 서의용의 어색한 얼굴\n\nLOCATION (lock): Beside the detective's workstation in the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Continue the dolly close behind 나상혁's extended arm and just outside his shoulder, framing 서의용's bewildered face at upper center with their clasped hands retained along the lower edge. 서의용 leans into the involuntary handshake while staring awkwardly at 나상혁, whose shoulder and partial profile provide a restrained foreground frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 나상혁 in the middle-left of the frame, foreground, reaches for 서의용's clasped hand; 서의용 in the middle-center of the frame, midground, reaches for 나상혁's clasped hand.\n- KEY BACKGROUND ELEMENTS: 서의용의 책상 (recently used by 나상혁) — Its working side is behind 나상혁 and its outer edge recedes into the frame; used as Keeps the displaced-workspace context visible behind the uneasy introduction; 컴퓨터 화면 (present on the desk) — The display face is only partially visible at an oblique angle; its specific contents are not emphasized; used as Provides subdued office context without drawing focus from the handshake.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office illumination is kept neutral-to-cool and naturalistic, with moderate contrast and no heightened dramatic source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면의 부서 안내판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 얼떨떨한 표정으로 나상혁의 손을 맞잡은 서의용의 어색한 얼굴\n\nLOCATION (lock): Beside the detective's workstation in the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Continue the dolly close behind 나상혁's extended arm and just outside his shoulder, framing 서의용's bewildered face at upper center with their clasped hands retained along the lower edge. 서의용 leans into the involuntary handshake while staring awkwardly at 나상혁, whose shoulder and partial profile provide a restrained foreground frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 나상혁 in the middle-left of the frame, foreground, reaches for 서의용's clasped hand; 서의용 in the middle-center of the frame, midground, reaches for 나상혁's clasped hand.\n- KEY BACKGROUND ELEMENTS: 서의용의 책상 (recently used by 나상혁) — Its working side is behind 나상혁 and its outer edge recedes into the frame; used as Keeps the displaced-workspace context visible behind the uneasy introduction; 컴퓨터 화면 (present on the desk) — The display face is only partially visible at an oblique angle; its specific contents are not emphasized; used as Provides subdued office context without drawing focus from the handshake.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office illumination is kept neutral-to-cool and naturalistic, with moderate contrast and no heightened dramatic source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면의 부서 안내판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "중경의 나상혁이 전경의 서의용을 바라보며 손을 맞잡고 있음.",
    "built_space": "사무실 내부의 책상, 모니터, 캐비닛 및 '강력1반' 안내판이 보임.",
    "entities": "전경 인물은 서의용(비니, 가죽재킷), 중경 인물은 나상혁(네이비 재킷)으로 외형은 일치함.",
    "hard_violations": [
     "프롬프트의 카메라 뷰와 인물 배치가 완전히 반전됨 (서의용이 중경, 나상혁이 전경이어야 하나 반대로 렌더링됨)"
    ],
    "physics": "두 인물이 서서 손을 잡고 있으며 물리적 지지는 정상적임."
   },
   {
    "label": "B",
    "direction": "중경의 나상혁이 전경의 서의용을 응시하며 악수하고 있음.",
    "built_space": "사무실 책상 위 모니터와 배경의 캐비닛, '강력1반' 안내판이 배치됨.",
    "entities": "전경은 서의용, 중경은 나상혁의 레퍼런스 외형과 일치함.",
    "hard_violations": [
     "프롬프트의 카메라 뷰와 인물 배치가 완전히 반전됨 (서의용이 중경, 나상혁이 전경이어야 하나 반대로 렌더링됨)"
    ],
    "physics": "바닥을 딛고 서서 악수하는 자세에 물리적 오류는 없음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 3,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "프롬프트가 요구한 두 인물의 위치(전경과 중경)를 완전히 반대로 배치하여 샷의 의도를 상실했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "두 인물의 전경 및 중경 배치가 프롬프트와 정반대로 렌더링되어 심각한 구도 오류가 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S13sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라 프레이밍 지시(나상혁의 어깨 너머로 서의용의 얼굴을 잡을 것)와 달리, 두 인물의 위치가 뒤바뀌어 서의용(비니와 가죽 재킷)이 전경에 배치되고 나상혁(네이비 재킷)의 얼굴이 화면 중앙에 나타났습니다.",
     "fix_en": "Redraw the characters to swap their identities: make the man facing the camera Seo Ui-yong (wearing a navy beanie and weathered black leather jacket), and make the foreground man Na Sang-hyuk (wearing a navy zip-up jacket and light shirt), preserving the handshake pose, the desk, the computer, and the office setting.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "클로즈업이 아니라 상반신과 책상·캐비닛이 넓게 보이는 미디엄 샷이다.",
     "fix_en": "Zoom in on the characters' faces and their clasped hands to create a close-up shot, cropping out the lower torso and the wider view of the cabinets and desk, while preserving the existing characters, expressions, and lighting.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "컴퓨터 화면이 사선으로 일부만 보여야 하는데 정면에 가깝게 내용이 드러나 있다.",
     "fix_en": "Redraw the computer monitor on the left to be turned at an oblique angle so its screen is barely visible, preserving the characters, the handshake, and the rest of the desk objects.",
     "severity": "minor",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라 프레이밍 지시(나상혁의 어깨 너머로 서의용의 얼굴을 잡을 것)와 달리, 두 인물의 위치가 뒤바뀌어 서의용(비니와 가죽 재킷)이 전경에 배치되고 나상혁(네이비 재킷)의 얼굴이 화면 중앙에 나타났습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "상단 중앙에 보여야 할 서의용의 얼굴 대신 나상혁의 얼굴이 화면 중앙·우측에 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 전경이 나상혁의 어깨·팔이 아니라 비니와 가죽재킷을 입은 서의용의 등이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "클로즈업이 아니라 상반신과 책상·캐비닛이 넓게 보이는 미디엄 샷이다.",
     "severity": "major"
    },
    {
     "issue_ko": "컴퓨터 화면이 사선으로 일부만 보여야 하는데 정면에 가깝게 내용이 드러나 있다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "클로즈업이 아니라 상반신과 책상·캐비닛이 넓게 보이는 미디엄 샷이다.",
    "fix_en": "Zoom in on the characters' faces and their clasped hands to create a close-up shot, cropping out the lower torso and the wider view of the cabinets and desk, while preserving the existing characters, expressions, and lighting.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "컴퓨터 화면이 사선으로 일부만 보여야 하는데 정면에 가깝게 내용이 드러나 있다.",
    "fix_en": "Redraw the computer monitor on the left to be turned at an oblique angle so its screen is barely visible, preserving the characters, the handshake, and the rest of the desk objects.",
    "severity": "minor",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the characters to swap their identities: make the man facing the camera Seo Ui-yong (wearing a navy beanie and weathered black leather jacket), and make the foreground man Na Sang-hyuk (wearing a navy zip-up jacket and light shirt), preserving the handshake pose, the desk, the computer, and the office setting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1375,
      "verdict_ko": "요구된 클로즈업 프레이밍 대신 미디엄 샷으로 렌더링되어 감점이 있으나, 인물의 좌우 배치와 중심 인물의 얼굴을 올바르게 노출하여 A보다 우세함."
     },
     {
      "label": "A",
      "score": 1600,
      "verdict_ko": "오버더숄더 클로즈업 구도를 시도했으나 전경과 중경의 인물 배치가 완전히 뒤바뀌어, 정작 보여주어야 할 서의용의 표정 대신 나상혁의 얼굴이 강조되는 치명적인 연출 오류가 발생함."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.6,
      "B": 1.375
     },
     "adjusted": {
      "A": 1.6,
      "B": 1.375
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.625,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1375,
      "verdict_ko": "요구된 클로즈업 프레이밍 대신 미디엄 샷으로 렌더링되어 감점이 있으나, 인물의 좌우 배치와 중심 인물의 얼굴을 올바르게 노출하여 A보다 우세함."
     },
     {
      "label": "A",
      "score": 1600,
      "verdict_ko": "오버더숄더 클로즈업 구도를 시도했으나 전경과 중경의 인물 배치가 완전히 뒤바뀌어, 정작 보여주어야 할 서의용의 표정 대신 나상혁의 얼굴이 강조되는 치명적인 연출 오류가 발생함."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "캐릭터의 외형과 배경 요소는 잘 반영했으나, 지정된 오버더숄더 클로즈업 구도를 무시하고 측면 투샷을 생성했으며 나상혁의 왼팔에 오른손이 달리는 치명적인 해부학적 오류가 발생했습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "지정된 전경과 원경의 인물 위치를 완전히 반대로 배치하여 샷의 핵심 대상인 서의용의 얼굴을 가려버렸으며, 악수하는 손의 구조가 심하게 뭉개졌습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "나상혁과 서의용이 서로를 마주보며 시선을 교환하고 있습니다. 카메라는 지시된 나상혁의 어깨 너머 시점이 아닌, 측면에서 두 사람을 모두 담는 넓은 구도를 취하고 있습니다.",
      "built_space": "전경에 전화기가 놓인 책상이 있고, 배경에 복사기와 철제 캐비닛이 보입니다. 왼쪽 벽면에 '강력1반' 안내판이 위치해 이전 샷의 공간 배치를 잘 유지하고 있습니다.",
      "entities": "서의용(비니, 가죽 재킷)과 나상혁(네이비 재킷, 단정한 머리)의 인상착의가 모두 참조 이미지 및 프롬프트와 일치합니다.",
      "hard_violations": [
       "물리적으로 불가능한 해부학 구조 (나상혁의 왼쪽 팔 끝에 엄지손가락 방향이 뒤집힌 오른손이 달려 있음)"
      ],
      "physics": "나상혁은 책상에 걸터앉아 체중을 지탱하고 있으며 서의용은 두 발로 서 있습니다. 뻗은 팔과 맞잡은 손은 공중에 유지되고 있으나 해부학적으로 성립하지 않습니다."
     },
     {
      "label": "B",
      "direction": "서의용이 나상혁을 바라보고 나상혁이 서의용을 마주봅니다. 카메라는 지시와 정반대로 서의용의 어깨 너머로 나상혁을 향하고 있어, 샷의 핵심인 서의용의 표정을 전혀 담지 못하고 있습니다.",
      "built_space": "전경에 책상 모서리와 켜진 모니터가 보이고, 배경에 복사기와 캐비닛, '강력1반' 안내판이 위치하여 공간의 구성 요소는 일치합니다.",
      "entities": "서의용(비니, 가죽 재킷)과 나상혁(네이비 재킷)의 인상착의가 참조 이미지와 일치합니다.",
      "hard_violations": [
       "물리적으로 불가능한 스테이징 (나상혁과 서의용의 전경/원경 배치를 완전히 뒤바꿔 프롬프트가 요구한 대상의 얼굴을 화면에서 배제함)",
       "물리적으로 불가능한 해부학 구조 (맞잡은 두 손의 엄지손가락 위치와 결합 형태가 논리적으로 불가능하게 뭉개짐)"
      ],
      "physics": "두 사람 모두 서서 바닥에 체중을 지탱하고 있습니다. 손을 맞잡은 동작 자체는 지탱되고 있으나 결합 부위의 해부학적 묘사가 불가능한 형태입니다."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "캐릭터의 외형과 배경 요소는 잘 반영했으나, 지정된 오버더숄더 클로즈업 구도를 무시하고 측면 투샷을 생성했으며 나상혁의 왼팔에 오른손이 달리는 치명적인 해부학적 오류가 발생했습니다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "지정된 전경과 원경의 인물 위치를 완전히 반대로 배치하여 샷의 핵심 대상인 서의용의 얼굴을 가려버렸으며, 악수하는 손의 구조가 심하게 뭉개졌습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "나상혁과 서의용이 서로를 마주보며 시선을 교환하고 있습니다. 카메라는 지시된 나상혁의 어깨 너머 시점이 아닌, 측면에서 두 사람을 모두 담는 넓은 구도를 취하고 있습니다.",
      "built_space": "전경에 전화기가 놓인 책상이 있고, 배경에 복사기와 철제 캐비닛이 보입니다. 왼쪽 벽면에 '강력1반' 안내판이 위치해 이전 샷의 공간 배치를 잘 유지하고 있습니다.",
      "entities": "서의용(비니, 가죽 재킷)과 나상혁(네이비 재킷, 단정한 머리)의 인상착의가 모두 참조 이미지 및 프롬프트와 일치합니다.",
      "hard_violations": [
       "물리적으로 불가능한 해부학 구조 (나상혁의 왼쪽 팔 끝에 엄지손가락 방향이 뒤집힌 오른손이 달려 있음)"
      ],
      "physics": "나상혁은 책상에 걸터앉아 체중을 지탱하고 있으며 서의용은 두 발로 서 있습니다. 뻗은 팔과 맞잡은 손은 공중에 유지되고 있으나 해부학적으로 성립하지 않습니다."
     },
     {
      "label": "A",
      "direction": "서의용이 나상혁을 바라보고 나상혁이 서의용을 마주봅니다. 카메라는 지시와 정반대로 서의용의 어깨 너머로 나상혁을 향하고 있어, 샷의 핵심인 서의용의 표정을 전혀 담지 못하고 있습니다.",
      "built_space": "전경에 책상 모서리와 켜진 모니터가 보이고, 배경에 복사기와 캐비닛, '강력1반' 안내판이 위치하여 공간의 구성 요소는 일치합니다.",
      "entities": "서의용(비니, 가죽 재킷)과 나상혁(네이비 재킷)의 인상착의가 참조 이미지와 일치합니다.",
      "hard_violations": [
       "물리적으로 불가능한 스테이징 (나상혁과 서의용의 전경/원경 배치를 완전히 뒤바꿔 프롬프트가 요구한 대상의 얼굴을 화면에서 배제함)",
       "물리적으로 불가능한 해부학 구조 (맞잡은 두 손의 엄지손가락 위치와 결합 형태가 논리적으로 불가능하게 뭉개짐)"
      ],
      "physics": "두 사람 모두 서서 바닥에 체중을 지탱하고 있습니다. 손을 맞잡은 동작 자체는 지탱되고 있으나 결합 부위의 해부학적 묘사가 불가능한 형태입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1601,
     "B": 1377
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S13sh6"
  }
 },
 "S22sh7::cine": {
  "applied": true,
  "fingerprint": "cf090a0b66070549579feb91b7e8126d5bb66048000638a32d0d9e6c34c8ea41",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S22sh7_sel.png",
  "source_sha256": "a47b41ee586554d868b7b89493c222a7fa2b6916ec83499fd4e523c425a31322",
  "file": "S22sh7_cine.png",
  "latency_ms": 10320
 },
 "S22sh10::signage": {
  "fp": "fda5b1fcdd397ce3",
  "inscriptions": [
   {
    "surface_native": "부서 현판",
    "text_native": "강력계",
    "reason_ko": "경찰서 강력계 사무실 입구임을 명확히 보여주고 극의 현실감을 더하기 위해 문에 부착된 부서 현판이 필요합니다."
   }
  ]
 },
 "S22sh10": {
  "input_fingerprint": "8588b0539798cfcf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 출입문 앞에 멈춰 선 채 사무실 안쪽을 응시하는 전택수의 전경\n\nLOCATION (lock): At the open doorway of the violent-crimes office, looking inward from the corridor threshold. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the dolly-out settles at waist height inside the office, hold a low wide view slightly off the doorway centerline, showing 전택수 from head to foot within the open entrance. He has stopped with his weight fixed at the threshold and watches the people deeper inside the office, while the doorway outlines him without turning the moment into a frontal portrait.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground; open office doorway around 전택수 in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 열린 출입문 (open) — The opening faces into the office, with 전택수 stopped at its threshold; used as Frames 전택수's full figure and marks his interruption of the office conversation; 사무실 책상 가장자리 (inside the office) — A side edge enters from one side of the interior foreground; used as Provides foreground depth as the camera looks outward from within the room; 강력반 사무실 내부 (visible around the camera-side edges); used as Establishes that 전택수 is looking into an occupied work area without bringing the others into this held frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is rendered in restrained neutral-to-cool tones with naturalistic contrast and procedural clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 부서 현판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 출입문 앞에 멈춰 선 채 사무실 안쪽을 응시하는 전택수의 전경\n\nLOCATION (lock): At the open doorway of the violent-crimes office, looking inward from the corridor threshold. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the dolly-out settles at waist height inside the office, hold a low wide view slightly off the doorway centerline, showing 전택수 from head to foot within the open entrance. He has stopped with his weight fixed at the threshold and watches the people deeper inside the office, while the doorway outlines him without turning the moment into a frontal portrait.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground; open office doorway around 전택수 in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 열린 출입문 (open) — The opening faces into the office, with 전택수 stopped at its threshold; used as Frames 전택수's full figure and marks his interruption of the office conversation; 사무실 책상 가장자리 (inside the office) — A side edge enters from one side of the interior foreground; used as Provides foreground depth as the camera looks outward from within the room; 강력반 사무실 내부 (visible around the camera-side edges); used as Establishes that 전택수 is looking into an occupied work area without bringing the others into this held frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is rendered in restrained neutral-to-cool tones with naturalistic contrast and procedural clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 부서 현판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 출입문 앞에 멈춰 선 채 사무실 안쪽을 응시하는 전택수의 전경\n\nLOCATION (lock): At the open doorway of the violent-crimes office, looking inward from the corridor threshold. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the dolly-out settles at waist height inside the office, hold a low wide view slightly off the doorway centerline, showing 전택수 from head to foot within the open entrance. He has stopped with his weight fixed at the threshold and watches the people deeper inside the office, while the doorway outlines him without turning the moment into a frontal portrait.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground; open office doorway around 전택수 in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 열린 출입문 (open) — The opening faces into the office, with 전택수 stopped at its threshold; used as Frames 전택수's full figure and marks his interruption of the office conversation; 사무실 책상 가장자리 (inside the office) — A side edge enters from one side of the interior foreground; used as Provides foreground depth as the camera looks outward from within the room; 강력반 사무실 내부 (visible around the camera-side edges); used as Establishes that 전택수 is looking into an occupied work area without bringing the others into this held frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is rendered in restrained neutral-to-cool tones with naturalistic contrast and procedural clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 부서 현판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수가 열린 출입문 문지방에 서서 사무실 안쪽(카메라 방향)을 향해 시선을 고정하고 있음.",
    "built_space": "카메라가 사무실 내부에 위치하며, 좌측 전경에 책상 가장자리, 우측에 회색 철제 캐비닛이 있고 중경에 열린 문이 정확한 위치 관계로 배치됨.",
    "entities": "인물은 제공된 캐릭터 레퍼런스의 얼굴, 남색 수사 모자, 재킷, 카키색 바지와 완벽히 일치함. 벽면 현판에 '강력계' 텍스트가 정확히 표기됨.",
    "hard_violations": [],
    "physics": "두 발이 문지방 바닥을 안정적으로 딛고 서 있으며, 의상과 소품의 무게감이 자연스러움."
   },
   {
    "label": "B",
    "direction": "인물이 열린 출입문 문지방에 서서 사무실 안쪽을 바라보고 있음.",
    "built_space": "사무실 내부의 좌측 책상과 우측 캐비닛, 중경의 열린 출입문이 적절하게 묘사됨.",
    "entities": "벽면 현판의 '강력계' 텍스트는 일치하나, 인물이 캐릭터 레퍼런스와 전혀 일치하지 않으며 이전 샷에 등장한 인물의 얼굴과 복장을 차용함.",
    "hard_violations": [
     "이전 샷의 인물 얼굴을 가져옴 (지시된 캐릭터 레퍼런스 불일치 및 차용 금지 규정 위반)"
    ],
    "physics": "인물이 바닥을 딛고 서 있으며 물리적으로 어긋난 부분은 없음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "캐릭터 레퍼런스의 인물 외형과 복장을 완벽히 재현했으며, 지시된 풀샷 프레이밍과 현판 텍스트까지 정확하게 구현함."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "이전 샷의 인물 얼굴을 차용하지 말라는 지시를 명백히 위반했으며, 캐릭터 레퍼런스와 전혀 다른 복장과 외형을 생성함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S22sh7_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 컷에서 벽면을 따라 나란히 배치되어 있던 회색 캐비닛과 철제 사물함 사이에 원래 없던 출입문이 임의로 생성되어 공간 구조가 일치하지 않습니다.",
     "fix_en": "Relocate the doorway so it does not split the grey cabinet and tall lockers; preserve the character, room elements, lighting, and framing.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "회색 캐비닛에 새로 추가되어 붙어 있는 종이들에 의미를 알 수 없는 글자들이 뭉개진 형태로 표기되어 있습니다.",
     "fix_en": "Remove the papers with illegible text from the tall grey cabinet, leaving its metal surface bare; keep the cabinet, surrounding room elements, and character unchanged.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "전택수가 문턱에서 양발에 균등히 서서 몸을 렌즈에 정면으로 맞추고 카메라를 응시하는 초상화 자세다",
     "fix_en": "Shift the character's weight to one leg and direct his head and gaze to look past the camera deeper into the room; maintain his identity, clothing, lighting, and the room's geometry.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "전택수의 시선이 사무실 안쪽이 아닌 카메라 렌즈를 정면으로 향하고 있어, 정면 초상화처럼 보이지 않게 하라는 프롬프트 지시를 위반했습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 컷에서 벽면을 따라 나란히 배치되어 있던 회색 캐비닛과 철제 사물함 사이에 원래 없던 출입문이 임의로 생성되어 공간 구조가 일치하지 않습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "회색 캐비닛에 새로 추가되어 붙어 있는 종이들에 의미를 알 수 없는 글자들이 뭉개진 형태로 표기되어 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "전택수가 문턱에서 양발에 균등히 서서 몸을 렌즈에 정면으로 맞추고 카메라를 응시하는 초상화 자세다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 컷에서 벽면을 따라 나란히 배치되어 있던 회색 캐비닛과 철제 사물함 사이에 원래 없던 출입문이 임의로 생성되어 공간 구조가 일치하지 않습니다.",
    "fix_en": "Relocate the doorway so it does not split the grey cabinet and tall lockers; preserve the character, room elements, lighting, and framing.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "회색 캐비닛에 새로 추가되어 붙어 있는 종이들에 의미를 알 수 없는 글자들이 뭉개진 형태로 표기되어 있습니다.",
    "fix_en": "Remove the papers with illegible text from the tall grey cabinet, leaving its metal surface bare; keep the cabinet, surrounding room elements, and character unchanged.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "전택수가 문턱에서 양발에 균등히 서서 몸을 렌즈에 정면으로 맞추고 카메라를 응시하는 초상화 자세다",
    "fix_en": "Shift the character's weight to one leg and direct his head and gaze to look past the camera deeper into the room; maintain his identity, clothing, lighting, and the room's geometry.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh7"
  }
 },
 "S22sh10::cine": {
  "applied": true,
  "fingerprint": "eb3d145cc51787dc7be4a6551d13dc38dea7775bede7d9406ff455c6d48c543b",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S22sh10_sel.png",
  "source_sha256": "c3eafa4604394b9acf9f461f6b9b939f4c5d276686444793fe13112b0a858748",
  "file": "S22sh10_cine.png",
  "latency_ms": 12877
 },
 "S23sh9::signage": {
  "fp": "4c8a869256919ce4",
  "inscriptions": [
   {
    "surface_native": "식당 벽면의 메뉴판",
    "text_native": "광어 우럭 매운탕",
    "reason_ko": "시장 골목의 허름한 횟집이라는 공간적 배경을 실감 나게 묘사하기 위해 벽면에 걸린 전형적인 메뉴 표기가 필요합니다."
   }
  ]
 },
 "era_assess::fe739f7b16bf9b6c": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 한국 시장 골목의 허름한 횟집 내부",
    "search_terms_native": [
     "시장 골목 횟집 내부",
     "오래된 횟집 실내",
     "횟집 실내포차 테이블",
     "횟집 벽 메뉴판"
    ],
    "language_lock_native": "검색 결과는 오직 한국어로 작성된 한국 웹사이트의 자료만을 포함해야 하며 다른 언어로 번역하거나 결합하지 마십시오.",
    "reason_ko": "2000년대 한국 시장 골목 횟집 특유의 플라스틱/스테인리스 테이블, 벽면의 종이 메뉴판, 한국식 소주병과 식기류의 형태는 일반적인 이미지 생성기가 정확히 묘사하기 어렵습니다."
   }
  ]
 },
 "era_fail::a553dc1cc8ae880a": {
  "stage": "research",
  "subject": "2000년대-2010년대 한국 시장 골목의 허름한 횟집 내부",
  "attempts": 4
 },
 "groupbg::시장 횟집 내부": {
  "input_fingerprint": "30fd7981f5d37da5",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "시장 횟집 내부",
    "tags": [
     "S23sh11",
     "S23sh9"
    ]
   },
   "context_sig": "a4458a5104aa1da6"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: At a corner table inside a shabby fish restaurant in the market district.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n횟집 내부: 시장 골목에 위치한 서민적인 분위기의 허름하고 비좁은 식당이다. (특징: 얇은 플라스틱이나 스테인리스 상판의 낡은 테이블; 벽면에 어지럽게 붙은 종이 메뉴판들; 테이블 위에 놓인 소주병과 소박한 안주 접시)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 횟집 - 밤\n- 시장통의 허름한 횟집. 구석 테이블에 택수, 주철, 의용, 상혁이 앉아있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: At a corner table inside a shabby fish restaurant in the market district.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n횟집 내부: 시장 골목에 위치한 서민적인 분위기의 허름하고 비좁은 식당이다. (특징: 얇은 플라스틱이나 스테인리스 상판의 낡은 테이블; 벽면에 어지럽게 붙은 종이 메뉴판들; 테이블 위에 놓인 소주병과 소박한 안주 접시)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 횟집 - 밤\n- 시장통의 허름한 횟집. 구석 테이블에 택수, 주철, 의용, 상혁이 앉아있다.\n\nTIME OF DAY (lock): night.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_시장_횟집_내부_3bc1f8.png",
  "asset_id": "141d32bc-4464-4303-a136-3c35b1679d34",
  "input_asset_ids": [
   "373f35cf-edcf-4da1-97a4-1d48bd84655a"
  ],
  "origin_tag": "S23sh9",
  "place_text": "At a corner table inside a shabby fish restaurant in the market district.",
  "origin_inputs": {
   "place_text": "At a corner table inside a shabby fish restaurant in the market district.",
   "time_of_day_en": "night",
   "conti_asset_id": "373f35cf-edcf-4da1-97a4-1d48bd84655a"
  }
 },
 "era_assess::f575a598f5c9a0ae": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 시장 횟집 내부",
    "search_terms_native": [
     "시장 횟집 내부",
     "횟집 비닐 테이블",
     "횟집 플라스틱 의자",
     "초장통 테이블"
    ],
    "language_lock_native": "모든 검색어는 오직 한국어로만 작성해야 하며 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "AI 모델은 한국 시장 횟집 특유의 비닐 씌운 테이블, 초록색 멜라민 식기, 플라스틱 의자 및 초장/간장 통이 올려진 전형적인 인테리어를 재현하지 못하고 서구식 또는 고급 일식 레스토랑처럼 그리기 쉽습니다."
   }
  ]
 },
 "era_ref::076e7972e7a5ef6f": {
  "subject": "2000년대 한국 시장 횟집 내부",
  "terms": [
   "시장 횟집 내부",
   "횟집 비닐 테이블",
   "횟집 플라스틱 의자",
   "초장통 테이블"
  ],
  "queries": [
   [
    "2000년대 한국 시장 횟집 내부 비닐 테이블 플라스틱 의자 초장통",
    "옛날 한국 재래시장 횟집 내부 초장통 테이블 플라스틱 의자"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://d12zq4w4guyljn.cloudfront.net/750_750_20250511054635777_photo_d4f75029907a.webp",
  "picked_reason_ko": "3번은 수산시장 안의 횟집 식사 공간을 주제로 삼아 금속 원형 탁자, 붉은 플라스틱 의자, 식기와 냉장고, 노출 설비 등 2000년대 한국 시장 횟집의 일상적인 내부 구성을 가장 명확하게 보여준다.",
  "sha256": "69d670c5bdd10d9672e8bda80b70ae18fdca26f9ab3bb7d7c5246d267f8bf3a0",
  "file": "eraref_076e7972e7a5ef6f.png"
 },
 "S23sh9::bgfirst_bg": {
  "input_fingerprint": "66c993344579b664",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블을 사이에 두고 서로의 눈을 가만히 마주 보는 전택수와 서의용의 팽팽한 시선 교환\n\nLOCATION (lock): At a corner table inside a shabby fish restaurant in the market district.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly at seated eye height beside the corner table, close to the gaze line but offset so 전택수 and 서의용 remain visible in opposing three-quarter angles across the table. Place 전택수 in the left foreground and 서의용 in the right midground, both momentarily still and absorbed in each other's eyes while the table edge holds the tension between them.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 전택수 across the table; corner table separating them in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 구석 테이블 (occupied by the group) — Its near edge crosses the lower frame and recedes between 전택수 and 서의용; used as Physically separates the two men and carries their glasses at the bottom of the composition; 소주잔들 (held or resting around the table) — Their open rims and side profiles are visible near the two seated men; used as Keeps the impending toast present without interrupting the sustained eye contact; 허름한 횟집 내부 (nighttime dining space around the corner table); used as Provides understated context for the private tension within the group gathering.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant illumination is kept subdued and naturalistic, with neutral-to-cool rendering and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 시장 횟집 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블을 사이에 두고 서로의 눈을 가만히 마주 보는 전택수와 서의용의 팽팽한 시선 교환\n\nLOCATION (lock): At a corner table inside a shabby fish restaurant in the market district.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Finish the dolly at seated eye height beside the corner table, close to the gaze line but offset so 전택수 and 서의용 remain visible in opposing three-quarter angles across the table. Place 전택수 in the left foreground and 서의용 in the right midground, both momentarily still and absorbed in each other's eyes while the table edge holds the tension between them.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 전택수 across the table; corner table separating them in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 구석 테이블 (occupied by the group) — Its near edge crosses the lower frame and recedes between 전택수 and 서의용; used as Physically separates the two men and carries their glasses at the bottom of the composition; 소주잔들 (held or resting around the table) — Their open rims and side profiles are visible near the two seated men; used as Keeps the impending toast present without interrupting the sustained eye contact; 허름한 횟집 내부 (nighttime dining space around the corner table); used as Provides understated context for the private tension within the group gathering.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant illumination is kept subdued and naturalistic, with neutral-to-cool rendering and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 시장 횟집 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S23sh9__bgfirst_bg.png",
  "asset_id": "692afe3c-5fe9-4c98-bfb7-3db9cf98432b",
  "input_asset_ids": [
   "373f35cf-edcf-4da1-97a4-1d48bd84655a",
   "141d32bc-4464-4303-a136-3c35b1679d34"
  ],
  "era_research": {
   "subject": "2000년대 한국 시장 횟집 내부",
   "queries": [
    [
     "2000년대 한국 시장 횟집 내부 비닐 테이블 플라스틱 의자 초장통",
     "옛날 한국 재래시장 횟집 내부 초장통 테이블 플라스틱 의자"
    ]
   ],
   "picked_url": "https://d12zq4w4guyljn.cloudfront.net/750_750_20250511054635777_photo_d4f75029907a.webp",
   "sha256": "69d670c5bdd10d9672e8bda80b70ae18fdca26f9ab3bb7d7c5246d267f8bf3a0",
   "file": "eraref_076e7972e7a5ef6f.png"
  }
 },
 "S23sh9": {
  "input_fingerprint": "25f8e2c985eee2b0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 서로의 눈을 가만히 마주 보는 전택수와 서의용의 팽팽한 시선 교환\n\nLOCATION (lock): At a corner table inside a shabby fish restaurant in the market district. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly at seated eye height beside the corner table, close to the gaze line but offset so 전택수 and 서의용 remain visible in opposing three-quarter angles across the table. Place 전택수 in the left foreground and 서의용 in the right midground, both momentarily still and absorbed in each other's eyes while the table edge holds the tension between them.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 전택수 across the table; corner table separating them in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 구석 테이블 (occupied by the group) — Its near edge crosses the lower frame and recedes between 전택수 and 서의용; used as Physically separates the two men and carries their glasses at the bottom of the composition; 소주잔들 (held or resting around the table) — Their open rims and side profiles are visible near the two seated men; used as Keeps the impending toast present without interrupting the sustained eye contact; 허름한 횟집 내부 (nighttime dining space around the corner table); used as Provides understated context for the private tension within the group gathering.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant illumination is kept subdued and naturalistic, with neutral-to-cool rendering and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 메뉴판: \"광어 우럭 매운탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 서로의 눈을 가만히 마주 보는 전택수와 서의용의 팽팽한 시선 교환\n\nLOCATION (lock): At a corner table inside a shabby fish restaurant in the market district. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly at seated eye height beside the corner table, close to the gaze line but offset so 전택수 and 서의용 remain visible in opposing three-quarter angles across the table. Place 전택수 in the left foreground and 서의용 in the right midground, both momentarily still and absorbed in each other's eyes while the table edge holds the tension between them.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 전택수 across the table; corner table separating them in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 구석 테이블 (occupied by the group) — Its near edge crosses the lower frame and recedes between 전택수 and 서의용; used as Physically separates the two men and carries their glasses at the bottom of the composition; 소주잔들 (held or resting around the table) — Their open rims and side profiles are visible near the two seated men; used as Keeps the impending toast present without interrupting the sustained eye contact; 허름한 횟집 내부 (nighttime dining space around the corner table); used as Provides understated context for the private tension within the group gathering.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant illumination is kept subdued and naturalistic, with neutral-to-cool rendering and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 메뉴판: \"광어 우럭 매운탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 서로의 눈을 가만히 마주 보는 전택수와 서의용의 팽팽한 시선 교환\n\nLOCATION (lock): At a corner table inside a shabby fish restaurant in the market district. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly at seated eye height beside the corner table, close to the gaze line but offset so 전택수 and 서의용 remain visible in opposing three-quarter angles across the table. Place 전택수 in the left foreground and 서의용 in the right midground, both momentarily still and absorbed in each other's eyes while the table edge holds the tension between them.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 전택수 across the table; corner table separating them in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 구석 테이블 (occupied by the group) — Its near edge crosses the lower frame and recedes between 전택수 and 서의용; used as Physically separates the two men and carries their glasses at the bottom of the composition; 소주잔들 (held or resting around the table) — Their open rims and side profiles are visible near the two seated men; used as Keeps the impending toast present without interrupting the sustained eye contact; 허름한 횟집 내부 (nighttime dining space around the corner table); used as Provides understated context for the private tension within the group gathering.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant illumination is kept subdued and naturalistic, with neutral-to-cool rendering and restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 메뉴판: \"광어 우럭 매운탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S23sh9__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S23sh9.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_시장_횟집_내부_3bc1f8.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 로케이션 레퍼런스를 완전히 무시한 다른 공간을 구현했으며, 프롬프트에 없는 배경 인물들을 다수 추가하여 치명적인 오류를 범했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "로케이션 레퍼런스의 공간(스테인리스 테이블, 벽면, 배경 구조)을 매우 정확하게 재현했으며, 두 인물의 배치, 의상, 시선 교환을 성공적으로 연출했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수와 서의용이 테이블을 사이에 두고 서로의 눈을 응시함.",
      "built_space": "목재 테이블과 의자가 배치되어 있으며, 배경의 벽면이나 출입구 구조가 로케이션 레퍼런스 사진과 전혀 다름.",
      "entities": "전택수(좌측)와 서의용(우측)의 외형 및 의상은 레퍼런스와 일치함. 명시되지 않은 여러 명의 배경 인물들이 테이블에 앉아 있음. 간판 텍스트가 나무 판자에 적혀 있음.",
      "hard_violations": [
       "프롬프트에 명시되지 않은 인물들(배경의 손님들) 추가",
       "LOCATION lock 위반 (제공된 로케이션 레퍼런스 사진과 다른 공간)"
      ],
      "physics": "두 인물 모두 의자에 앉아 팔을 테이블 등에 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "전택수와 서의용이 서로의 눈을 팽팽하게 마주보고 있음.",
      "built_space": "로케이션 레퍼런스와 일치하는 스테인리스 스틸 테이블, 검은색 의자, 벽면 질감, 창문, 포스터 및 식당 내부 구조(냉장고 등)가 정확히 배치됨.",
      "entities": "전택수(좌측)와 서의용(우측)의 외형 및 의상이 레퍼런스와 일치하며, 테이블 위에 소주잔들이 놓여 있음. 추가된 인물 없음. 벽면 메뉴판에 텍스트가 구현됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 정상적으로 의자에 앉아 자세를 유지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 로케이션 레퍼런스를 완전히 무시한 다른 공간을 구현했으며, 프롬프트에 없는 배경 인물들을 다수 추가하여 치명적인 오류를 범했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "로케이션 레퍼런스의 공간(스테인리스 테이블, 벽면, 배경 구조)을 매우 정확하게 재현했으며, 두 인물의 배치, 의상, 시선 교환을 성공적으로 연출했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수와 서의용이 테이블을 사이에 두고 서로의 눈을 응시함.",
      "built_space": "목재 테이블과 의자가 배치되어 있으며, 배경의 벽면이나 출입구 구조가 로케이션 레퍼런스 사진과 전혀 다름.",
      "entities": "전택수(좌측)와 서의용(우측)의 외형 및 의상은 레퍼런스와 일치함. 명시되지 않은 여러 명의 배경 인물들이 테이블에 앉아 있음. 간판 텍스트가 나무 판자에 적혀 있음.",
      "hard_violations": [
       "프롬프트에 명시되지 않은 인물들(배경의 손님들) 추가",
       "LOCATION lock 위반 (제공된 로케이션 레퍼런스 사진과 다른 공간)"
      ],
      "physics": "두 인물 모두 의자에 앉아 팔을 테이블 등에 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "전택수와 서의용이 서로의 눈을 팽팽하게 마주보고 있음.",
      "built_space": "로케이션 레퍼런스와 일치하는 스테인리스 스틸 테이블, 검은색 의자, 벽면 질감, 창문, 포스터 및 식당 내부 구조(냉장고 등)가 정확히 배치됨.",
      "entities": "전택수(좌측)와 서의용(우측)의 외형 및 의상이 레퍼런스와 일치하며, 테이블 위에 소주잔들이 놓여 있음. 추가된 인물 없음. 벽면 메뉴판에 텍스트가 구현됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 정상적으로 의자에 앉아 자세를 유지하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 장소 레퍼런스를 완벽히 구현했으며, 프롬프트에 명시된 두 인물만의 팽팽한 시선 교환과 메뉴판 텍스트를 정확하게 연출함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트에 없는 배경 인물들을 임의로 다수 추가하였고, 고정된 장소 레퍼런스(테이블, 구조)를 무시하여 핵심 규칙을 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽의 전택수와 오른쪽의 서의용이 서로의 눈을 똑바로 응시하고 있음.",
      "built_space": "레퍼런스와 동일한 스테인리스 재질의 원형 화구가 있는 테이블, 검은색 의자, 창문 및 벽면 구조가 정확하게 배치됨.",
      "entities": "전택수와 서의용의 외모 및 의상이 레퍼런스와 일치함. 명시된 인물 외에 다른 사람은 없음. 테이블 위에 소주잔들이 놓여 있고, 왼쪽 벽면에 '광어 우럭 매운탕' 텍스트가 적힌 메뉴판이 있음.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 안정적으로 앉아 체중을 지탱하고 있으며, 손과 팔도 자연스럽게 놓여 있음."
     },
     {
      "label": "B",
      "direction": "전택수와 서의용이 테이블을 사이에 두고 서로를 바라봄.",
      "built_space": "레퍼런스와 전혀 다른 나무 테이블과 의자가 배치되었으며, 식당 내부 공간 구조와 조명도 임의로 변경됨.",
      "entities": "전택수와 서의용의 외모와 복장은 일치함. 상단에 '광어 우럭 매운탕' 간판이 있으나, 배경에 지시문에 없는 다수의 엑스트라 인물들이 존재함.",
      "hard_violations": [
       "지시문에 없는 배경 인물 다수 추가 (invented people)",
       "고정된 장소 레퍼런스(LOCATION lock) 위반 (다른 재질의 테이블과 공간 구조)"
      ],
      "physics": "두 인물과 배경 인물들 모두 의자에 정상적으로 앉아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 장소 레퍼런스를 완벽히 구현했으며, 프롬프트에 명시된 두 인물만의 팽팽한 시선 교환과 메뉴판 텍스트를 정확하게 연출함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트에 없는 배경 인물들을 임의로 다수 추가하였고, 고정된 장소 레퍼런스(테이블, 구조)를 무시하여 핵심 규칙을 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "왼쪽의 전택수와 오른쪽의 서의용이 서로의 눈을 똑바로 응시하고 있음.",
      "built_space": "레퍼런스와 동일한 스테인리스 재질의 원형 화구가 있는 테이블, 검은색 의자, 창문 및 벽면 구조가 정확하게 배치됨.",
      "entities": "전택수와 서의용의 외모 및 의상이 레퍼런스와 일치함. 명시된 인물 외에 다른 사람은 없음. 테이블 위에 소주잔들이 놓여 있고, 왼쪽 벽면에 '광어 우럭 매운탕' 텍스트가 적힌 메뉴판이 있음.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 안정적으로 앉아 체중을 지탱하고 있으며, 손과 팔도 자연스럽게 놓여 있음."
     },
     {
      "label": "A",
      "direction": "전택수와 서의용이 테이블을 사이에 두고 서로를 바라봄.",
      "built_space": "레퍼런스와 전혀 다른 나무 테이블과 의자가 배치되었으며, 식당 내부 공간 구조와 조명도 임의로 변경됨.",
      "entities": "전택수와 서의용의 외모와 복장은 일치함. 상단에 '광어 우럭 매운탕' 간판이 있으나, 배경에 지시문에 없는 다수의 엑스트라 인물들이 존재함.",
      "hard_violations": [
       "지시문에 없는 배경 인물 다수 추가 (invented people)",
       "고정된 장소 레퍼런스(LOCATION lock) 위반 (다른 재질의 테이블과 공간 구조)"
      ],
      "physics": "두 인물과 배경 인물들 모두 의자에 정상적으로 앉아 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수와 서의용이 테이블을 사이에 두고 서로의 눈을 응시함.",
    "built_space": "목재 테이블과 의자가 배치되어 있으며, 배경의 벽면이나 출입구 구조가 로케이션 레퍼런스 사진과 전혀 다름.",
    "entities": "전택수(좌측)와 서의용(우측)의 외형 및 의상은 레퍼런스와 일치함. 명시되지 않은 여러 명의 배경 인물들이 테이블에 앉아 있음. 간판 텍스트가 나무 판자에 적혀 있음.",
    "hard_violations": [
     "프롬프트에 명시되지 않은 인물들(배경의 손님들) 추가",
     "LOCATION lock 위반 (제공된 로케이션 레퍼런스 사진과 다른 공간)"
    ],
    "physics": "두 인물 모두 의자에 앉아 팔을 테이블 등에 지지하고 있음."
   },
   {
    "label": "B",
    "direction": "전택수와 서의용이 서로의 눈을 팽팽하게 마주보고 있음.",
    "built_space": "로케이션 레퍼런스와 일치하는 스테인리스 스틸 테이블, 검은색 의자, 벽면 질감, 창문, 포스터 및 식당 내부 구조(냉장고 등)가 정확히 배치됨.",
    "entities": "전택수(좌측)와 서의용(우측)의 외형 및 의상이 레퍼런스와 일치하며, 테이블 위에 소주잔들이 놓여 있음. 추가된 인물 없음. 벽면 메뉴판에 텍스트가 구현됨.",
    "hard_violations": [],
    "physics": "두 인물 모두 정상적으로 의자에 앉아 자세를 유지하고 있음."
   }
  ],
  "totals": {
   "A": 6,
   "B": 16
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "지정된 로케이션 레퍼런스를 완전히 무시한 다른 공간을 구현했으며, 프롬프트에 없는 배경 인물들을 다수 추가하여 치명적인 오류를 범했습니다."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "로케이션 레퍼런스의 공간(스테인리스 테이블, 벽면, 배경 구조)을 매우 정확하게 재현했으며, 두 인물의 배치, 의상, 시선 교환을 성공적으로 연출했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_시장_횟집_내부_3bc1f8.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "왼쪽 인물(전택수)의 재킷 소매와 가슴에 원본과 다른 왜곡된 영문 텍스트('PROICE', 'SALICS')가 생성됨.",
     "fix_en": "Remove the distorted English letters 'PROICE' and 'SALICS' from the left man's jacket, leaving the fabric as plain dark blue. Preserve the two men, their poses, the table, the glasses, and the background.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "왼쪽 벽면 텍스트('광어 우럭 매운탕') 하단과 중앙 벽면 메뉴판의 글자들이 알아볼 수 없게 뭉개지고 왜곡됨.",
     "fix_en": "Replace the garbled text below '광어 우럭 매운탕' and on the center wall menu with illegible, natural-looking blurred strokes. Preserve the men, the main readable signage, the table, and the room's lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "테이블 위에 놓인 잔들이 일반적인 소주잔 형태가 아닌 과도하게 큰 크기의 유리잔으로 묘사됨.",
     "fix_en": "Replace the oversized glasses on the table with standard, small, clear glass soju shot glasses. Preserve the men, the table surface, and the background.",
     "severity": "minor",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "왼쪽 인물(전택수)의 재킷 소매와 가슴에 원본과 다른 왜곡된 영문 텍스트('PROICE', 'SALICS')가 생성됨.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 벽면 텍스트('광어 우럭 매운탕') 하단과 중앙 벽면 메뉴판의 글자들이 알아볼 수 없게 뭉개지고 왜곡됨.",
     "severity": "major"
    },
    {
     "issue_ko": "테이블 위에 놓인 잔들이 일반적인 소주잔 형태가 아닌 과도하게 큰 크기의 유리잔으로 묘사됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "왼쪽 벽 작은 메뉴 종이의 한글 일부가 깨지거나 의미 없는 문구로 되어 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 인물(전택수)의 재킷 소매와 가슴에 원본과 다른 왜곡된 영문 텍스트('PROICE', 'SALICS')가 생성됨.",
    "fix_en": "Remove the distorted English letters 'PROICE' and 'SALICS' from the left man's jacket, leaving the fabric as plain dark blue. Preserve the two men, their poses, the table, the glasses, and the background.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "왼쪽 벽면 텍스트('광어 우럭 매운탕') 하단과 중앙 벽면 메뉴판의 글자들이 알아볼 수 없게 뭉개지고 왜곡됨.",
    "fix_en": "Replace the garbled text below '광어 우럭 매운탕' and on the center wall menu with illegible, natural-looking blurred strokes. Preserve the men, the main readable signage, the table, and the room's lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "테이블 위에 놓인 잔들이 일반적인 소주잔 형태가 아닌 과도하게 큰 크기의 유리잔으로 묘사됨.",
    "fix_en": "Replace the oversized glasses on the table with standard, small, clear glass soju shot glasses. Preserve the men, the table surface, and the background.",
    "severity": "minor",
    "observation_index": 2
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S23sh9__bgfirst_bg.png",
   "bg_asset_id": "692afe3c-5fe9-4c98-bfb7-3db9cf98432b",
   "bg_record_key": "S23sh9::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "시장 횟집 내부",
   "groupbg_asset_id": "141d32bc-4464-4303-a136-3c35b1679d34"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S23sh9::cine": {
  "applied": true,
  "fingerprint": "f32d5254d80b7728f4e0b63bd01383734235c99862dd5a2b408c4f2e8509ec91",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S23sh9_sel.png",
  "source_sha256": "b35fdc225a4ba7819e19f2eab59501d7f49b4ab9d9056aa250f771eacd4d1813",
  "file": "S23sh9_cine.png",
  "latency_ms": 12262
 },
 "S23sh11::signage": {
  "fp": "968271cddbc1f517",
  "inscriptions": [
   {
    "surface_native": "식당 벽면의 메뉴판",
    "text_native": "광어\n우럭\n매운탕",
    "reason_ko": "2000년대 한국 횟집 구석 자리 벽면에 걸려 있는 전형적인 메뉴판을 표현하여 사실적인 식당 내부 분위기를 더하기 위함."
   }
  ]
 },
 "S23sh11": {
  "input_fingerprint": "127f8937fb12cf7f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수, 서의용, 강주철, 나상혁의 네 손이 쥔 소주잔들이 한데 모여 겹쳐진 순간\n\nLOCATION (lock): Over the corner dining table inside the fish restaurant. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane close above the center of the corner table at a steep high angle, aiming downward as four naturally varied arms enter from different edges and bring their soju glasses into contact at center. Keep the hands asynchronous in wrist angle and reach—전택수 from upper left, 서의용 from lower right, 강주철 from lower left, and 나상혁 from upper right—while retaining enough table around them to preserve realistic scale.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 네 개의 소주잔 (touching in a shared toast) — Their rims meet near the center while each glass remains angled from a different side of the table; used as Primary action detail at the table center, with four separate hands and rims remaining distinguishable; 구석 테이블 상판 (supporting the group setting) — Its upper face is seen from the steep high angle around the four glasses; used as Provides scale reference and negative space around the converging hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant light remains subdued and naturalistic, preserving restrained tonal separation among the four hands and glasses.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용, 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same worn seafood-restaurant interior, table setting, dim night lighting, and the four seated people's clothing. Exclude the earlier locked gaze and focus only on the four raised shot glasses.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수, 서의용, 강주철, 나상혁 right now — 전택수, 서의용, 강주철, 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수, 서의용, 강주철, 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 메뉴판: \"광어\n우럭\n매운탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수, 서의용, 강주철, 나상혁의 네 손이 쥔 소주잔들이 한데 모여 겹쳐진 순간\n\nLOCATION (lock): Over the corner dining table inside the fish restaurant. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane close above the center of the corner table at a steep high angle, aiming downward as four naturally varied arms enter from different edges and bring their soju glasses into contact at center. Keep the hands asynchronous in wrist angle and reach—전택수 from upper left, 서의용 from lower right, 강주철 from lower left, and 나상혁 from upper right—while retaining enough table around them to preserve realistic scale.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 네 개의 소주잔 (touching in a shared toast) — Their rims meet near the center while each glass remains angled from a different side of the table; used as Primary action detail at the table center, with four separate hands and rims remaining distinguishable; 구석 테이블 상판 (supporting the group setting) — Its upper face is seen from the steep high angle around the four glasses; used as Provides scale reference and negative space around the converging hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant light remains subdued and naturalistic, preserving restrained tonal separation among the four hands and glasses.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용, 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same worn seafood-restaurant interior, table setting, dim night lighting, and the four seated people's clothing. Exclude the earlier locked gaze and focus only on the four raised shot glasses.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수, 서의용, 강주철, 나상혁 right now — 전택수, 서의용, 강주철, 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수, 서의용, 강주철, 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 메뉴판: \"광어\n우럭\n매운탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수, 서의용, 강주철, 나상혁의 네 손이 쥔 소주잔들이 한데 모여 겹쳐진 순간\n\nLOCATION (lock): Over the corner dining table inside the fish restaurant. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane close above the center of the corner table at a steep high angle, aiming downward as four naturally varied arms enter from different edges and bring their soju glasses into contact at center. Keep the hands asynchronous in wrist angle and reach—전택수 from upper left, 서의용 from lower right, 강주철 from lower left, and 나상혁 from upper right—while retaining enough table around them to preserve realistic scale.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 네 개의 소주잔 (touching in a shared toast) — Their rims meet near the center while each glass remains angled from a different side of the table; used as Primary action detail at the table center, with four separate hands and rims remaining distinguishable; 구석 테이블 상판 (supporting the group setting) — Its upper face is seen from the steep high angle around the four glasses; used as Provides scale reference and negative space around the converging hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime restaurant light remains subdued and naturalistic, preserving restrained tonal separation among the four hands and glasses.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용, 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same worn seafood-restaurant interior, table setting, dim night lighting, and the four seated people's clothing. Exclude the earlier locked gaze and focus only on the four raised shot glasses.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수, 서의용, 강주철, 나상혁 right now — 전택수, 서의용, 강주철, 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수, 서의용, 강주철, 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면의 메뉴판: \"광어\n우럭\n매운탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "gq": {
   "route": "combined",
   "gap": 0.4,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.667
   },
   "adjusted": {
    "A": 0.6,
    "B": 0.417
   },
   "violations": {
    "A": [
     "[gemini-pro] 이전 샷에서 고정된 철제 테이블을 둥근 나무 테이블로 임의 변경함 (장소 설정 위반).",
     "[gemini-pro] 벽에 붙은 종이 메뉴판을 바닥에 세워진 입간판 형태로 렌더링함 (구조 위반).",
     "[gemini-pro] 손과 잔에만 집중하는 '인서트 클로즈업' 지시를 무시하고 인물들의 상반신과 얼굴을 포함한 넓은 화각으로 연출함 (프레이밍 위반).",
     "[gemini-pro] 서의용(우측 하단 지정)과 전택수(좌측 상단 지정)의 위치가 잘못 배치됨 (스테이징 위반)."
    ],
    "B": [
     "[gemini-pro] 이전 샷에서 고정된 철제 테이블을 나무 테이블로 임의 변경함 (장소 설정 위반).",
     "[gemini-pro] 벽에 붙은 종이 메뉴판을 세워진 형태로 렌더링함 (구조 위반).",
     "[gemini-pro] '인서트 클로즈업' 지시를 무시하고 인물들의 얼굴과 상반신을 모두 노출시킴 (프레이밍 위반).",
     "[gemini-pro] 모든 인물의 지정된 등장 위치(전택수 좌측 상단, 서의용 우측 하단 등)를 어김 (스테이징 위반).",
     "[gemini-pro] 우측 상단 인물에게 지정되지 않은 베이지색 스웨터를 입힘 (의상 위반)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 600,
   "B": 417
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 600,
    "verdict_ko": "요구된 인서트 클로즈업을 무시하고 인물의 상반신을 포함한 넓은 화각으로 연출했으며, 이전 샷에서 확정된 철제 테이블 대신 나무 테이블을 배치하고 인물들의 지정된 위치를 어겨 핵심 지시사항을 모두 위반했습니다.  ★위반: [gemini-pro] 이전 샷에서 고정된 철제 테이블을 둥근 나무 테이블로 임의 변경함 (장소 설정 위반). / [gemini-pro] 벽에 붙은 종이 메뉴판을 바닥에 세워진 입간판 형태로 렌더링함 (구조 위반). / [gemini-pro] 손과 잔에만 집중하는 '인서트 클로즈업' 지시를 무시하고 인물들의 상반신과 얼굴을 포함한 넓은 화각으로 연출함 (프레이밍 위반). / [gemini-pro] 서의용(우측 하단 지정)과 전택수(좌측 상단 지정)의 위치가 잘못 배치됨 (스테이징 위반)."
   },
   {
    "label": "B",
    "score": 417,
    "verdict_ko": "인서트 클로즈업 지시와 이전 샷의 테이블 재질(철제)을 모두 무시했으며, 우측 상단 인물에게 지정되지 않은 베이지색 스웨터를 입히고 전체 인물의 배치 위치를 틀리게 렌더링하여 실격입니다.  ★위반: [gemini-pro] 이전 샷에서 고정된 철제 테이블을 나무 테이블로 임의 변경함 (장소 설정 위반). / [gemini-pro] 벽에 붙은 종이 메뉴판을 세워진 형태로 렌더링함 (구조 위반). / [gemini-pro] '인서트 클로즈업' 지시를 무시하고 인물들의 얼굴과 상반신을 모두 노출시킴 (프레이밍 위반). / [gemini-pro] 모든 인물의 지정된 등장 위치(전택수 좌측 상단, 서의용 우측 하단 등)를 어김 (스테이징 위반). / [gemini-pro] 우측 상단 인물에게 지정되지 않은 베이지색 스웨터를 입힘 (의상 위반)."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 서의용, 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S23sh9_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "샷 텍스트와 프레이밍 규정에서 'insert close-up'으로 손과 소주잔의 디테일만 잡도록 지시했으나, 인물들의 상반신과 얼굴이 넓게 프레임에 포함됨.",
     "fix_en": "Apply a heavy vignette and blur to the upper bodies and faces to artificially focus attention solely on the hands and glasses in the center.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이전 샷(Previous Shot Still)에서 확정된 은색 금속 사각형 테이블이 아닌, 원목 재질의 둥근 테이블로 렌더링됨.",
     "fix_en": "Replace the wooden texture of the table surface with a brushed stainless steel texture, maintaining its current round shape, the hands, and glasses.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "프롬프트 상 상단 좌측(upper left)은 전택수(경찰 재킷, 모자)가 위치해야 하나, 이전 샷에서 비니를 썼던 서의용의 가죽 재킷을 입은 인물(비니 미착용)이 배치됨.",
     "fix_en": "Change the upper-left person's clothing to a navy blue police jacket over a plaid shirt and add a navy cap, matching 전택수, preserving his pose, the table, and lighting.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "상단 우측(upper right)은 나상혁(네이비 보머 재킷)이 위치해야 하나, 레퍼런스에 없는 어두운 무지 스웨터를 입은 인물이 배치됨.",
     "fix_en": "Change the upper-right person's clothing to a navy bomber jacket over a light blue collared shirt, matching 나상혁, preserving his pose, the other people, and the background.",
     "severity": "critical",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "하단 우측(lower right)은 서의용(가죽 재킷)이 위치해야 하나, 나상혁의 의상인 네이비 재킷과 하늘색 셔츠 소매를 입은 팔이 배치됨.",
     "fix_en": "Change the lower-right arm's clothing to a black leather jacket over a dark grey hoodie, matching 서의용, preserving the hand's pose, the glass, and the rest of the scene.",
     "severity": "critical",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "벽면에 걸려 있어야 할 '광어 우럭 매운탕' 메뉴판이 식당 벽이 아닌 상단 좌측 인물 바로 뒤 낮은 위치에 비정상적으로 놓여 있음.",
     "fix_en": "Remove the freestanding menu sign behind the upper-left person and place a flat menu sign on the background wall, preserving the characters and lighting.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "우측 상단 인물이 나상혁 참조의 해사한 얼굴·단정한 머리가 아니라 다른 얼굴이다.",
     "fix_en": "Replace the face of the upper-right person with the youthful face and neat short black hair matching 나상혁, preserving his gaze, pose, and the lighting.",
     "severity": "critical",
     "observation_index": 7,
     "needs_regeneration": true
    },
    {
     "issue_ko": "좌측 하단 인물이 강주철 참조의 둥근 얼굴이 아니라 다른 얼굴이다.",
     "fix_en": "Replace the face of the lower-left person with the round face and short black hair matching 강주철, preserving his gaze, pose, and the lighting.",
     "severity": "critical",
     "observation_index": 8,
     "needs_regeneration": true
    },
    {
     "issue_ko": "좌측 하단 의상이 강주철 참조의 주황·회색 등산 재킷이 아니라 녹·주황 배색 재킷이다.",
     "fix_en": "Change the lower-left person's jacket to an orange and dark grey hiking jacket matching 강주철, preserving his pose, the table, and the rest of the scene.",
     "severity": "major",
     "observation_index": 12,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "샷 텍스트와 프레이밍 규정에서 'insert close-up'으로 손과 소주잔의 디테일만 잡도록 지시했으나, 인물들의 상반신과 얼굴이 넓게 프레임에 포함됨.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷(Previous Shot Still)에서 확정된 은색 금속 사각형 테이블이 아닌, 원목 재질의 둥근 테이블로 렌더링됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "프롬프트 상 상단 좌측(upper left)은 전택수(경찰 재킷, 모자)가 위치해야 하나, 이전 샷에서 비니를 썼던 서의용의 가죽 재킷을 입은 인물(비니 미착용)이 배치됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "상단 우측(upper right)은 나상혁(네이비 보머 재킷)이 위치해야 하나, 레퍼런스에 없는 어두운 무지 스웨터를 입은 인물이 배치됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "하단 우측(lower right)은 서의용(가죽 재킷)이 위치해야 하나, 나상혁의 의상인 네이비 재킷과 하늘색 셔츠 소매를 입은 팔이 배치됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "벽면에 걸려 있어야 할 '광어 우럭 매운탕' 메뉴판이 식당 벽이 아닌 상단 좌측 인물 바로 뒤 낮은 위치에 비정상적으로 놓여 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "테이블이 이전 장면의 모서리 스테인리스 식탁이 아니라 원형 나무 식탁이다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 상단 인물이 나상혁 참조의 해사한 얼굴·단정한 머리가 아니라 다른 얼굴이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "좌측 하단 인물이 강주철 참조의 둥근 얼굴이 아니라 다른 얼굴이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "카메라가 테이블 중앙 직상방의 가파른 하향 클로즈업이 아니라 네 사람의 상반신까지 보이는 넓은 구도이다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 상단 소매가 나상혁의 네이비 집업 재킷이 아니라 회색 니트이다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 하단 소매가 나상혁의 재킷이 아니라 흰 셔츠 커프가 보이는 네이비 소매이다.",
     "severity": "major"
    },
    {
     "issue_ko": "좌측 하단 의상이 강주철 참조의 주황·회색 등산 재킷이 아니라 녹·주황 배색 재킷이다.",
     "severity": "major"
    },
    {
     "issue_ko": "네 손이 지정된 모서리(전택수 좌상·서의용 우하·강주철 좌하·나상혁 우상)에서 들어오지 않는다.",
     "severity": "major"
    },
    {
     "issue_ko": "메뉴판 글자가 벽면이 아니라 테이블 뒤 액자형 간판에 있다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 6,
    "openrouter:x-ai/grok-4.6": 9
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "샷 텍스트와 프레이밍 규정에서 'insert close-up'으로 손과 소주잔의 디테일만 잡도록 지시했으나, 인물들의 상반신과 얼굴이 넓게 프레임에 포함됨.",
    "fix_en": "Apply a heavy vignette and blur to the upper bodies and faces to artificially focus attention solely on the hands and glasses in the center.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "벽면에 걸려 있어야 할 '광어 우럭 매운탕' 메뉴판이 식당 벽이 아닌 상단 좌측 인물 바로 뒤 낮은 위치에 비정상적으로 놓여 있음.",
    "fix_en": "Remove the freestanding menu sign behind the upper-left person and place a flat menu sign on the background wall, preserving the characters and lighting.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "좌측 하단 의상이 강주철 참조의 주황·회색 등산 재킷이 아니라 녹·주황 배색 재킷이다.",
    "fix_en": "Change the lower-left person's jacket to an orange and dark grey hiking jacket matching 강주철, preserving his pose, the table, and the rest of the scene.",
    "severity": "major",
    "observation_index": 12,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 6,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the wooden texture of the table surface with a brushed stainless steel texture, maintaining its current round shape, the hands, and glasses.\n- Change the upper-left person's clothing to a navy blue police jacket over a plaid shirt and add a navy cap, matching 전택수, preserving his pose, the table, and lighting.\n- Change the upper-right person's clothing to a navy bomber jacket over a light blue collared shirt, matching 나상혁, preserving his pose, the other people, and the background.\n- Change the lower-right arm's clothing to a black leather jacket over a dark grey hoodie, matching 서의용, preserving the hand's pose, the glass, and the rest of the scene.\n- Replace the face of the upper-right person with the youthful face and neat short black hair matching 나상혁, preserving his gaze, pose, and the lighting.\n- Replace the face of the lower-left person with the round face and short black hair matching 강주철, preserving his gaze, pose, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "하이 앵글과 소주잔을 부딪치는 핵심 액션은 시도했으나, 전택수가 누락되고 인서트 클로즈업 프레이밍 지시를 어겼으며 테이블 설정이 잘못되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "지시된 하이 앵글 인서트 프레이밍과 건배 액션을 완전히 무시했으며, 강주철을 누락하고 엉뚱한 인물을 추가하여 지침을 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "4개의 팔이 프레임 가장자리에서 중앙으로 모여 소주잔을 맞부딪치고 있음.",
      "built_space": "이전 샷의 사각 철제 테이블이 아닌 둥근 원목 테이블이 배치됨. 메뉴판이 배경 바닥에 부자연스럽게 세워져 있음.",
      "entities": "전택수 누락. 좌측 상단은 서의용(비니 누락), 좌측 하단은 강주철(비니 누락), 우측 상단은 나상혁 얼굴에 프롬프트에 없는 스웨터를 착용함. 우측 하단에서는 나상혁의 원래 의상(네이비 재킷과 셔츠)을 입은 4번째 팔만 등장함.",
      "hard_violations": [
       "필수 인물(전택수) 누락",
       "인물과 의상 매칭 오류(나상혁의 얼굴과 의상이 분리됨)",
       "이전 샷에 고정된 테이블 재질 및 형태(사각 철제) 위반"
      ],
      "physics": "4개의 손이 공중에서 소주잔을 쥐고 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "4명의 인물이 테이블에 앉아 정면 카메라를 응시하며, 잔을 부딪치지 않음.",
      "built_space": "둥근 철제 테이블 주변으로 인물들이 앉아 있음.",
      "entities": "강주철 누락. 좌측 끝에 프롬프트에 없는 새로운 인물이 경찰 점퍼를 입고 등장함. 전택수, 나상혁, 서의용(후드끈 누락)은 올바르게 착석해 있음.",
      "hard_violations": [
       "필수 인물(강주철) 누락",
       "프롬프트에 없는 새로운 인물 추가",
       "하이 앵글 인서트 클로즈업 프레이밍 완전 무시 (아이 레벨 미디엄 샷으로 렌더링)",
       "핵심 액션(소주잔을 한데 모아 겹치는 순간) 누락"
      ],
      "physics": "소주잔과 손 모두 테이블 상판 위에 놓여 지탱됨."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "하이 앵글과 소주잔을 부딪치는 핵심 액션은 시도했으나, 전택수가 누락되고 인서트 클로즈업 프레이밍 지시를 어겼으며 테이블 설정이 잘못되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "지시된 하이 앵글 인서트 프레이밍과 건배 액션을 완전히 무시했으며, 강주철을 누락하고 엉뚱한 인물을 추가하여 지침을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "4개의 팔이 프레임 가장자리에서 중앙으로 모여 소주잔을 맞부딪치고 있음.",
      "built_space": "이전 샷의 사각 철제 테이블이 아닌 둥근 원목 테이블이 배치됨. 메뉴판이 배경 바닥에 부자연스럽게 세워져 있음.",
      "entities": "전택수 누락. 좌측 상단은 서의용(비니 누락), 좌측 하단은 강주철(비니 누락), 우측 상단은 나상혁 얼굴에 프롬프트에 없는 스웨터를 착용함. 우측 하단에서는 나상혁의 원래 의상(네이비 재킷과 셔츠)을 입은 4번째 팔만 등장함.",
      "hard_violations": [
       "필수 인물(전택수) 누락",
       "인물과 의상 매칭 오류(나상혁의 얼굴과 의상이 분리됨)",
       "이전 샷에 고정된 테이블 재질 및 형태(사각 철제) 위반"
      ],
      "physics": "4개의 손이 공중에서 소주잔을 쥐고 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "4명의 인물이 테이블에 앉아 정면 카메라를 응시하며, 잔을 부딪치지 않음.",
      "built_space": "둥근 철제 테이블 주변으로 인물들이 앉아 있음.",
      "entities": "강주철 누락. 좌측 끝에 프롬프트에 없는 새로운 인물이 경찰 점퍼를 입고 등장함. 전택수, 나상혁, 서의용(후드끈 누락)은 올바르게 착석해 있음.",
      "hard_violations": [
       "필수 인물(강주철) 누락",
       "프롬프트에 없는 새로운 인물 추가",
       "하이 앵글 인서트 클로즈업 프레이밍 완전 무시 (아이 레벨 미디엄 샷으로 렌더링)",
       "핵심 액션(소주잔을 한데 모아 겹치는 순간) 누락"
      ],
      "physics": "소주잔과 손 모두 테이블 상판 위에 놓여 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "하이 앵글과 건배 액션, 메뉴판 텍스트를 준수했으나 지정된 인물 배치와 테이블의 연속성을 위반함."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "요구된 하이 앵글, 인서트 구도, 핵심 액션(건배)을 모두 무시하고 정면 응시를 연출하여 실패함."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "네 손이 중앙으로 모여 소주잔을 부딪치고 있으며, 인물들의 시선도 중앙을 향함.",
      "built_space": "원형 나무 테이블. 배경에 세로형 메뉴판이 보임.",
      "entities": "좌측 상단 서의용, 좌측 하단 강주철, 우측 상단 나상혁(의상 불일치). 우측 하단은 팔만 보이며 전택수는 식별 불가. 메뉴판의 '광어 우럭 매운탕' 텍스트는 정확함.",
      "hard_violations": [
       "프롬프트가 지정한 인물별 화면 내 위치 불일치 (예: 서의용이 우하단이 아닌 좌상단에 위치).",
       "레퍼런스의 사각형 금속 테이블이 아닌 원형 나무 테이블 사용 (연속성 위반)."
      ],
      "physics": "공중에서 손이 소주잔을 쥐고 서로 맞대어 지지함."
     },
     {
      "label": "A",
      "direction": "네 명의 인물이 모두 카메라 렌즈를 정면으로 응시. 잔들은 흩어져 있음.",
      "built_space": "원형 금속 테이블이 중심에 있음.",
      "entities": "가장 좌측에 신원 불상의 인물이 경찰복 착용. 그 외 전택수, 나상혁, 서의용 식별됨. 소주잔 4개가 떨어져 있음.",
      "hard_violations": [
       "명시된 수직 하이 앵글 카메라 위치 무시 (정면 아이레벨 연출).",
       "인서트 클로즈업 프레이밍 무시 (인물 상반신이 모두 나오는 미디엄 숏).",
       "잔을 중앙에 모아 부딪치는 핵심 액션 누락."
      ],
      "physics": "손과 잔이 모두 테이블 상판에 놓여 지지됨."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "하이 앵글과 건배 액션, 메뉴판 텍스트를 준수했으나 지정된 인물 배치와 테이블의 연속성을 위반함."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "요구된 하이 앵글, 인서트 구도, 핵심 액션(건배)을 모두 무시하고 정면 응시를 연출하여 실패함."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "네 손이 중앙으로 모여 소주잔을 부딪치고 있으며, 인물들의 시선도 중앙을 향함.",
      "built_space": "원형 나무 테이블. 배경에 세로형 메뉴판이 보임.",
      "entities": "좌측 상단 서의용, 좌측 하단 강주철, 우측 상단 나상혁(의상 불일치). 우측 하단은 팔만 보이며 전택수는 식별 불가. 메뉴판의 '광어 우럭 매운탕' 텍스트는 정확함.",
      "hard_violations": [
       "프롬프트가 지정한 인물별 화면 내 위치 불일치 (예: 서의용이 우하단이 아닌 좌상단에 위치).",
       "레퍼런스의 사각형 금속 테이블이 아닌 원형 나무 테이블 사용 (연속성 위반)."
      ],
      "physics": "공중에서 손이 소주잔을 쥐고 서로 맞대어 지지함."
     },
     {
      "label": "B",
      "direction": "네 명의 인물이 모두 카메라 렌즈를 정면으로 응시. 잔들은 흩어져 있음.",
      "built_space": "원형 금속 테이블이 중심에 있음.",
      "entities": "가장 좌측에 신원 불상의 인물이 경찰복 착용. 그 외 전택수, 나상혁, 서의용 식별됨. 소주잔 4개가 떨어져 있음.",
      "hard_violations": [
       "명시된 수직 하이 앵글 카메라 위치 무시 (정면 아이레벨 연출).",
       "인서트 클로즈업 프레이밍 무시 (인물 상반신이 모두 나오는 미디엄 숏).",
       "잔을 중앙에 모아 부딪치는 핵심 액션 누락."
      ],
      "physics": "손과 잔이 모두 테이블 상판에 놓여 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S23sh9"
  }
 },
 "S23sh11::cine": {
  "applied": true,
  "fingerprint": "f310f4021cb84062aca28d388c2605914306fcc26429c2e76e18eadda2e9f8c2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S23sh11_sel.png",
  "source_sha256": "5a2a479feb324224f9827582c52da67b6df0983688c4426b6f30693d5a655d4d",
  "file": "S23sh11_cine.png",
  "latency_ms": 10850
 },
 "S24sh4::signage": {
  "fp": "92997dd62b945a0a",
  "inscriptions": [
   {
    "surface_native": "수사 서류 표지",
    "text_native": "수사보고",
    "reason_ko": "강력반 형사의 책상 위에 부검 사진과 함께 놓여 있는 서류 봉투의 현실감을 높이기 위해 '수사보고'라는 표기가 필요합니다."
   }
  ]
 },
 "S24sh4": {
  "input_fingerprint": "70b6818d6427a768",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상 옆에 서서 책상 위의 부검 사진들을 향해 시선을 던지는 나상혁의 모습\n\nLOCATION (lock): Beside the detective's cluttered desk in the violent-crimes office, where copied autopsy photographs are spread out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At standing chest height on the desk’s opposite diagonal, the camera completes its track into a medium view of 나상혁 at frame right, with the piled records and indistinct autopsy photocopies extending across the lower-left foreground. He shifts his weight beside the desk and glances down at the photographs, while the composition leaves space beyond his near shoulder for the coming arc toward 서의용.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 의용의 책상 (Covered with records and autopsy photocopies) — Its near edge angles from the lower-left foreground toward 나상혁 at frame right; used as The desk creates a diagonal leading line from the photographic evidence toward 나상혁; 부검 사진 복사본 (Laid on the desk as unclear photocopies) — Their image-bearing faces lie upward toward the camera, but the copied images are darkened and indistinct; used as Visible evidence in the lower foreground that motivates 나상혁's downward glance; 사건 기록들 (Piled high on the desk) — Pages and covers face upward at mixed angles across the desk; used as Provides dense procedural context behind the photographs without obscuring them.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient office illumination is rendered in a subdued neutral-to-cool palette with moderate contrast and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same busy detective-office layout, worn desks, doorway, and daytime fluorescent light. Exclude the senior officer standing in the doorway and retain only the younger investigator beside the autopsy-photo desk.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dark, indistinct photocopied autopsy photographs remain spread on Euiyong's desk as Sanghyeok examines them.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사 서류 표지: \"수사보고\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상 옆에 서서 책상 위의 부검 사진들을 향해 시선을 던지는 나상혁의 모습\n\nLOCATION (lock): Beside the detective's cluttered desk in the violent-crimes office, where copied autopsy photographs are spread out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At standing chest height on the desk’s opposite diagonal, the camera completes its track into a medium view of 나상혁 at frame right, with the piled records and indistinct autopsy photocopies extending across the lower-left foreground. He shifts his weight beside the desk and glances down at the photographs, while the composition leaves space beyond his near shoulder for the coming arc toward 서의용.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 의용의 책상 (Covered with records and autopsy photocopies) — Its near edge angles from the lower-left foreground toward 나상혁 at frame right; used as The desk creates a diagonal leading line from the photographic evidence toward 나상혁; 부검 사진 복사본 (Laid on the desk as unclear photocopies) — Their image-bearing faces lie upward toward the camera, but the copied images are darkened and indistinct; used as Visible evidence in the lower foreground that motivates 나상혁's downward glance; 사건 기록들 (Piled high on the desk) — Pages and covers face upward at mixed angles across the desk; used as Provides dense procedural context behind the photographs without obscuring them.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient office illumination is rendered in a subdued neutral-to-cool palette with moderate contrast and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same busy detective-office layout, worn desks, doorway, and daytime fluorescent light. Exclude the senior officer standing in the doorway and retain only the younger investigator beside the autopsy-photo desk.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dark, indistinct photocopied autopsy photographs remain spread on Euiyong's desk as Sanghyeok examines them.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사 서류 표지: \"수사보고\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상 옆에 서서 책상 위의 부검 사진들을 향해 시선을 던지는 나상혁의 모습\n\nLOCATION (lock): Beside the detective's cluttered desk in the violent-crimes office, where copied autopsy photographs are spread out. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At standing chest height on the desk’s opposite diagonal, the camera completes its track into a medium view of 나상혁 at frame right, with the piled records and indistinct autopsy photocopies extending across the lower-left foreground. He shifts his weight beside the desk and glances down at the photographs, while the composition leaves space beyond his near shoulder for the coming arc toward 서의용.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 의용의 책상 (Covered with records and autopsy photocopies) — Its near edge angles from the lower-left foreground toward 나상혁 at frame right; used as The desk creates a diagonal leading line from the photographic evidence toward 나상혁; 부검 사진 복사본 (Laid on the desk as unclear photocopies) — Their image-bearing faces lie upward toward the camera, but the copied images are darkened and indistinct; used as Visible evidence in the lower foreground that motivates 나상혁's downward glance; 사건 기록들 (Piled high on the desk) — Pages and covers face upward at mixed angles across the desk; used as Provides dense procedural context behind the photographs without obscuring them.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient office illumination is rendered in a subdued neutral-to-cool palette with moderate contrast and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same busy detective-office layout, worn desks, doorway, and daytime fluorescent light. Exclude the senior officer standing in the doorway and retain only the younger investigator beside the autopsy-photo desk.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dark, indistinct photocopied autopsy photographs remain spread on Euiyong's desk as Sanghyeok examines them.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사 서류 표지: \"수사보고\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "나상혁의 시선이 책상 위에 놓인 부검 사진 복사본들을 향해 아래로 내려다보고 있음.",
    "built_space": "형사과 사무실의 열린 문, 캐비닛, 컴퓨터 모니터와 복합기가 있는 책상 등 이전 샷의 배경 요소들이 프레임 내에 정확한 위치와 스케일로 배치되어 있음.",
    "entities": "나상혁(30대 한국 남성, 단정한 검은 머리, 네이비 재킷, 하늘색 셔츠)이 정확히 구현되었으며, 책상 위에는 '수사보고'라고 적힌 서류와 흑백으로 흐릿하게 복사된 부검 사진들이 놓여 있음.",
    "hard_violations": [],
    "physics": "나상혁은 두 발로 바닥을 딛고 자연스럽게 서 있으며, 책상 위의 서류와 사진들도 표면 위에 안정적으로 놓여 있어 물리적 오류가 없음."
   },
   {
    "label": "B",
    "direction": "나상혁의 시선이 책상 위의 사진들을 향해 아래로 향하고 있음.",
    "built_space": "이전 샷과 동일한 형사과 사무실의 캐비닛, 열린 문, 복합기 등의 요소가 프레임 안에 적절하게 배치되어 있음.",
    "entities": "나상혁의 외형과 복장이 레퍼런스와 일치하며 '수사보고' 텍스트도 정확히 표기되었으나, 부검 사진 중 일부가 '희미한 흑백 복사본'이 아닌 컬러 사진으로 등장함.",
    "hard_violations": [],
    "physics": "인물이 땅에 안정적으로 서 있으며, 책상 위 소품들의 배치나 중력 적용에 있어 어색함이 없음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 10,
   "B": 8
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 10,
    "verdict_ko": "프롬프트가 요구한 '어둡고 희미한 부검 사진 복사본'의 디테일을 정확히 묘사하였으며, 카메라 구도와 배경의 연속성도 완벽하게 구현되었습니다."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "구도와 배경, 인물의 외형은 훌륭하나, 책상 위의 부검 사진이 흑백 복사본이 아닌 원본 컬러 사진으로 묘사되어 프롬프트의 지시와 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S22sh10_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   },
   {
    "label": "PROP REFERENCE — 피해자 부검 사진 및 복사본: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1017104>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "캐릭터의 재킷 하단 중앙 바지 부위에 옷 구조와 분리된 정체불명의 지퍼 부속품이 잘못 생성되었습니다.",
     "fix_en": "Remove the stray zipper pull hanging off the crotch area of the grey trousers, replacing it with smooth grey fabric matching the pants. Keep the character's face, jacket, shirt, and background exactly as they are.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 숏 레퍼런스에는 존재하지 않았던 종이 상자가 화면 중앙 문 왼쪽의 회색 캐비닛 위에 임의로 추가되었습니다.",
     "fix_en": "Remove the brown cardboard box from the top of the grey cabinet next to the doorway, leaving the cabinet's top surface completely bare. Keep the room layout, lighting, character, and desk exactly as they are.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "화면 좌측 책상 위의 서류 더미 단면이 낱장 종이가 아닌 통짜 블록처럼 뭉개졌으며, 모니터 아래 키보드 형태가 심하게 왜곡되었습니다.",
     "fix_en": "Redraw the edges of the paper stacks on the left desk to show distinct individual sheets rather than solid blocks, and straighten the black keyboard to a normal rectangle with clear keys. Keep the autopsy photos, the room, and the character exactly as they are.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "오른쪽 나상혁이 양팔을 늘어뜨리고 양발에 균등히 선 채 경직되어, 체중을 옮기며 책상 사진을 보는 중간 동작이 아니다.",
     "fix_en": "Redraw the character to shift his weight onto one leg with a more relaxed, asymmetric stance, such as slightly bending one knee or engaging an arm, rather than standing stiffly at attention. Keep his identity, clothing, the desk items, and framing exactly as they are.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "캐릭터의 재킷 하단 중앙 바지 부위에 옷 구조와 분리된 정체불명의 지퍼 부속품이 잘못 생성되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 숏 레퍼런스에는 존재하지 않았던 종이 상자가 화면 중앙 문 왼쪽의 회색 캐비닛 위에 임의로 추가되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 좌측 책상 위의 서류 더미 단면이 낱장 종이가 아닌 통짜 블록처럼 뭉개졌으며, 모니터 아래 키보드 형태가 심하게 왜곡되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "미디엄 샷이어야 하는데 나상혁의 발끝과 사무실 전경이 다 보이는 풀샷·와이드 구도이다.",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 나상혁이 양팔을 늘어뜨리고 양발에 균등히 선 채 경직되어, 체중을 옮기며 책상 사진을 보는 중간 동작이 아니다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "캐릭터의 재킷 하단 중앙 바지 부위에 옷 구조와 분리된 정체불명의 지퍼 부속품이 잘못 생성되었습니다.",
    "fix_en": "Remove the stray zipper pull hanging off the crotch area of the grey trousers, replacing it with smooth grey fabric matching the pants. Keep the character's face, jacket, shirt, and background exactly as they are.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "이전 숏 레퍼런스에는 존재하지 않았던 종이 상자가 화면 중앙 문 왼쪽의 회색 캐비닛 위에 임의로 추가되었습니다.",
    "fix_en": "Remove the brown cardboard box from the top of the grey cabinet next to the doorway, leaving the cabinet's top surface completely bare. Keep the room layout, lighting, character, and desk exactly as they are.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "화면 좌측 책상 위의 서류 더미 단면이 낱장 종이가 아닌 통짜 블록처럼 뭉개졌으며, 모니터 아래 키보드 형태가 심하게 왜곡되었습니다.",
    "fix_en": "Redraw the edges of the paper stacks on the left desk to show distinct individual sheets rather than solid blocks, and straighten the black keyboard to a normal rectangle with clear keys. Keep the autopsy photos, the room, and the character exactly as they are.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "오른쪽 나상혁이 양팔을 늘어뜨리고 양발에 균등히 선 채 경직되어, 체중을 옮기며 책상 사진을 보는 중간 동작이 아니다.",
    "fix_en": "Redraw the character to shift his weight onto one leg with a more relaxed, asymmetric stance, such as slightly bending one knee or engaging an arm, rather than standing stiffly at attention. Keep his identity, clothing, the desk items, and framing exactly as they are.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S22sh10"
  }
 },
 "S24sh4::cine": {
  "applied": true,
  "fingerprint": "81ce24ff7ce4c47547dd84a51d1d6896e8bd4c23282baea07442443cb494992a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S24sh4_sel.png",
  "source_sha256": "36be4b1e8de96ae32881c08fccb666fe81f243ba9802ab8b01aa39e0a144a255",
  "file": "S24sh4_cine.png",
  "latency_ms": 10103
 },
 "S24sh5::signage": {
  "fp": "4ef3d6c2450ed4f6",
  "inscriptions": [
   {
    "surface_native": "사무실 벽면 부서 안내판",
    "text_native": "강력1반",
    "reason_ko": "강력계 형사들이 근무하는 사무실 내부라는 공간적 배경을 사실적으로 묘사하기 위해 부서 안내판이 필요합니다."
   }
  ]
 },
 "S24sh5": {
  "input_fingerprint": "bc0e23539b4e6f11",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 의자에 앉은 채 나상혁 쪽으로 고개를 돌린 자세로 눈을 크게 뜬 서의용\n\nLOCATION (lock): At the detective's desk and chair inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind and outside 나상혁's near shoulder at standing shoulder height, hold a slightly elevated close over-the-shoulder view toward the seated 서의용. 서의용 occupies the center-left in three-quarter view, twisting toward 나상혁 with widened eyes, while 나상혁's soft shoulder edge remains at frame right as the sole spatial reference.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 의용의 책상 (Covered with case records) — Only its far corner and upper edge enter the lower frame beneath 서의용; used as A narrow lower-frame reminder of the records that prompted the exchange; 사무용 의자 (Occupied by 서의용) — Its seat faces the desk while 서의용 rotates his upper body toward frame right; used as Supports 서의용's arrested seated turn toward 나상혁.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same subdued daytime office ambience holds neutral-to-cool tones and restrained contrast on 서의용's startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same desk area, office materials, fluorescent lighting, and the standing investigator's clothing. Exclude the autopsy-photo-focused pose and show the seated detective turning toward him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The indistinct autopsy-photo copies remain laid out on the desk between Euiyong and Sanghyeok.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 부서 안내판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 의자에 앉은 채 나상혁 쪽으로 고개를 돌린 자세로 눈을 크게 뜬 서의용\n\nLOCATION (lock): At the detective's desk and chair inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind and outside 나상혁's near shoulder at standing shoulder height, hold a slightly elevated close over-the-shoulder view toward the seated 서의용. 서의용 occupies the center-left in three-quarter view, twisting toward 나상혁 with widened eyes, while 나상혁's soft shoulder edge remains at frame right as the sole spatial reference.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 의용의 책상 (Covered with case records) — Only its far corner and upper edge enter the lower frame beneath 서의용; used as A narrow lower-frame reminder of the records that prompted the exchange; 사무용 의자 (Occupied by 서의용) — Its seat faces the desk while 서의용 rotates his upper body toward frame right; used as Supports 서의용's arrested seated turn toward 나상혁.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same subdued daytime office ambience holds neutral-to-cool tones and restrained contrast on 서의용's startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same desk area, office materials, fluorescent lighting, and the standing investigator's clothing. Exclude the autopsy-photo-focused pose and show the seated detective turning toward him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The indistinct autopsy-photo copies remain laid out on the desk between Euiyong and Sanghyeok.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 부서 안내판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 의자에 앉은 채 나상혁 쪽으로 고개를 돌린 자세로 눈을 크게 뜬 서의용\n\nLOCATION (lock): At the detective's desk and chair inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind and outside 나상혁's near shoulder at standing shoulder height, hold a slightly elevated close over-the-shoulder view toward the seated 서의용. 서의용 occupies the center-left in three-quarter view, twisting toward 나상혁 with widened eyes, while 나상혁's soft shoulder edge remains at frame right as the sole spatial reference.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 의용의 책상 (Covered with case records) — Only its far corner and upper edge enter the lower frame beneath 서의용; used as A narrow lower-frame reminder of the records that prompted the exchange; 사무용 의자 (Occupied by 서의용) — Its seat faces the desk while 서의용 rotates his upper body toward frame right; used as Supports 서의용's arrested seated turn toward 나상혁.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same subdued daytime office ambience holds neutral-to-cool tones and restrained contrast on 서의용's startled expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same desk area, office materials, fluorescent lighting, and the standing investigator's clothing. Exclude the autopsy-photo-focused pose and show the seated detective turning toward him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The indistinct autopsy-photo copies remain laid out on the desk between Euiyong and Sanghyeok.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 벽면 부서 안내판: \"강력1반\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "서의용의 시선이 화면 우측 전경에 위치한 나상혁의 얼굴과 어깨 쪽을 정확히 향하고 있습니다.",
    "built_space": "사무실 내부 책상, 모니터, 캐비닛, 열린 문의 위치가 이전 샷과 정확히 일치하며, 서의용이 단일 사무용 의자에 바르게 착석해 있습니다.",
    "entities": "서의용(짧은 검은 머리)이 앉아 있으나 캐릭터 레퍼런스의 복장(비니, 가죽재킷)이 전경 인물에게 잘못 입혀졌습니다. 벽면에 '강력1반' 안내판과 책상 위 4장의 부검 사진이 정확히 존재합니다.",
    "hard_violations": [],
    "physics": "서의용은 의자에 자연스럽게 앉아 체중을 지탱하고 있으며, 책상 위 사물들도 바닥면에 정상적으로 밀착되어 있습니다."
   },
   {
    "label": "B",
    "direction": "서의용의 시선이 우측 전경의 인물을 향해 자연스럽게 이어지고 있습니다.",
    "built_space": "이전 샷과 유사한 사무실 배경과 책상 배치를 유지하며, 인물이 의자에 위치해 있습니다.",
    "entities": "서의용이 정장 재킷을 입고 있으며 레퍼런스의 의상이 전경 인물에게 적용된 오류가 동일하게 나타납니다. '강력1반' 안내판은 정확하나, 부검 사진들에 임의의 흰색 테두리가 추가되었습니다.",
    "hard_violations": [],
    "physics": "인물이 의자에 안정적으로 착석해 있고, 모든 사물이 지지면 위에 물리적 오류 없이 놓여 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 6,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "지정된 숄더뷰 화각, 배경의 디테일, 놀란 표정 및 간판 텍스트('강력1반')를 훌륭히 구현했으나, 레퍼런스의 의상(비니, 가죽재킷)이 서의용이 아닌 전경 인물에게 잘못 적용되었습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "화각과 간판 텍스트는 잘 반영되었으나, A에 비해 놀란 표정의 강도가 약하고 책상 위 부검 사진의 디테일이 기준 이미지와 다소 차이가 있으며 동일한 의상 오류가 존재합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S24sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "의자에 앉아 있는 인물(서의용)의 얼굴이 캐릭터 레퍼런스의 얼굴과 일치하지 않는 전혀 다른 인물입니다.",
     "fix_en": "Replace the seated man's face with the reference wide-faced Korean man in his mid-40s. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "의자에 앉아 있는 인물(서의용)이 레퍼런스 의상(비니, 후드, 가죽 재킷) 대신 셔츠와 넥타이를 입고 있습니다.",
     "fix_en": "Change the seated man's clothing to a dark beanie, grey hoodie, and black leather jacket. Preserve the people present and their positions, the foreground character's clothing, the set, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "우측 전경에 걸쳐진 인물(나상혁)이 프롬프트가 지시한 이전 샷 스틸의 의상(남색 재킷)이 아닌, 서의용의 레퍼런스 의상(비니, 후드)을 입고 있어 인물의 외형이 뒤바뀌었습니다.",
     "fix_en": "Change the right foreground character's clothing to a light shirt and navy jacket, replacing the beanie with bare dark hair. Preserve the people present and their positions, the seated man's clothing, the set, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "나상혁이 오른쪽 어깨 가장자리만이 아니라 머리와 등 전체로 크게 들어와 있다",
     "fix_en": "Darken the foreground character to reduce his visual footprint. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    },
    {
     "issue_ko": "책상이 먼 모서리·상단만 들어와야 하는데 모니터·프린터·서류 더미까지 하단을 넓게 차지한다",
     "fix_en": "Blur the monitor and lower desk area to reduce their prominence. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.",
     "severity": "major",
     "observation_index": 6,
     "needs_regeneration": true
    },
    {
     "issue_ko": "클로즈업이 아니라 상반신과 사무실 배경이 많이 보이는 미디엄 샷이다",
     "fix_en": "Apply a vignette to focus on the seated man's face. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.",
     "severity": "major",
     "observation_index": 7,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "의자에 앉아 있는 인물(서의용)의 얼굴이 캐릭터 레퍼런스의 얼굴과 일치하지 않는 전혀 다른 인물입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "의자에 앉아 있는 인물(서의용)이 레퍼런스 의상(비니, 후드, 가죽 재킷) 대신 셔츠와 넥타이를 입고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "우측 전경에 걸쳐진 인물(나상혁)이 프롬프트가 지시한 이전 샷 스틸의 의상(남색 재킷)이 아닌, 서의용의 레퍼런스 의상(비니, 후드)을 입고 있어 인물의 외형이 뒤바뀌었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 중앙에 앉은 남자가 서의용 레퍼런스와 얼굴·체형·의상이 전혀 다른 인물이다",
     "severity": "critical"
    },
    {
     "issue_ko": "오른쪽 오버숄더 인물이 나상혁의 남색 점퍼가 아니라 서의용 레퍼런스의 비니·가죽재킷·후드를 입고 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "나상혁이 오른쪽 어깨 가장자리만이 아니라 머리와 등 전체로 크게 들어와 있다",
     "severity": "major"
    },
    {
     "issue_ko": "책상이 먼 모서리·상단만 들어와야 하는데 모니터·프린터·서류 더미까지 하단을 넓게 차지한다",
     "severity": "major"
    },
    {
     "issue_ko": "클로즈업이 아니라 상반신과 사무실 배경이 많이 보이는 미디엄 샷이다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "나상혁이 오른쪽 어깨 가장자리만이 아니라 머리와 등 전체로 크게 들어와 있다",
    "fix_en": "Darken the foreground character to reduce his visual footprint. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   },
   {
    "issue_ko": "책상이 먼 모서리·상단만 들어와야 하는데 모니터·프린터·서류 더미까지 하단을 넓게 차지한다",
    "fix_en": "Blur the monitor and lower desk area to reduce their prominence. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.",
    "severity": "major",
    "observation_index": 6,
    "needs_regeneration": true
   },
   {
    "issue_ko": "클로즈업이 아니라 상반신과 사무실 배경이 많이 보이는 미디엄 샷이다",
    "fix_en": "Apply a vignette to focus on the seated man's face. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.",
    "severity": "major",
    "observation_index": 7,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the seated man's face with the reference wide-faced Korean man in his mid-40s. Preserve the people present and their positions, their clothing, the set, the lighting, and the framing.\n- Change the seated man's clothing to a dark beanie, grey hoodie, and black leather jacket. Preserve the people present and their positions, the foreground character's clothing, the set, the lighting, and the framing.\n- Change the right foreground character's clothing to a light shirt and navy jacket, replacing the beanie with bare dark hair. Preserve the people present and their positions, the seated man's clothing, the set, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "지정된 오버더숄더 클로즈업 구도, 눈을 크게 뜬 표정 연기, '강력1반' 간판을 완벽히 구현했으나 두 인물의 의상이 서로 뒤바뀐 점이 아쉽습니다.  ★위반: [openrouter:x-ai/grok-4.6] 앉은 인물이 서의용이 아닌 다른 얼굴·의상의 창작 인물이다 / [openrouter:x-ai/grok-4.6] 서의용 의상이 나상혁이어야 할 오른쪽 어깨에 잘못 올라갔다"
     },
     {
      "label": "B",
      "score": 536,
      "verdict_ko": "서의용의 의상은 레퍼런스와 일치하나, 프레이밍 지시를 어기고 금지된 인물을 전신으로 노출시켰으며 인물이 중복되는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 중복되거나 추가된 인물 (서 있는 인물과 전경의 어깨가 동시에 존재함) / [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 전신 인물과 세 번째 인물을 추가했다 / [openrouter:x-ai/grok-4.6] 이전 스틸 인물의 얼굴과 전신을 이 샷에 그대로 반입했다"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.286
     },
     "adjusted": {
      "A": 1.0,
      "B": 0.536
     },
     "violations": {
      "B": [
       "[gemini-pro] 중복되거나 추가된 인물 (서 있는 인물과 전경의 어깨가 동시에 존재함)",
       "[openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 전신 인물과 세 번째 인물을 추가했다",
       "[openrouter:x-ai/grok-4.6] 이전 스틸 인물의 얼굴과 전신을 이 샷에 그대로 반입했다"
      ],
      "A": [
       "[openrouter:x-ai/grok-4.6] 앉은 인물이 서의용이 아닌 다른 얼굴·의상의 창작 인물이다",
       "[openrouter:x-ai/grok-4.6] 서의용 의상이 나상혁이어야 할 오른쪽 어깨에 잘못 올라갔다"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "지정된 오버더숄더 클로즈업 구도, 눈을 크게 뜬 표정 연기, '강력1반' 간판을 완벽히 구현했으나 두 인물의 의상이 서로 뒤바뀐 점이 아쉽습니다.  ★위반: [openrouter:x-ai/grok-4.6] 앉은 인물이 서의용이 아닌 다른 얼굴·의상의 창작 인물이다 / [openrouter:x-ai/grok-4.6] 서의용 의상이 나상혁이어야 할 오른쪽 어깨에 잘못 올라갔다"
     },
     {
      "label": "B",
      "score": 536,
      "verdict_ko": "서의용의 의상은 레퍼런스와 일치하나, 프레이밍 지시를 어기고 금지된 인물을 전신으로 노출시켰으며 인물이 중복되는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 중복되거나 추가된 인물 (서 있는 인물과 전경의 어깨가 동시에 존재함) / [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 전신 인물과 세 번째 인물을 추가했다 / [openrouter:x-ai/grok-4.6] 이전 스틸 인물의 얼굴과 전신을 이 샷에 그대로 반입했다"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "서의용의 신원과 복장은 레퍼런스를 유지했으나, 프레이밍이 와이드 샷으로 잘못되었고 샷 텍스트에 없는 인물(이전 컷의 인물)이 버젓이 등장하여 실패한 결과물입니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "오버더숄더 프레이밍과 안내판 텍스트(강력1반)는 지시를 따랐으나, 정작 앉아 있어야 할 서의용의 복장(비니, 가죽재킷)이 전경의 어깨로 가고 다른 인물이 앉아 있는 치명적인 신원 뒤바뀜이 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "앉아 있는 서의용은 화면 우측 전경의 어깨가 아닌, 화면 중앙에 서 있는 남자를 빤히 바라보고 있음.",
      "built_space": "화면 좌측에 수사 서류가 쌓인 책상이 있으며, 서의용이 그 앞 사무용 의자에 앉아 있음.",
      "entities": "서의용은 비니와 가죽 재킷을 착용해 캐릭터 레퍼런스와 일치함. 그러나 배제되어야 할 이전 컷의 인물(파란 재킷을 입은 남성)이 프레임 중앙에 온전히 서 있으며, 우측 전경에 세 번째 인물의 어깨가 걸쳐 있음. 벽의 안내판은 '강력1반'이 아닌 이전 컷의 '강력계'임. 책상 위 부검 사진들은 잘 놓여 있음.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(이전 컷의 서 있는 남성 및 전경의 추가 인물) 렌더링으로 인한 엑스트라 바디 추가",
       "클로즈업 및 지정된 오버더숄더 구도를 완전히 벗어난 와이드 샷 프레이밍"
      ],
      "physics": "앉은 서의용은 의자에, 서 있는 남성은 바닥에 정상적으로 지지되어 체중을 싣고 있음."
     },
     {
      "label": "B",
      "direction": "의자에 앉은 남자가 우측 전경에 위치한 인물(어깨) 쪽으로 몸을 돌려 눈을 크게 뜨고 쳐다봄.",
      "built_space": "좌측에 책상의 모서리와 서류들이 보이고, 남자가 사무용 의자에 앉아 있음.",
      "entities": "안내판 텍스트는 '강력1반'으로 완벽히 수정됨. 책상 위 서류와 부검 사진도 묘사됨. 그러나 중심 인물인 앉아 있는 남자는 서의용 레퍼런스와 전혀 다른 얼굴에 셔츠와 넥타이를 착용하고 있음. 반면 프레임 우측 전경의 어깨(나상혁 역이어야 함)가 서의용의 상징인 파란색 비니와 후드, 가죽 재킷을 입고 있음.",
      "hard_violations": [
       "인물 신원 및 복장 배치 오류 (지정된 서의용이 의자에 없고 엉뚱한 인물이 앉아 있으며, 서의용의 복장이 전경의 어깨로 뒤바뀜)"
      ],
      "physics": "앉은 인물은 의자에 등을 돌린 채 정상적으로 앉아 자세를 유지하고 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "서의용의 신원과 복장은 레퍼런스를 유지했으나, 프레이밍이 와이드 샷으로 잘못되었고 샷 텍스트에 없는 인물(이전 컷의 인물)이 버젓이 등장하여 실패한 결과물입니다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "오버더숄더 프레이밍과 안내판 텍스트(강력1반)는 지시를 따랐으나, 정작 앉아 있어야 할 서의용의 복장(비니, 가죽재킷)이 전경의 어깨로 가고 다른 인물이 앉아 있는 치명적인 신원 뒤바뀜이 발생했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "앉아 있는 서의용은 화면 우측 전경의 어깨가 아닌, 화면 중앙에 서 있는 남자를 빤히 바라보고 있음.",
      "built_space": "화면 좌측에 수사 서류가 쌓인 책상이 있으며, 서의용이 그 앞 사무용 의자에 앉아 있음.",
      "entities": "서의용은 비니와 가죽 재킷을 착용해 캐릭터 레퍼런스와 일치함. 그러나 배제되어야 할 이전 컷의 인물(파란 재킷을 입은 남성)이 프레임 중앙에 온전히 서 있으며, 우측 전경에 세 번째 인물의 어깨가 걸쳐 있음. 벽의 안내판은 '강력1반'이 아닌 이전 컷의 '강력계'임. 책상 위 부검 사진들은 잘 놓여 있음.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(이전 컷의 서 있는 남성 및 전경의 추가 인물) 렌더링으로 인한 엑스트라 바디 추가",
       "클로즈업 및 지정된 오버더숄더 구도를 완전히 벗어난 와이드 샷 프레이밍"
      ],
      "physics": "앉은 서의용은 의자에, 서 있는 남성은 바닥에 정상적으로 지지되어 체중을 싣고 있음."
     },
     {
      "label": "A",
      "direction": "의자에 앉은 남자가 우측 전경에 위치한 인물(어깨) 쪽으로 몸을 돌려 눈을 크게 뜨고 쳐다봄.",
      "built_space": "좌측에 책상의 모서리와 서류들이 보이고, 남자가 사무용 의자에 앉아 있음.",
      "entities": "안내판 텍스트는 '강력1반'으로 완벽히 수정됨. 책상 위 서류와 부검 사진도 묘사됨. 그러나 중심 인물인 앉아 있는 남자는 서의용 레퍼런스와 전혀 다른 얼굴에 셔츠와 넥타이를 착용하고 있음. 반면 프레임 우측 전경의 어깨(나상혁 역이어야 함)가 서의용의 상징인 파란색 비니와 후드, 가죽 재킷을 입고 있음.",
      "hard_violations": [
       "인물 신원 및 복장 배치 오류 (지정된 서의용이 의자에 없고 엉뚱한 인물이 앉아 있으며, 서의용의 복장이 전경의 어깨로 뒤바뀜)"
      ],
      "physics": "앉은 인물은 의자에 등을 돌린 채 정상적으로 앉아 자세를 유지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1001,
     "B": 538
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S24sh4"
  }
 },
 "S24sh5::cine": {
  "applied": true,
  "fingerprint": "b2975735e9411677d775e0bd49cf182da4071f03e266b05deb88dcc1cebb4a89",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S24sh5_sel.png",
  "source_sha256": "a8749315115b84a8aacbb717107aec97ae64dd92478c19f1920fc2ca858a7d63",
  "file": "S24sh5_cine.png",
  "latency_ms": 10691
 },
 "S25sh3::confined_fp_apt": {
  "applies": true,
  "reason_ko": "자동차 내부 조수석이라는 제한된 공간에서 운전자와 탑승자 간의 좌석 배치와 고개를 돌리는 시선 방향이 정확히 연출되어야 이야기의 흐름이 깨지지 않으므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "964c91cf3c68cb34"
 },
 "S25sh3::signage": {
  "fp": "3c4c972addc134a4",
  "inscriptions": []
 },
 "confinedfp::f2fd29404415": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/confinedfp_base_f2fd29404415.png",
  "place_text": "Inside the moving car at the front passenger seat, closely bordered by the driver's seat, dashboard, window, and controls.",
  "input_fingerprint": "5a876e7ad56266f5"
 },
 "S25sh3::confined_fp": {
  "reads": {
   "controls": "A steering wheel is located in front of the Driver Seat, attached to the dashboard on the left side of the cabin.",
   "mirrors": "No mirrors are present or marked in the diagram.",
   "camera": "The camera is positioned in the center-rear of the cabin space, behind the front seats, pointing straight forward towards the dashboard.",
   "occupants": "Nasanghyeok occupies the Front Passenger Seat on the right. An unnamed occupant icon is in the Driver Seat on the left."
  },
  "mismatches": [
   "The text describes the camera dolly starting 'just forward of the gap' for a close-up on Nasanghyeok, but the diagram places the camera behind the gap pointing straight forward at the dashboard."
  ],
  "scene_description_en": "The camera is positioned centrally in the rear of the cabin, pointing straight forward. In the near-left midground is the driver's seat, facing forward, with a steering wheel visible on the dashboard directly in front of it. On the near-right midground, the front passenger seat is occupied by Nasanghyeok, who is positioned facing forward. A wide dashboard spans the entire background across the center. Elongated windows enclose the far left and far right edges of the view.",
  "fixed": true,
  "input_fingerprint": "a38d2580477e09f8"
 },
 "era_assess::bdd01684e315b26d": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 초반 한국 승용차 내부 (조수석 및 대시보드)",
    "search_terms_native": [
     "국산차 내부 2000년대",
     "소나타 센터페시아 실내",
     "아반떼 내부 대시보드",
     "NF소나타 조수석 실내"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며, 영어나 다른 언어로 번역하거나 추가 검색어를 덧붙이지 마십시오.",
    "reason_ko": "2000년대 및 2010년대 초반 한국에서 흔히 볼 수 있었던 현대/기아 등 국산 승용차의 조수석, 대시보드 및 센터페시아의 특유의 플라스틱 마감, 버튼 배열, 에어컨 송풍구 형태가 일반적인 현대식 외산 차량과 크게 다르기 때문입니다."
   }
  ]
 },
 "era_ref::d2c0f3604665c955": {
  "subject": "2000년대~2010년대 초반 한국 승용차 내부 (조수석 및 대시보드)",
  "terms": [
   "국산차 내부 2000년대",
   "소나타 센터페시아 실내",
   "아반떼 내부 대시보드",
   "NF소나타 조수석 실내"
  ],
  "queries": [
   [
    "소나타 센터페시아 실내 2000년대 NF소나타 조수석 실내",
    "아반떼 내부 대시보드 2000년대 국산차 내부"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://cnp1753-prod.s3.amazonaws.com/cars/6WtIiVv3QDxSCo/lV193WpxMvMU9Qmukhrt.jpg",
  "picked_reason_ko": "2번은 2000년대 한국 현대 승용차의 실제 사용된 실내로, 조수석과 대시보드 전체의 비례·재료·구성과 사용 흔적을 가장 명확하게 보여준다.",
  "sha256": "e09777f553775806211e4acb725692a2bbad21d6b74a849daeb83527754f7145",
  "file": "eraref_d2c0f3604665c955.png"
 },
 "S25sh3": {
  "input_fingerprint": "1b07506beca15187",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽으로 고개를 돌린 채 서늘한 눈빛으로 입을 벌린 나상혁의 얼굴\n\nLOCATION (lock): Inside the moving car at the front passenger seat, closely bordered by the driver's seat, dashboard, window, and controls. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Dolly in from just forward of the gap between the front seats to a close view at 나상혁's face height, keeping him slightly right of center and isolating him from the driver. His torso remains seated forward but his head turns toward 서의용 off frame left, mouth opening on the cool reply, with a narrow strip of the seat gap preserving the cabin geography.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 나상혁 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 앞좌석 사이 공간 (Occupied by the two seated men) — The inner edges of the two front seats converge around 나상혁's close framing; used as The gap provides a restrained foreground frame around 나상혁 without introducing the driver; 조수석 (Occupied by 나상혁) — Its back and side align toward the front of the moving car, with 나상혁 seated against it; used as Places 나상혁 physically in the passenger position while allowing his head turn toward the driver.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination inside the car remains subdued and neutral-to-cool, with natural facial contrast rather than dramatic modeling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned centrally in the rear of the cabin, pointing straight forward. In the near-left midground is the driver's seat, facing forward, with a steering wheel visible on the dashboard directly in front of it. On the near-right midground, the front passenger seat is occupied by Nasanghyeok, who is positioned facing forward. A wide dashboard spans the entire background across the center. Elongated windows enclose the far left and far right edges of the view.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽으로 고개를 돌린 채 서늘한 눈빛으로 입을 벌린 나상혁의 얼굴\n\nLOCATION (lock): Inside the moving car at the front passenger seat, closely bordered by the driver's seat, dashboard, window, and controls. The shot takes place here.\n",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is positioned centrally in the rear of the cabin, pointing straight forward. In the near-left midground is the driver's seat, facing forward, with a steering wheel visible on the dashboard directly in front of it. On the near-right midground, the front passenger seat is occupied by Nasanghyeok, who is positioned facing forward. A wide dashboard spans the entire background across the center. Elongated windows enclose the far left and far right edges of the view.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽으로 고개를 돌린 채 서늘한 눈빛으로 입을 벌린 나상혁의 얼굴\n\nLOCATION (lock): Inside the moving car at the front passenger seat, closely bordered by the driver's seat, dashboard, window, and controls. The shot takes place here.\n\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S25sh3_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대~2010년대 초반 한국 승용차 내부 (조수석 및 대시보드): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_d2c0f3604665c955.png"
    },
    {
     "label": "나상혁",
     "path": "<bytes:747016>"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S25sh3_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대~2010년대 초반 한국 승용차 내부 (조수석 및 대시보드): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_d2c0f3604665c955.png"
    },
    {
     "label": "나상혁",
     "path": "<bytes:747016>"
    }
   ]
  },
  "readings": [
   {
    "label": "A",
    "direction": "나상혁이 좌측 운전석의 인물을 향해 고개를 돌리고 시선을 향함.",
    "built_space": "1개의 스티어링 휠, 좌우 시트 및 중앙 콘솔 등 차량 내부 구조가 정상 배치됨.",
    "entities": "나상혁의 외모와 의상은 일치하나, 지시와 달리 프레임 좌측에 운전자가 포함됨.",
    "hard_violations": [],
    "physics": "두 인물 모두 시트에 안정적으로 착석했으며 운전자는 핸들을 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "나상혁이 화면 좌측 밖을 향해 시선을 던짐.",
    "built_space": "좌측 전경의 핸들 외에 우측 조수석 대시보드에도 두 번째 스티어링 휠이 렌더링됨.",
    "entities": "나상혁의 외양은 일치하나 차량 구조 왜곡으로 인해 부자연스러운 공간에 놓임.",
    "hard_violations": [
     "중복된 제어 장치 (조수석 대시보드에 두 번째 스티어링 휠 생성)"
    ],
    "physics": "인물이 시트에 착석해 자세를 유지함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 4,
   "B": 2
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "클로즈업 프레이밍 및 운전자 배제 지시를 어기고 와이드 샷으로 렌더링했으나, 치명적 위반은 없음."
   },
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "조수석에 두 번째 스티어링 휠이 생성되는 치명적인 구조적 오류(중복된 제어 장치)가 발생함."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S25sh3_confinedfp.png"
   },
   {
    "label": "PERIOD REFERENCE — 2000년대~2010년대 초반 한국 승용차 내부 (조수석 및 대시보드): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_d2c0f3604665c955.png"
   },
   {
    "label": "나상혁",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "대시보드의 중앙 화면, 조작 버튼, 기어 변속기 주변, 선바이저 경고문 등의 텍스트와 기호가 읽을 수 없게 뭉개져 있음.",
     "fix_en": "Replace the illegible gibberish text on the dashboard screen, controls, gear shifter, and sun visor with generic, unreadable UI elements or blank shapes. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "차량 내부 레퍼런스 이미지에 명시된 대시보드와 차문의 우드 트림(나무 무늬 장식) 무늬가 누락되고 붉은색 선형 장식으로 잘못 대체됨.",
     "fix_en": "Change the red linear trim along the dashboard and doors to a brown wood grain trim that matches the period reference. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "minor",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "대시보드의 중앙 화면, 조작 버튼, 기어 변속기 주변, 선바이저 경고문 등의 텍스트와 기호가 읽을 수 없게 뭉개져 있음.",
     "severity": "minor"
    },
    {
     "issue_ko": "차량 내부 레퍼런스 이미지에 명시된 대시보드와 차문의 우드 트림(나무 무늬 장식) 무늬가 누락되고 붉은색 선형 장식으로 잘못 대체됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "샷 텍스트가 나상혁의 얼굴을 지시하는데 운전석까지 포함한 차량 앞쪽 전체 와이드 샷이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "대시보드의 중앙 화면, 조작 버튼, 기어 변속기 주변, 선바이저 경고문 등의 텍스트와 기호가 읽을 수 없게 뭉개져 있음.",
    "fix_en": "Replace the illegible gibberish text on the dashboard screen, controls, gear shifter, and sun visor with generic, unreadable UI elements or blank shapes. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "차량 내부 레퍼런스 이미지에 명시된 대시보드와 차문의 우드 트림(나무 무늬 장식) 무늬가 누락되고 붉은색 선형 장식으로 잘못 대체됨.",
    "fix_en": "Change the red linear trim along the dashboard and doors to a brown wood grain trim that matches the period reference. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "minor",
    "observation_index": 1
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "confined_fp": {
   "base_key": "confinedfp::f2fd29404415",
   "apt_reason": "자동차 내부 조수석이라는 제한된 공간에서 운전자와 탑승자 간의 좌석 배치와 고개를 돌리는 시선 방향이 정확히 연출되어야 이야기의 흐름이 깨지지 않으므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": true,
   "mismatches": [
    "The text describes the camera dolly starting 'just forward of the gap' for a close-up on Nasanghyeok, but the diagram places the camera behind the gap pointing straight forward at the dashboard."
   ],
   "era_research": {
    "subject": "2000년대~2010년대 초반 한국 승용차 내부 (조수석 및 대시보드)",
    "queries": [
     [
      "소나타 센터페시아 실내 2000년대 NF소나타 조수석 실내",
      "아반떼 내부 대시보드 2000년대 국산차 내부"
     ]
    ],
    "picked_url": "https://cnp1753-prod.s3.amazonaws.com/cars/6WtIiVv3QDxSCo/lV193WpxMvMU9Qmukhrt.jpg",
    "sha256": "e09777f553775806211e4acb725692a2bbad21d6b74a849daeb83527754f7145",
    "file": "eraref_d2c0f3604665c955.png"
   }
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S25sh3::cine": {
  "applied": true,
  "fingerprint": "0b705e2232a8b2faff6cc3c104b04b5dd6320ef11a1a274c34e58c1b4fc28b1f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S25sh3_sel.png",
  "source_sha256": "d7d516b973bfb72f0fa2fcbe1de44fc43361b96d8231697434865df64cf9ac4c",
  "file": "S25sh3_cine.png",
  "latency_ms": 12335
 },
 "S25sh4::confined_fp_apt": {
  "applies": true,
  "reason_ko": "차량 내부 운전석이라는 통제 공간에서 운전자와 조수석 인물의 위치 관계와 시선 처리가 샷의 상황을 전달하는 데 매우 중요하기 때문입니다.",
  "input_fingerprint": "1118b8e602ff6a94"
 },
 "S25sh4::signage": {
  "fp": "808a5393a1bba9d1",
  "inscriptions": []
 },
 "confinedfp::c09d00655cb2": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/confinedfp_base_c09d00655cb2.png",
  "place_text": "Inside the moving car at the driver's seat, with the passenger, steering wheel, mirrors, and dashboard close around him.",
  "input_fingerprint": "9779cb9b8e12b67e"
 },
 "S25sh4::confined_fp": {
  "reads": {
   "controls": "STEERING WHEEL attached to DRIVER (SEAT).",
   "mirrors": "DRIVER SIDE MIRROR on the left facing inward/right; REARVIEW MIRROR top center facing backward/down; PASSENGER SIDE MIRROR on the right facing inward/left.",
   "camera": "Positioned on the far right (outside the passenger area), pointing directly left towards the driver.",
   "occupants": "DRIVER (SEAT) is occupied by 서의용. PASSENGER (SEAT) is empty. An unmarked figure's shoulder is drawn on the far left outside the car."
  },
  "mismatches": [
   "The text describes 나상혁's turned shoulder in the near-left foreground, but the diagram places the camera on the right pointing left, which puts the only drawn shoulder figure in the far background outside the driver's window.",
   "The text states the shot is inside the car 'with the passenger', but the diagram shows the PASSENGER (SEAT) as empty."
  ],
  "scene_description_en": "The camera is stationed on the passenger side, looking directly left across the cabin interior. The empty passenger seat occupies the immediate near foreground. In the center midground, 서의용 sits in the driver's seat; his body is oriented toward the right side of the frame (the front of the car), but his head is turned to look toward the camera. To the right of the driver in the frame, the steering wheel is mounted at the driver's station. In the background directly behind the driver, the driver's side mirror is visible, facing inward toward the cabin. On the far right edge of the frame, the rearview mirror faces the left side of the screen, while an unidentified, large figure's shoulder appears in the extreme far background beyond the driver's side window.",
  "fixed": true,
  "input_fingerprint": "86aa1577db90ddc6"
 },
 "era_assess::bca3b9ca333e4400": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 한국 자동차 내부 (대시보드 및 운전석)",
    "search_terms_native": [
     "국산차 내부 운전석 2000년대",
     "쏘나타 아반떼 실내 대시보드",
     "자동차 센터페시아 2010년"
    ],
    "language_lock_native": "검색어 제안 및 검색은 오직 한국어로만 진행해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대와 2010년대 한국에서 흔히 볼 수 있었던 국산 차량(현대 쏘나타, 아반떼 등)의 센터페시아 레이아웃, 스티어링 휠 로고, 내부 마감재 특징을 정확히 재현하기 위함입니다."
   }
  ]
 },
 "era_ref::60e46dd7bf83a084": {
  "subject": "2000년대-2010년대 한국 자동차 내부 (대시보드 및 운전석)",
  "terms": [
   "국산차 내부 운전석 2000년대",
   "쏘나타 아반떼 실내 대시보드",
   "자동차 센터페시아 2010년"
  ],
  "queries": [
   [
    "2000년대 국산차 실내 운전석 대시보드 쏘나타 아반떼 센터페시아",
    "2010년 국산 자동차 실내 센터페시아 운전석"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://img.carswp.com/hyundai/sonata/photos_hyundai_sonata_2001_2_b.jpg",
  "picked_reason_ko": "1번은 2000년대 초반 한국 현대 승용차의 운전석과 대시보드를 정면에서 크게 담아 계기판, 센터페시아, 변속기, 재료와 마감이 가장 명확하게 읽힌다.",
  "sha256": "efe1db749013811045cffda5d48617c65e98edadbb104733460ff60ad0e2e292",
  "file": "eraref_60e46dd7bf83a084.png"
 },
 "S25sh4": {
  "input_fingerprint": "d7d859b343c2406a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나상혁의 대답에 눈을 크게 뜨고 입술이 살짝 벌어진 채 굳어버린 서의용의 옆얼굴\n\nLOCATION (lock): Inside the moving car at the driver's seat, with the passenger, steering wheel, mirrors, and dashboard close around him. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The lateral track settles near the passenger side at slightly below 서의용's face height, observing his right profile across the cabin at close range. 서의용 holds the driving posture but freezes with widened eyes and parted lips as his glance flicks toward 나상혁; the soft edge of 나상혁's turned shoulder stays in the near-left foreground.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 운전석 (Occupied by 서의용) — The seat faces the front of the car while 서의용's head turns briefly toward the passenger side; used as Keeps 서의용's driving position readable beneath his arrested reaction; 차량 앞좌석 공간 (Vehicle in motion) — The passenger side lies close to camera and the driver sits across the cabin; used as Creates the lateral depth separating the foreground passenger shoulder from the driver's profile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime cabin ambience keeps the palette neutral-to-cool and the surprise readable without heightened or stylized lighting.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed on the passenger side, looking directly left across the cabin interior. The empty passenger seat occupies the immediate near foreground. In the center midground, 서의용 sits in the driver's seat; his body is oriented toward the right side of the frame (the front of the car), but his head is turned to look toward the camera. To the right of the driver in the frame, the steering wheel is mounted at the driver's station. In the background directly behind the driver, the driver's side mirror is visible, facing inward toward the cabin. On the far right edge of the frame, the rearview mirror faces the left side of the screen, while an unidentified, large figure's shoulder appears in the extreme far background beyond the driver's side window.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나상혁의 대답에 눈을 크게 뜨고 입술이 살짝 벌어진 채 굳어버린 서의용의 옆얼굴\n\nLOCATION (lock): Inside the moving car at the driver's seat, with the passenger, steering wheel, mirrors, and dashboard close around him. The shot takes place here.\n",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed on the passenger side, looking directly left across the cabin interior. The empty passenger seat occupies the immediate near foreground. In the center midground, 서의용 sits in the driver's seat; his body is oriented toward the right side of the frame (the front of the car), but his head is turned to look toward the camera. To the right of the driver in the frame, the steering wheel is mounted at the driver's station. In the background directly behind the driver, the driver's side mirror is visible, facing inward toward the cabin. On the far right edge of the frame, the rearview mirror faces the left side of the screen, while an unidentified, large figure's shoulder appears in the extreme far background beyond the driver's side window.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나상혁의 대답에 눈을 크게 뜨고 입술이 살짝 벌어진 채 굳어버린 서의용의 옆얼굴\n\nLOCATION (lock): Inside the moving car at the driver's seat, with the passenger, steering wheel, mirrors, and dashboard close around him. The shot takes place here.\n\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S25sh4_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대-2010년대 한국 자동차 내부 (대시보드 및 운전석): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_60e46dd7bf83a084.png"
    },
    {
     "label": "서의용",
     "path": "<bytes:929743>"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S25sh4_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 2000년대-2010년대 한국 자동차 내부 (대시보드 및 운전석): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_60e46dd7bf83a084.png"
    },
    {
     "label": "서의용",
     "path": "<bytes:929743>"
    }
   ]
  },
  "readings": [
   {
    "label": "A",
    "direction": "서의용의 시선이 조수석이 아닌 차량 정면을 향하고 있습니다.",
    "built_space": "조수석 위치에서 운전석을 바라보는 구도이며, 스티어링 휠과 대시보드가 보입니다.",
    "entities": "서의용의 인상과 복장은 일치하나, 전경에 나상혁의 어깨가 뚜렷하지 않고 우측 창밖에 의도되지 않은 인물 형상이 떠 있습니다.",
    "hard_violations": [
     "창밖에 물리적으로 불가능한 정체불명의 인물 얼굴 형상이 생성됨 (invented person)"
    ],
    "physics": "인물이 운전석 시트에 정상적으로 앉아 지지받고 있습니다."
   },
   {
    "label": "B",
    "direction": "서의용이 고개를 돌려 조수석(카메라) 방향을 바라보고 있습니다.",
    "built_space": "조수석 측에서 운전석을 향하는 구도이며 휠과 중앙 대시보드가 알맞게 배치되어 있습니다.",
    "entities": "서의용의 인상과 의상이 참조와 일치하며, 프롬프트대로 전경 좌측에 나상혁의 흐릿한 어깨 형태가 존재합니다.",
    "hard_violations": [],
    "physics": "인물이 좌석에 올바르게 착석하여 물리적 지지를 받고 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지시된 대로 운전자가 조수석 쪽으로 고개를 돌려 놀란 표정을 짓고 있으며, 전경에 흐릿한 어깨가 잘 연출되었습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "시선이 정면을 향해 있어 프롬프트의 상황과 맞지 않으며, 창밖에 유령 같은 형상이 나타나는 심각한 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S25sh4_confinedfp.png"
   },
   {
    "label": "PERIOD REFERENCE — 2000년대-2010년대 한국 자동차 내부 (대시보드 및 운전석): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_60e46dd7bf83a084.png"
   },
   {
    "label": "서의용",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "운전석에 앉은 서의용의 얼굴이 카메라 쪽(조수석)이 아닌 반대편 창문 밖을 향하고 있음",
     "fix_en": "Redraw the driver's head to turn towards the camera with wide eyes and slightly parted lips. Preserve his body position, clothing, car interior, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 왼쪽 가까이에 조수석이 위치해야 하나, 현재 구도상 카메라가 조수석 위치보다 뒤쪽이나 다른 위치에 있어 좌석 배치가 레이아웃과 일치하지 않음",
     "fix_en": "Redraw the left foreground to prominently show the empty passenger seat close to the camera. Preserve the driver, dashboard, lighting, and framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "스티어링 휠 등 대시보드 구조가 왼쪽(운전석 측)이 아닌 오른쪽(조수석 측)에 배치되어 차량 내부 구조가 반전됨",
     "fix_en": "Redraw the right dashboard to remove the steering wheel and instrument cluster, leaving a plain passenger-side surface. Preserve the driver, clothing, lighting, and framing.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 오른쪽 끝에 나타나야 할 신원 미상 인물의 어깨가 보이지 않음",
     "fix_en": "Add the shoulder of an unidentified large figure in the far background visible through the right window. Preserve the driver, interior, lighting, and framing.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "운전석에 앉은 서의용의 얼굴이 카메라 쪽(조수석)이 아닌 반대편 창문 밖을 향하고 있음",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 왼쪽 가까이에 조수석이 위치해야 하나, 현재 구도상 카메라가 조수석 위치보다 뒤쪽이나 다른 위치에 있어 좌석 배치가 레이아웃과 일치하지 않음",
     "severity": "major"
    },
    {
     "issue_ko": "스티어링 휠 등 대시보드 구조가 왼쪽(운전석 측)이 아닌 오른쪽(조수석 측)에 배치되어 차량 내부 구조가 반전됨",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 오른쪽 끝에 나타나야 할 신원 미상 인물의 어깨가 보이지 않음",
     "severity": "major"
    },
    {
     "issue_ko": "중앙의 서의용 얼굴이 옆얼굴이 아니라 카메라를 향한 거의 정면이다.",
     "severity": "major"
    },
    {
     "issue_ko": "계기판이 스티어링 휠 뒤가 아니라 오른쪽에 붙어 참조 차량 배치와 다르다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 왼쪽 가까이에 조수석이 위치해야 하나, 현재 구도상 카메라가 조수석 위치보다 뒤쪽이나 다른 위치에 있어 좌석 배치가 레이아웃과 일치하지 않음",
    "fix_en": "Redraw the left foreground to prominently show the empty passenger seat close to the camera. Preserve the driver, dashboard, lighting, and framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "화면 오른쪽 끝에 나타나야 할 신원 미상 인물의 어깨가 보이지 않음",
    "fix_en": "Add the shoulder of an unidentified large figure in the far background visible through the right window. Preserve the driver, interior, lighting, and framing.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the driver's head to turn towards the camera with wide eyes and slightly parted lips. Preserve his body position, clothing, car interior, lighting, and framing.\n- Redraw the right dashboard to remove the steering wheel and instrument cluster, leaving a plain passenger-side surface. Preserve the driver, clothing, lighting, and framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "레퍼런스의 대시보드와 차량 내부 디테일을 놀랍도록 정확하게 구현했으며, 조수석을 전경에 둔 평면도의 구도와 인물의 복장, 표정 지시를 충실히 따랐으나 창밖의 미확인 인물 형상은 누락되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 복장과 표정은 묘사되었으나, 평면도에서 지시한 조수석 전경이 센터 콘솔로 대체되었고 샷 텍스트가 요구한 '옆얼굴' 대신 정면을 응시하고 있으며 배경이 백색으로 비어있어 몰입을 깹니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "서의용이 조수석 쪽(카메라 방향)으로 고개를 돌려 바라보고 있음.",
      "built_space": "카메라가 조수석 쪽에 위치해 전경에 조수석 시트가 보이며, 프레임 우측에 위치한 운전대와 대시보드가 시대적 레퍼런스와 매우 정확히 일치함. 운전석 창밖으로 사이드 미러가 보임.",
      "entities": "서의용은 비니, 후드, 가죽 재킷, 경찰 배지 등 레퍼런스의 복장을 정확히 착용하고 있으며, 눈을 크게 뜨고 입술이 벌어진 옆얼굴/반측면 얼굴을 보여줌. 창밖의 거대한 인물 어깨는 보이지 않음.",
      "hard_violations": [],
      "physics": "인물이 운전석에 안정적으로 앉아 있으며, 창밖 풍경의 모션 블러를 통해 차량이 이동 중임이 물리적으로 자연스럽게 표현됨."
     },
     {
      "label": "B",
      "direction": "서의용이 렌즈를 정면으로 응시하고 있음.",
      "built_space": "카메라 뷰 앞쪽에 조수석 시트 대신 센터 콘솔이 위치해 평면도 지시를 어겼으며, 대시보드의 형태가 잘려있고 레퍼런스와 다름. 창밖 배경이 완전히 하얗게 날아감.",
      "entities": "서의용의 신원과 복장(비니, 가죽 재킷, 배지 등)은 일치하나, 샷 텍스트가 요구한 옆얼굴이 아닌 정면 얼굴임. 창밖 인물 없음.",
      "hard_violations": [
       "평면도에서 지시한 카메라 위치 및 전경 레이어(조수석 시트) 위반"
      ],
      "physics": "인물은 좌석에 앉아 운전대를 잡고 있으나, 차량 외부가 백색 공간으로 처리되어 이동 중이라는 물리적 상황을 전혀 설명하지 못함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "레퍼런스의 대시보드와 차량 내부 디테일을 놀랍도록 정확하게 구현했으며, 조수석을 전경에 둔 평면도의 구도와 인물의 복장, 표정 지시를 충실히 따랐으나 창밖의 미확인 인물 형상은 누락되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 복장과 표정은 묘사되었으나, 평면도에서 지시한 조수석 전경이 센터 콘솔로 대체되었고 샷 텍스트가 요구한 '옆얼굴' 대신 정면을 응시하고 있으며 배경이 백색으로 비어있어 몰입을 깹니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서의용이 조수석 쪽(카메라 방향)으로 고개를 돌려 바라보고 있음.",
      "built_space": "카메라가 조수석 쪽에 위치해 전경에 조수석 시트가 보이며, 프레임 우측에 위치한 운전대와 대시보드가 시대적 레퍼런스와 매우 정확히 일치함. 운전석 창밖으로 사이드 미러가 보임.",
      "entities": "서의용은 비니, 후드, 가죽 재킷, 경찰 배지 등 레퍼런스의 복장을 정확히 착용하고 있으며, 눈을 크게 뜨고 입술이 벌어진 옆얼굴/반측면 얼굴을 보여줌. 창밖의 거대한 인물 어깨는 보이지 않음.",
      "hard_violations": [],
      "physics": "인물이 운전석에 안정적으로 앉아 있으며, 창밖 풍경의 모션 블러를 통해 차량이 이동 중임이 물리적으로 자연스럽게 표현됨."
     },
     {
      "label": "B",
      "direction": "서의용이 렌즈를 정면으로 응시하고 있음.",
      "built_space": "카메라 뷰 앞쪽에 조수석 시트 대신 센터 콘솔이 위치해 평면도 지시를 어겼으며, 대시보드의 형태가 잘려있고 레퍼런스와 다름. 창밖 배경이 완전히 하얗게 날아감.",
      "entities": "서의용의 신원과 복장(비니, 가죽 재킷, 배지 등)은 일치하나, 샷 텍스트가 요구한 옆얼굴이 아닌 정면 얼굴임. 창밖 인물 없음.",
      "hard_violations": [
       "평면도에서 지시한 카메라 위치 및 전경 레이어(조수석 시트) 위반"
      ],
      "physics": "인물은 좌석에 앉아 운전대를 잡고 있으나, 차량 외부가 백색 공간으로 처리되어 이동 중이라는 물리적 상황을 전혀 설명하지 못함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "샷 텍스트의 '옆얼굴'과 놀란 표정을 적절히 묘사했으며, 이동 중인 차량 내부라는 장소 고정 조건을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물의 복장과 표정은 양호하나, 창밖 배경이 하얗게 비워져 있어 장소 설정을 심각하게 위반했고 지시된 '옆얼굴' 대신 정면을 그렸습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선이 카메라가 위치한 왼쪽 방향을 향하고 있음.",
      "built_space": "조수석 너머로 촬영된 구도. 전경에 조수석 일부가 보이며, 우측에 운전대와 대시보드가 배치됨. 창밖으로 이동 중인 외부 풍경이 보임.",
      "entities": "서의용의 인상착의(비니, 가죽 재킷, 후드 등)가 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "차량 시트에 체중을 싣고 안정적으로 앉아 있음."
     },
     {
      "label": "A",
      "direction": "렌즈를 정면으로 응시하고 있음.",
      "built_space": "조수석 위치에서 촬영. 우측에 운전대 배치. 창밖 배경이 완전히 하얀 스튜디오 환경처럼 묘사됨.",
      "entities": "서의용의 인상착의가 레퍼런스와 일치함.",
      "hard_violations": [
       "'이동 중인 차량 내부'라는 장소 고정 조건 위반 (창밖이 텅 빈 흰색 배경임)"
      ],
      "physics": "차량 시트에 앉아 스티어링 휠을 잡고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "샷 텍스트의 '옆얼굴'과 놀란 표정을 적절히 묘사했으며, 이동 중인 차량 내부라는 장소 고정 조건을 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 복장과 표정은 양호하나, 창밖 배경이 하얗게 비워져 있어 장소 설정을 심각하게 위반했고 지시된 '옆얼굴' 대신 정면을 그렸습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 카메라가 위치한 왼쪽 방향을 향하고 있음.",
      "built_space": "조수석 너머로 촬영된 구도. 전경에 조수석 일부가 보이며, 우측에 운전대와 대시보드가 배치됨. 창밖으로 이동 중인 외부 풍경이 보임.",
      "entities": "서의용의 인상착의(비니, 가죽 재킷, 후드 등)가 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "차량 시트에 체중을 싣고 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "렌즈를 정면으로 응시하고 있음.",
      "built_space": "조수석 위치에서 촬영. 우측에 운전대 배치. 창밖 배경이 완전히 하얀 스튜디오 환경처럼 묘사됨.",
      "entities": "서의용의 인상착의가 레퍼런스와 일치함.",
      "hard_violations": [
       "'이동 중인 차량 내부'라는 장소 고정 조건 위반 (창밖이 텅 빈 흰색 배경임)"
      ],
      "physics": "차량 시트에 앉아 스티어링 휠을 잡고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "confined_fp": {
   "base_key": "confinedfp::c09d00655cb2",
   "apt_reason": "차량 내부 운전석이라는 통제 공간에서 운전자와 조수석 인물의 위치 관계와 시선 처리가 샷의 상황을 전달하는 데 매우 중요하기 때문입니다.",
   "fixed": true,
   "mismatches": [
    "The text describes 나상혁's turned shoulder in the near-left foreground, but the diagram places the camera on the right pointing left, which puts the only drawn shoulder figure in the far background outside the driver's window.",
    "The text states the shot is inside the car 'with the passenger', but the diagram shows the PASSENGER (SEAT) as empty."
   ],
   "era_research": {
    "subject": "2000년대-2010년대 한국 자동차 내부 (대시보드 및 운전석)",
    "queries": [
     [
      "2000년대 국산차 실내 운전석 대시보드 쏘나타 아반떼 센터페시아",
      "2010년 국산 자동차 실내 센터페시아 운전석"
     ]
    ],
    "picked_url": "https://img.carswp.com/hyundai/sonata/photos_hyundai_sonata_2001_2_b.jpg",
    "sha256": "efe1db749013811045cffda5d48617c65e98edadbb104733460ff60ad0e2e292",
    "file": "eraref_60e46dd7bf83a084.png"
   }
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S25sh3"
  }
 },
 "S25sh4::cine": {
  "applied": true,
  "fingerprint": "d8420d30a35baa065a1c28189390dc67dc415f0f204bd662411fa4007b4d023b",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S25sh4_sel.png",
  "source_sha256": "ccfa8b01ede559ad817ec2b5a310568ccd040342858a0ab58727e14e5842adfa",
  "file": "S25sh4_cine.png",
  "latency_ms": 12837
 },
 "S26sh1::signage": {
  "fp": "b8d6b6eec211acf1",
  "inscriptions": [
   {
    "surface_native": "모니터 화면의 문서 상단",
    "text_native": "부검보고서",
    "reason_ko": "부검의가 분석 중인 모니터 화면 속 문서가 실제 부검 문서임을 직관적으로 보여주기 위해 제목이 필요합니다."
   }
  ]
 },
 "era_assess::991b1ecfb8b9ffa6": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 대학 의과대학 법의학 연구실",
    "search_terms_native": [
     "의과대학 연구실",
     "법의학교실 내부",
     "대학병원 진료실 2010년",
     "의대 교수실"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성되어야 하며, 영어 등 다른 언어로의 번역이나 혼용은 금지됩니다.",
    "reason_ko": "2000년대에서 2010년대 초반 사이 한국 의과대학 연구실 특유의 한국형 사무 가구 배치, 한글 의학 서적의 형태 및 병원 진료실 특유의 실내 마감 분위기를 정확하게 묘사하기 위함입니다."
   }
  ]
 },
 "era_ref::1b38c72de354ac18": {
  "subject": "2000년대~2010년대 한국 대학 의과대학 법의학 연구실",
  "terms": [
   "의과대학 연구실",
   "법의학교실 내부",
   "대학병원 진료실 2010년",
   "의대 교수실"
  ],
  "queries": [
   [
    "2010년 한국 의과대학 법의학교실 연구실 내부 의대 교수실",
    "2010년 한국 대학병원 진료실 내부 의과대학 연구실"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://medix.hanyang.ac.kr/documents/331011283/338799794/%EB%B3%91%EB%A6%AC%ED%95%99%EA%B5%90%EC%8B%A4_%EC%9D%B4%EB%AF%B8%EC%A7%80.jpg/5502f118-f878-1d6f-d63c-6bd350f23eec?download=true&t=1760575260059",
  "picked_reason_ko": "4번은 현미경, 표본·서류 작업대, 컴퓨터와 전문서적이 갖춰진 한국 대학의 일상적인 법의학 연구실 내부를 가장 명확하게 보여준다.",
  "sha256": "cf440e27d4f628875a6d3c1beaef444fdab3a2d14c20f960bcca6eb544bb0293",
  "file": "eraref_1b38c72de354ac18.png"
 },
 "groupbg::법의학 연구진찰실": {
  "input_fingerprint": "35d7f6a0a8b2e606",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "법의학 연구진찰실",
    "tags": [
     "S26sh1",
     "S26sh6"
    ]
   },
   "context_sig": "20ad35511b6106b9",
   "era_research_sha": "cf440e27d4f628875a6d3c1beaef444fdab3a2d14c20f960bcca6eb544bb0293"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: At the monitor workstation inside a medical-school research room resembling a hospital examination room.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n전남대학교 법의학교실: 병원 진찰실처럼 깔끔하고 임상적인 분위기의 대학교 법의학 연구실이다. (특징: 흰색 가운이 걸린 진료실 스타일의 깨끗한 내벽; 부검 사진이 크게 띄워져 있는 고해상도 컴퓨터 듀얼 모니터; 의학 전문 서적이 빽빽한 책장)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 전남대 법의학교실 - 낮\n- 병원 진찰실 느낌의 연구실.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 대학 의과대학 법의학 연구실: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: At the monitor workstation inside a medical-school research room resembling a hospital examination room.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n전남대학교 법의학교실: 병원 진찰실처럼 깔끔하고 임상적인 분위기의 대학교 법의학 연구실이다. (특징: 흰색 가운이 걸린 진료실 스타일의 깨끗한 내벽; 부검 사진이 크게 띄워져 있는 고해상도 컴퓨터 듀얼 모니터; 의학 전문 서적이 빽빽한 책장)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 전남대 법의학교실 - 낮\n- 병원 진찰실 느낌의 연구실.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 대학 의과대학 법의학 연구실: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_법의학_연구진찰실_510618.png",
  "asset_id": "064c1a0c-3de6-4208-ab1d-c3dbf69519d5",
  "input_asset_ids": [
   "ad98f13a-f3d2-492d-a963-fed2f0bd42fc"
  ],
  "origin_tag": "S26sh1",
  "place_text": "At the monitor workstation inside a medical-school research room resembling a hospital examination room.",
  "origin_inputs": {
   "place_text": "At the monitor workstation inside a medical-school research room resembling a hospital examination room.",
   "time_of_day_en": "day",
   "conti_asset_id": "ad98f13a-f3d2-492d-a963-fed2f0bd42fc"
  },
  "era_research": {
   "subject": "2000년대~2010년대 한국 대학 의과대학 법의학 연구실",
   "terms": [
    "의과대학 연구실",
    "법의학교실 내부",
    "대학병원 진료실 2010년",
    "의대 교수실"
   ],
   "queries": [
    [
     "2010년 한국 의과대학 법의학교실 연구실 내부 의대 교수실",
     "2010년 한국 대학병원 진료실 내부 의과대학 연구실"
    ]
   ],
   "candidates": 4,
   "picked_index": 4,
   "picked_url": "https://medix.hanyang.ac.kr/documents/331011283/338799794/%EB%B3%91%EB%A6%AC%ED%95%99%EA%B5%90%EC%8B%A4_%EC%9D%B4%EB%AF%B8%EC%A7%80.jpg/5502f118-f878-1d6f-d63c-6bd350f23eec?download=true&t=1760575260059",
   "picked_reason_ko": "4번은 현미경, 표본·서류 작업대, 컴퓨터와 전문서적이 갖춰진 한국 대학의 일상적인 법의학 연구실 내부를 가장 명확하게 보여준다.",
   "sha256": "cf440e27d4f628875a6d3c1beaef444fdab3a2d14c20f960bcca6eb544bb0293",
   "file": "eraref_1b38c72de354ac18.png",
   "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_1b38c72de354ac18.png"
  }
 },
 "era_assess::121e47831beab0a4": {
  "subjects": [
   {
    "subject_native": "2000년대 대한민국 대학병원 진료실 모니터와 의료 시스템 컴퓨터",
    "search_terms_native": [
     "2000년대 병원 진료실 컴퓨터",
     "대학병원 PACS 모니터",
     "의료 정보 시스템 화면"
    ],
    "language_lock_native": "이 검색은 오직 한국어 검색어로만 수행되어야 하며, 영어 등 다른 언어로 번역하거나 결합하여 검색해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 병원 진료실에서 사용하던 특유의 EMR/PACS 의료 프로그램 화면(UI) 및 가로/세로 듀얼 모니터 배치 방식은 일반적인 현대 모니터 및 서구권 병원 이미지와 크게 다릅니다."
   }
  ]
 },
 "era_ref::483c9a13a91174d8": {
  "subject": "2000년대 대한민국 대학병원 진료실 모니터와 의료 시스템 컴퓨터",
  "terms": [
   "2000년대 병원 진료실 컴퓨터",
   "대학병원 PACS 모니터",
   "의료 정보 시스템 화면"
  ],
  "queries": [
   [
    "2000년대 병원 진료실 컴퓨터 대학병원 PACS 모니터 의료 정보 시스템 화면",
    "2000년대 대한민국 대학병원 진료실 PACS 의료정보시스템 컴퓨터 화면"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://cdn.jhealthmedia.joins.com/news/photo/202108/24217_22724_1034.jpg",
  "picked_reason_ko": "대한민국 대학병원에서 실제 사용하는 다중 모니터 의료 영상·전산 워크스테이션의 장비 구성과 배치, 재질을 가장 명확하게 보여주는 실사진이다.",
  "sha256": "a639aa707f935ba6126ae1a2421ecd7eeb16d9c8bb6c79705f616c08cce9e1fb",
  "file": "eraref_483c9a13a91174d8.png"
 },
 "S26sh1::bgfirst_bg": {
  "input_fingerprint": "1f83bd5835bb79df",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 모니터 화면 속 훼손된 부검 사진 위를 손가락으로 가리킨 채 입을 벌린 50대 한국인 남성 부검의의 상체\n\nLOCATION (lock): At the monitor workstation inside a medical-school research room resembling a hospital examination room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height on the monitor side, frame 전남대 부검의 in a medium three-quarter view on the right and the monitor on the left, both sharing the discussion plane. He leans toward the display with one arm raised, his finger indicating an indistinct area of the autopsy photograph while his attention follows downward to the detail he is explaining.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전남대 부검의 in the middle-right of the frame, midground, points to monitor display; monitor display in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모니터 화면 (Displaying autopsy photographs) — The active display faces partly toward camera and shows multiple autopsy photographs without emphasizing graphic detail; used as Shares the frame with the speaker and serves as the precise anchor for his pointing gesture; 연구실 진찰 공간 (In use for the consultation) — The visible room plane recedes behind the monitor and the seated discussion area; used as Establishes the room as a research space with the clinical feel stated in the scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination uses restrained neutral-to-cool color and naturalistic contrast, keeping the forensic imagery subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 대학병원 진료실 모니터와 의료 시스템 컴퓨터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 모니터 화면 속 훼손된 부검 사진 위를 손가락으로 가리킨 채 입을 벌린 50대 한국인 남성 부검의의 상체\n\nLOCATION (lock): At the monitor workstation inside a medical-school research room resembling a hospital examination room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height on the monitor side, frame 전남대 부검의 in a medium three-quarter view on the right and the monitor on the left, both sharing the discussion plane. He leans toward the display with one arm raised, his finger indicating an indistinct area of the autopsy photograph while his attention follows downward to the detail he is explaining.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전남대 부검의 in the middle-right of the frame, midground, points to monitor display; monitor display in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모니터 화면 (Displaying autopsy photographs) — The active display faces partly toward camera and shows multiple autopsy photographs without emphasizing graphic detail; used as Shares the frame with the speaker and serves as the precise anchor for his pointing gesture; 연구실 진찰 공간 (In use for the consultation) — The visible room plane recedes behind the monitor and the seated discussion area; used as Establishes the room as a research space with the clinical feel stated in the scene.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination uses restrained neutral-to-cool color and naturalistic contrast, keeping the forensic imagery subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 대학병원 진료실 모니터와 의료 시스템 컴퓨터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S26sh1__bgfirst_bg.png",
  "asset_id": "88b3665f-fab6-4531-8a48-57a034593ed1",
  "input_asset_ids": [
   "ad98f13a-f3d2-492d-a963-fed2f0bd42fc",
   "064c1a0c-3de6-4208-ab1d-c3dbf69519d5"
  ],
  "era_research": {
   "subject": "2000년대 대한민국 대학병원 진료실 모니터와 의료 시스템 컴퓨터",
   "queries": [
    [
     "2000년대 병원 진료실 컴퓨터 대학병원 PACS 모니터 의료 정보 시스템 화면",
     "2000년대 대한민국 대학병원 진료실 PACS 의료정보시스템 컴퓨터 화면"
    ]
   ],
   "picked_url": "https://cdn.jhealthmedia.joins.com/news/photo/202108/24217_22724_1034.jpg",
   "sha256": "a639aa707f935ba6126ae1a2421ecd7eeb16d9c8bb6c79705f616c08cce9e1fb",
   "file": "eraref_483c9a13a91174d8.png"
  }
 },
 "S26sh1": {
  "input_fingerprint": "83ca1dd32197f1fb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 속 훼손된 부검 사진 위를 손가락으로 가리킨 채 입을 벌린 50대 한국인 남성 부검의의 상체\n\nLOCATION (lock): At the monitor workstation inside a medical-school research room resembling a hospital examination room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height on the monitor side, frame 전남대 부검의 in a medium three-quarter view on the right and the monitor on the left, both sharing the discussion plane. He leans toward the display with one arm raised, his finger indicating an indistinct area of the autopsy photograph while his attention follows downward to the detail he is explaining.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전남대 부검의 in the middle-right of the frame, midground, points to monitor display; monitor display in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모니터 화면 (Displaying autopsy photographs) — The active display faces partly toward camera and shows multiple autopsy photographs without emphasizing graphic detail; used as Shares the frame with the speaker and serves as the precise anchor for his pointing gesture; 연구실 진찰 공간 (In use for the consultation) — The visible room plane recedes behind the monitor and the seated discussion area; used as Establishes the room as a research space with the clinical feel stated in the scene.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination uses restrained neutral-to-cool color and naturalistic contrast, keeping the forensic imagery subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photographs remain displayed on the monitor while the medical examiner points out and explains the injuries.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전남대 부검의 (Korean 남성, 50대 중반, 타원형 얼굴, 짧은 검은색 머리) — wearing: 법의학자 및 부검의가 착용하는 흰색 실험실 가운. 안에는 일상적인 셔츠와 슬랙스를 받쳐 입는다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 모니터 화면의 문서 상단: \"부검보고서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 속 훼손된 부검 사진 위를 손가락으로 가리킨 채 입을 벌린 50대 한국인 남성 부검의의 상체\n\nLOCATION (lock): At the monitor workstation inside a medical-school research room resembling a hospital examination room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height on the monitor side, frame 전남대 부검의 in a medium three-quarter view on the right and the monitor on the left, both sharing the discussion plane. He leans toward the display with one arm raised, his finger indicating an indistinct area of the autopsy photograph while his attention follows downward to the detail he is explaining.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전남대 부검의 in the middle-right of the frame, midground, points to monitor display; monitor display in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모니터 화면 (Displaying autopsy photographs) — The active display faces partly toward camera and shows multiple autopsy photographs without emphasizing graphic detail; used as Shares the frame with the speaker and serves as the precise anchor for his pointing gesture; 연구실 진찰 공간 (In use for the consultation) — The visible room plane recedes behind the monitor and the seated discussion area; used as Establishes the room as a research space with the clinical feel stated in the scene.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination uses restrained neutral-to-cool color and naturalistic contrast, keeping the forensic imagery subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photographs remain displayed on the monitor while the medical examiner points out and explains the injuries.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전남대 부검의 (Korean 남성, 50대 중반, 타원형 얼굴, 짧은 검은색 머리) — wearing: 법의학자 및 부검의가 착용하는 흰색 실험실 가운. 안에는 일상적인 셔츠와 슬랙스를 받쳐 입는다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 모니터 화면의 문서 상단: \"부검보고서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 모니터 화면 속 훼손된 부검 사진 위를 손가락으로 가리킨 채 입을 벌린 50대 한국인 남성 부검의의 상체\n\nLOCATION (lock): At the monitor workstation inside a medical-school research room resembling a hospital examination room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height on the monitor side, frame 전남대 부검의 in a medium three-quarter view on the right and the monitor on the left, both sharing the discussion plane. He leans toward the display with one arm raised, his finger indicating an indistinct area of the autopsy photograph while his attention follows downward to the detail he is explaining.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전남대 부검의 in the middle-right of the frame, midground, points to monitor display; monitor display in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 모니터 화면 (Displaying autopsy photographs) — The active display faces partly toward camera and shows multiple autopsy photographs without emphasizing graphic detail; used as Shares the frame with the speaker and serves as the precise anchor for his pointing gesture; 연구실 진찰 공간 (In use for the consultation) — The visible room plane recedes behind the monitor and the seated discussion area; used as Establishes the room as a research space with the clinical feel stated in the scene.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination uses restrained neutral-to-cool color and naturalistic contrast, keeping the forensic imagery subdued.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photographs remain displayed on the monitor while the medical examiner points out and explains the injuries.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전남대 부검의 (Korean 남성, 50대 중반, 타원형 얼굴, 짧은 검은색 머리) — wearing: 법의학자 및 부검의가 착용하는 흰색 실험실 가운. 안에는 일상적인 셔츠와 슬랙스를 받쳐 입는다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 모니터 화면의 문서 상단: \"부검보고서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S26sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S26sh1.png"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_법의학_연구진찰실_510618.png"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "모니터 화면에 요구된 부검 사진과 '부검보고서' 텍스트를 정확히 묘사했으며, 인물의 외형과 행동 지시도 훌륭하게 반영함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물과 배경의 구현은 우수하나, 모니터 화면에 부검 사진이 아닌 작업실 CCTV 풍경이 출력되어 핵심 프롬프트를 벗어남."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "남성의 시선과 가리키는 손가락이 오른쪽 모니터의 부검 사진을 정확히 향하고 있음.",
      "built_space": "참조 이미지의 책상, 듀얼 모니터, 뒤편 책장 및 진찰용 침대가 올바른 위치에 구현됨.",
      "entities": "50대 한국인 남성, 흰색 가운과 셔츠, 입을 벌린 표정이 일치함. 모니터에 훼손된 신체가 담긴 부검 사진과 '부검보고서' 텍스트가 존재함.",
      "hard_violations": [],
      "physics": "의자에 앉아 체중이 실려 있으며, 들어 올린 팔과 손가락의 자세가 물리적으로 자연스러움."
     },
     {
      "label": "B",
      "direction": "남성의 시선과 손가락이 오른쪽 모니터 화면을 향하고 있음.",
      "built_space": "참조 이미지의 연구실 공간(책상, 모니터, 책장, 창문 등)이 정확한 비율과 위치로 묘사됨.",
      "entities": "인물의 외형과 복장은 프롬프트와 일치하나, 모니터 화면에 부검 사진이 아닌 사람들이 있는 작업실/CCTV 화면이 나타남.",
      "hard_violations": [],
      "physics": "앉은 자세와 모니터를 가리키기 위해 뻗은 팔의 무게 중심이 자연스럽게 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "모니터 화면에 요구된 부검 사진과 '부검보고서' 텍스트를 정확히 묘사했으며, 인물의 외형과 행동 지시도 훌륭하게 반영함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물과 배경의 구현은 우수하나, 모니터 화면에 부검 사진이 아닌 작업실 CCTV 풍경이 출력되어 핵심 프롬프트를 벗어남."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "남성의 시선과 가리키는 손가락이 오른쪽 모니터의 부검 사진을 정확히 향하고 있음.",
      "built_space": "참조 이미지의 책상, 듀얼 모니터, 뒤편 책장 및 진찰용 침대가 올바른 위치에 구현됨.",
      "entities": "50대 한국인 남성, 흰색 가운과 셔츠, 입을 벌린 표정이 일치함. 모니터에 훼손된 신체가 담긴 부검 사진과 '부검보고서' 텍스트가 존재함.",
      "hard_violations": [],
      "physics": "의자에 앉아 체중이 실려 있으며, 들어 올린 팔과 손가락의 자세가 물리적으로 자연스러움."
     },
     {
      "label": "B",
      "direction": "남성의 시선과 손가락이 오른쪽 모니터 화면을 향하고 있음.",
      "built_space": "참조 이미지의 연구실 공간(책상, 모니터, 책장, 창문 등)이 정확한 비율과 위치로 묘사됨.",
      "entities": "인물의 외형과 복장은 프롬프트와 일치하나, 모니터 화면에 부검 사진이 아닌 사람들이 있는 작업실/CCTV 화면이 나타남.",
      "hard_violations": [],
      "physics": "앉은 자세와 모니터를 가리키기 위해 뻗은 팔의 무게 중심이 자연스럽게 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "손상된 부검 사진이 띄워진 모니터를 가리키며 설명하는 부검의의 모습과 지시된 공간적 배경을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "모니터 화면에 지시된 부검 사진 대신 엉뚱한 작업자들의 사진이 생성되어 핵심 프롬프트 내용과 불일치합니다.  ★위반: [gemini-pro] 모니터 화면에 지시된 '훼손된 부검 사진' 대신 프롬프트에 없는 작업자들의 모습을 그려 넣음"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.25,
      "B": 1.714
     },
     "violations": {
      "A": [
       "[gemini-pro] 모니터 화면에 지시된 '훼손된 부검 사진' 대신 프롬프트에 없는 작업자들의 모습을 그려 넣음"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "손상된 부검 사진이 띄워진 모니터를 가리키며 설명하는 부검의의 모습과 지시된 공간적 배경을 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 1250,
      "verdict_ko": "모니터 화면에 지시된 부검 사진 대신 엉뚱한 작업자들의 사진이 생성되어 핵심 프롬프트 내용과 불일치합니다.  ★위반: [gemini-pro] 모니터 화면에 지시된 '훼손된 부검 사진' 대신 프롬프트에 없는 작업자들의 모습을 그려 넣음"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1721,
     "B": 1254
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "남성의 시선과 가리키는 손가락이 오른쪽 모니터의 부검 사진을 정확히 향하고 있음.",
    "built_space": "참조 이미지의 책상, 듀얼 모니터, 뒤편 책장 및 진찰용 침대가 올바른 위치에 구현됨.",
    "entities": "50대 한국인 남성, 흰색 가운과 셔츠, 입을 벌린 표정이 일치함. 모니터에 훼손된 신체가 담긴 부검 사진과 '부검보고서' 텍스트가 존재함.",
    "hard_violations": [],
    "physics": "의자에 앉아 체중이 실려 있으며, 들어 올린 팔과 손가락의 자세가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "남성의 시선과 손가락이 오른쪽 모니터 화면을 향하고 있음.",
    "built_space": "참조 이미지의 연구실 공간(책상, 모니터, 책장, 창문 등)이 정확한 비율과 위치로 묘사됨.",
    "entities": "인물의 외형과 복장은 프롬프트와 일치하나, 모니터 화면에 부검 사진이 아닌 사람들이 있는 작업실/CCTV 화면이 나타남.",
    "hard_violations": [],
    "physics": "앉은 자세와 모니터를 가리키기 위해 뻗은 팔의 무게 중심이 자연스럽게 지탱됨."
   }
  ],
  "totals": {
   "A": 1721,
   "B": 1254
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "모니터 화면에 요구된 부검 사진과 '부검보고서' 텍스트를 정확히 묘사했으며, 인물의 외형과 행동 지시도 훌륭하게 반영함."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "인물과 배경의 구현은 우수하나, 모니터 화면에 부검 사진이 아닌 작업실 CCTV 풍경이 출력되어 핵심 프롬프트를 벗어남."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_법의학_연구진찰실_510618.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "남성이 앉은 파란색 의자의 하단 기둥과 바퀴가 렌더링되지 않아 상판이 공중에 떠 있음.",
     "fix_en": "Render a central metal pillar and a five-point wheeled base connecting the bottom of the man's floating blue seat to the floor. Keep the man, his clothing, the monitors, and the entire room background completely unchanged.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "원본 배경 중앙에 있던 원형 테이블과 파란색 의자가 누락됨.",
     "fix_en": "Insert the missing round table and its blue stool into the midground floor area behind the man. Keep the man, the monitors, the desk, and the rest of the background intact.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "우측 모니터 상단의 '부검보고서' 텍스트의 글자가 알아볼 수 없게 깨져서 출력됨.",
     "fix_en": "Render the text area at the top of the right monitor slightly out of focus or obscured by screen glare so the garbled characters cannot be read. Keep the man, his hand, the autopsy images, and the monitors exactly as they are.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "남성이 앉은 파란색 의자의 하단 기둥과 바퀴가 렌더링되지 않아 상판이 공중에 떠 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "원본 배경 중앙에 있던 원형 테이블과 파란색 의자가 누락됨.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 모니터 상단의 '부검보고서' 텍스트의 글자가 알아볼 수 없게 깨져서 출력됨.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 참조 중앙에 있던 원형 테이블과 스툴이 프레임 한가운데에서 완전히 사라졌다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "원본 배경 중앙에 있던 원형 테이블과 파란색 의자가 누락됨.",
    "fix_en": "Insert the missing round table and its blue stool into the midground floor area behind the man. Keep the man, the monitors, the desk, and the rest of the background intact.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "우측 모니터 상단의 '부검보고서' 텍스트의 글자가 알아볼 수 없게 깨져서 출력됨.",
    "fix_en": "Render the text area at the top of the right monitor slightly out of focus or obscured by screen glare so the garbled characters cannot be read. Keep the man, his hand, the autopsy images, and the monitors exactly as they are.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Render a central metal pillar and a five-point wheeled base connecting the bottom of the man's floating blue seat to the floor. Keep the man, his clothing, the monitors, and the entire room background completely unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "요구된 인물의 외형, 복장, 배경, 자세를 정확히 구현했으며 지정된 텍스트 표기와 인체 해부학적 구조도 완벽하게 준수했습니다."
     },
     {
      "label": "A",
      "score": 1050,
      "verdict_ko": "전체적인 구도와 배경은 일치하나, 손가락이 6개로 묘사되는 치명적 해부학 오류와 텍스트 오타가 발생하여 탈락입니다.  ★위반: [gemini-pro] 손가락 개수가 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.3,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.05,
      "B": 1.714
     },
     "violations": {
      "A": [
       "[gemini-pro] 손가락 개수가 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "요구된 인물의 외형, 복장, 배경, 자세를 정확히 구현했으며 지정된 텍스트 표기와 인체 해부학적 구조도 완벽하게 준수했습니다."
     },
     {
      "label": "A",
      "score": 1050,
      "verdict_ko": "전체적인 구도와 배경은 일치하나, 손가락이 6개로 묘사되는 치명적 해부학 오류와 텍스트 오타가 발생하여 탈락입니다.  ★위반: [gemini-pro] 손가락 개수가 6개로 렌더링된 물리적으로 불가능한 해부학적 오류"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1750,
      "verdict_ko": "지시된 프레이밍과 인물의 자세를 정확히 구현했으며, 손의 해부학적 구조와 모니터 상단의 '부검보고서' 텍스트 묘사가 매우 자연스럽고 선명함."
     },
     {
      "label": "B",
      "score": 1778,
      "verdict_ko": "프롬프트의 전반적인 요구사항은 충족했으나, 모니터를 가리키는 오른손 손목이 해부학적으로 지나치게 얇고 납작하며 허벅지에 올린 왼손의 형태가 뭉개짐."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.778
     },
     "adjusted": {
      "A": 1.75,
      "B": 1.778
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "지시된 프레이밍과 인물의 자세를 정확히 구현했으며, 손의 해부학적 구조와 모니터 상단의 '부검보고서' 텍스트 묘사가 매우 자연스럽고 선명함."
     },
     {
      "label": "A",
      "score": 1778,
      "verdict_ko": "프롬프트의 전반적인 요구사항은 충족했으나, 모니터를 가리키는 오른손 손목이 해부학적으로 지나치게 얇고 납작하며 허벅지에 올린 왼손의 형태가 뭉개짐."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2828,
     "B": 3464
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S26sh1__bgfirst_bg.png",
   "bg_asset_id": "88b3665f-fab6-4531-8a48-57a034593ed1",
   "bg_record_key": "S26sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "법의학 연구진찰실",
   "groupbg_asset_id": "064c1a0c-3de6-4208-ab1d-c3dbf69519d5"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S26sh1::cine": {
  "applied": true,
  "fingerprint": "360be04264447e477bce70b6083a07d63a31cdc3f847638a8b570744790c8ace",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S26sh1_sel.png",
  "source_sha256": "73d4403de8e4084a899319f712d34c32473b83195495d6263903a15fd1c703a9",
  "file": "S26sh1_cine.png",
  "latency_ms": 9531
 },
 "S26sh6::signage": {
  "fp": "8affb8e2841000aa",
  "inscriptions": [
   {
    "surface_native": "명함",
    "text_native": "서울지방경찰청 경위 나상혁",
    "reason_ko": "부검의에게 건네는 명함에 신뢰감을 주는 소속과 직함, 이름을 표시하여 인물의 신분을 명확히 나타내기 위함입니다."
   }
  ]
 },
 "S26sh6": {
  "input_fingerprint": "a1b205a00ffa6469",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 부검의 쪽을 향해 하얀 명함 한 장을 깔끔하게 뻗은 나상혁의 손\n\nLOCATION (lock): Beside the desk and monitor inside the forensic medicine research room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height beside 서의용's position, track obliquely along 나상혁's extended arm into a close view of the single white business card held near the lower center, with the card remaining a modest portion of the frame. 나상혁 reaches cleanly toward 전남대 부검의 at frame right, whose upper body inclines toward the offered card, preserving both giver and recipient along the handoff axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 나상혁's extended hand in the lower-left of the frame, foreground, moves toward 전남대 부검의; 전남대 부검의 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 하얀 명함 한 장 (Extended toward the coroner) — One white face is visible at an oblique angle between 나상혁's fingers; its details are not presented frontally; used as The sharp focal point of the handoff, with the recipient providing human scale behind it; 모니터 화면 (Displaying autopsy photographs) — The display face remains visible only as a soft background plane beyond the handoff; used as Provides procedural context behind the hands without competing with the card.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime ambience gives the white card clean separation while retaining restrained, natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same clinical research-room look, monitor setup, cool daylight, and the examiner's position. Exclude the earlier pointing hand and show only the younger investigator extending a white business card.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok has taken one card from his business-card case and holds it out to the medical examiner.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 나상혁 right now — 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 명함: \"서울지방경찰청 경위 나상혁\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 부검의 쪽을 향해 하얀 명함 한 장을 깔끔하게 뻗은 나상혁의 손\n\nLOCATION (lock): Beside the desk and monitor inside the forensic medicine research room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height beside 서의용's position, track obliquely along 나상혁's extended arm into a close view of the single white business card held near the lower center, with the card remaining a modest portion of the frame. 나상혁 reaches cleanly toward 전남대 부검의 at frame right, whose upper body inclines toward the offered card, preserving both giver and recipient along the handoff axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 나상혁's extended hand in the lower-left of the frame, foreground, moves toward 전남대 부검의; 전남대 부검의 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 하얀 명함 한 장 (Extended toward the coroner) — One white face is visible at an oblique angle between 나상혁's fingers; its details are not presented frontally; used as The sharp focal point of the handoff, with the recipient providing human scale behind it; 모니터 화면 (Displaying autopsy photographs) — The display face remains visible only as a soft background plane beyond the handoff; used as Provides procedural context behind the hands without competing with the card.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime ambience gives the white card clean separation while retaining restrained, natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same clinical research-room look, monitor setup, cool daylight, and the examiner's position. Exclude the earlier pointing hand and show only the younger investigator extending a white business card.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok has taken one card from his business-card case and holds it out to the medical examiner.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 나상혁 right now — 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 명함: \"서울지방경찰청 경위 나상혁\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 부검의 쪽을 향해 하얀 명함 한 장을 깔끔하게 뻗은 나상혁의 손\n\nLOCATION (lock): Beside the desk and monitor inside the forensic medicine research room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height beside 서의용's position, track obliquely along 나상혁's extended arm into a close view of the single white business card held near the lower center, with the card remaining a modest portion of the frame. 나상혁 reaches cleanly toward 전남대 부검의 at frame right, whose upper body inclines toward the offered card, preserving both giver and recipient along the handoff axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 나상혁's extended hand in the lower-left of the frame, foreground, moves toward 전남대 부검의; 전남대 부검의 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 하얀 명함 한 장 (Extended toward the coroner) — One white face is visible at an oblique angle between 나상혁's fingers; its details are not presented frontally; used as The sharp focal point of the handoff, with the recipient providing human scale behind it; 모니터 화면 (Displaying autopsy photographs) — The display face remains visible only as a soft background plane beyond the handoff; used as Provides procedural context behind the hands without competing with the card.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime ambience gives the white card clean separation while retaining restrained, natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same clinical research-room look, monitor setup, cool daylight, and the examiner's position. Exclude the earlier pointing hand and show only the younger investigator extending a white business card.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok has taken one card from his business-card case and holds it out to the medical examiner.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 나상혁 right now — 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 명함: \"서울지방경찰청 경위 나상혁\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.571,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.429
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy (명함 우측 하단을 잡고 있는 연결되지 않은 여분의 손가락들)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1429,
   "A": 1250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1429,
    "verdict_ko": "손의 구조와 텍스트는 정확하게 구현되었으나, 명함 클로즈업 대신 인물의 상체와 얼굴을 포함하여 카메라 및 프레이밍 지시를 따르지 못함."
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "명함 우측 하단에 허공에 뜬 여분의 손가락이 생성되는 치명적인 해부학적 오류가 있으며, 명함을 비스듬히 배치하라는 지시도 어김.  ★위반: [gemini-pro] physically impossible anatomy (명함 우측 하단을 잡고 있는 연결되지 않은 여분의 손가락들)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S26sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프레임 우측의 부검의가 레퍼런스(이전 샷) 이미지 속 인물의 얼굴 및 복장(회색 셔츠, 흰색 가운)과 완전히 동일하게 렌더링되어, 이전 샷 인물 디자인 배제 지시를 위반함.",
     "fix_en": "Redraw the doctor on the right to be a different actor with a new face, hairstyle, and a different shirt (e.g., light blue) under his white coat, ensuring he does not resemble the man from the previous still. Preserve the framing, lighting, the background room, the monitors, the left investigator's face and clothing, and the hand holding the business card exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "명함이 수령인을 향한 비스듬한 각도가 아닌 카메라 렌즈를 향해 완전히 정면으로 배치되어, 명함의 디테일이 정면으로 제시되지 않아야 한다는 지시 및 소품 방향 규칙을 위반함.",
     "fix_en": "Redraw the business card to be angled obliquely toward the doctor on the right, turning its face away from the camera lens so the text is illegible and in shallow focus. Preserve the hand holding the card, both characters, the monitors, and the room.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프레임 우측의 부검의가 레퍼런스(이전 샷) 이미지 속 인물의 얼굴 및 복장(회색 셔츠, 흰색 가운)과 완전히 동일하게 렌더링되어, 이전 샷 인물 디자인 배제 지시를 위반함.",
     "severity": "critical"
    },
    {
     "issue_ko": "명함이 수령인을 향한 비스듬한 각도가 아닌 카메라 렌즈를 향해 완전히 정면으로 배치되어, 명함의 디테일이 정면으로 제시되지 않아야 한다는 지시 및 소품 방향 규칙을 위반함.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 스틸의 흰 가운 남성이 오른쪽 중경에 같은 얼굴·의상으로 등장했다.",
     "severity": "critical"
    },
    {
     "issue_ko": "명함이 비스듬하지 않고 거의 정면으로 잡혀 글자가 완전히 읽힌다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "명함이 수령인을 향한 비스듬한 각도가 아닌 카메라 렌즈를 향해 완전히 정면으로 배치되어, 명함의 디테일이 정면으로 제시되지 않아야 한다는 지시 및 소품 방향 규칙을 위반함.",
    "fix_en": "Redraw the business card to be angled obliquely toward the doctor on the right, turning its face away from the camera lens so the text is illegible and in shallow focus. Preserve the hand holding the card, both characters, the monitors, and the room.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the doctor on the right to be a different actor with a new face, hairstyle, and a different shirt (e.g., light blue) under his white coat, ensuring he does not resemble the man from the previous still. Preserve the framing, lighting, the background room, the monitors, the left investigator's face and clothing, and the hand holding the business card exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1350,
      "verdict_ko": "이전 샷의 연속성을 잘 유지하여 부검의의 복장(회색 셔츠)과 외모(안경 없음)를 그대로 재현했으며, 명함의 텍스트와 인물들의 구도도 프롬프트와 완벽하게 일치합니다.  ★위반: [openrouter:x-ai/grok-4.6] 이전 샷 인물의 얼굴과 의상을 부검의에게 그대로 옮겨 금지된 인물 재사용"
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "명함 텍스트와 전체적인 구도는 양호하나, 이전 샷과 달리 부검의가 안경을 쓰고 파란색 셔츠를 입고 있어 샷 연속성에서 감점 요인이 됩니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.6,
      "B": 1.571
     },
     "adjusted": {
      "A": 1.35,
      "B": 1.571
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 이전 샷 인물의 얼굴과 의상을 부검의에게 그대로 옮겨 금지된 인물 재사용"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.4,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1350,
      "verdict_ko": "이전 샷의 연속성을 잘 유지하여 부검의의 복장(회색 셔츠)과 외모(안경 없음)를 그대로 재현했으며, 명함의 텍스트와 인물들의 구도도 프롬프트와 완벽하게 일치합니다.  ★위반: [openrouter:x-ai/grok-4.6] 이전 샷 인물의 얼굴과 의상을 부검의에게 그대로 옮겨 금지된 인물 재사용"
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "명함 텍스트와 전체적인 구도는 양호하나, 이전 샷과 달리 부검의가 안경을 쓰고 파란색 셔츠를 입고 있어 샷 연속성에서 감점 요인이 됩니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "명함의 텍스트를 정확히 구현했으며, 이전 컷 레퍼런스의 부검의 얼굴과 의상을 완벽하게 유지하여 지시사항을 훌륭히 이행함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 텍스트는 양호하나, 이전 컷에 등장한 부검의의 신원(안경 추가, 얼굴 변경)과 의상을 임의로 변경하여 연속성을 크게 훼손함."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "나상혁의 손이 프레임 우측의 부검의를 향해 명함을 뻗고 있으며, 부검의의 시선이 명함에 안착함.",
      "built_space": "연구실 책상, 부검 사진이 띄워진 두 대의 모니터, 배경의 책장 등 이전 컷의 공간 배치를 일관되게 유지함.",
      "entities": "명함 텍스트('서울지방경찰청 경위 나상혁')가 정확함. 나상혁의 재킷과 부검의의 얼굴, 셔츠, 가운이 레퍼런스와 완벽히 일치함.",
      "hard_violations": [],
      "physics": "손가락이 물리적인 위화감 없이 명함을 안정적으로 쥐고 허공에 지탱함."
     },
     {
      "label": "A",
      "direction": "나상혁이 부검의 쪽으로 명함을 내밀고, 부검의가 명함을 바라보고 있음.",
      "built_space": "모니터와 책상 구조가 이전 컷의 배치와 동일하게 렌더링됨.",
      "entities": "명함 텍스트와 나상혁의 의상은 일치하나, 부검의의 얼굴이 변형되었고 안경이 추가되었으며 셔츠 색상이 달라짐.",
      "hard_violations": [
       "이전 컷에서 확립된 인물(부검의)의 신원 및 의상 임의 변경"
      ],
      "physics": "엄지와 다른 손가락들이 명함의 양면을 정상적으로 파지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "명함의 텍스트를 정확히 구현했으며, 이전 컷 레퍼런스의 부검의 얼굴과 의상을 완벽하게 유지하여 지시사항을 훌륭히 이행함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "구도와 텍스트는 양호하나, 이전 컷에 등장한 부검의의 신원(안경 추가, 얼굴 변경)과 의상을 임의로 변경하여 연속성을 크게 훼손함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "나상혁의 손이 프레임 우측의 부검의를 향해 명함을 뻗고 있으며, 부검의의 시선이 명함에 안착함.",
      "built_space": "연구실 책상, 부검 사진이 띄워진 두 대의 모니터, 배경의 책장 등 이전 컷의 공간 배치를 일관되게 유지함.",
      "entities": "명함 텍스트('서울지방경찰청 경위 나상혁')가 정확함. 나상혁의 재킷과 부검의의 얼굴, 셔츠, 가운이 레퍼런스와 완벽히 일치함.",
      "hard_violations": [],
      "physics": "손가락이 물리적인 위화감 없이 명함을 안정적으로 쥐고 허공에 지탱함."
     },
     {
      "label": "B",
      "direction": "나상혁이 부검의 쪽으로 명함을 내밀고, 부검의가 명함을 바라보고 있음.",
      "built_space": "모니터와 책상 구조가 이전 컷의 배치와 동일하게 렌더링됨.",
      "entities": "명함 텍스트와 나상혁의 의상은 일치하나, 부검의의 얼굴이 변형되었고 안경이 추가되었으며 셔츠 색상이 달라짐.",
      "hard_violations": [
       "이전 컷에서 확립된 인물(부검의)의 신원 및 의상 임의 변경"
      ],
      "physics": "엄지와 다른 손가락들이 명함의 양면을 정상적으로 파지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1358,
     "B": 1575
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S26sh1"
  }
 },
 "S26sh6::cine": {
  "applied": true,
  "fingerprint": "8680b77965cc43015bc6fed05387e002f1ad4ae85d205f1af206e5c1e8beb459",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S26sh6_sel.png",
  "source_sha256": "c18ebcc6786b60ea20241ae1bf22001b4dfb600193b0b9370e8bca016f018db6",
  "file": "S26sh6_cine.png",
  "latency_ms": 11939
 },
 "S27sh6::signage": {
  "fp": "e2a11414d7302ac3",
  "inscriptions": []
 },
 "S27sh6": {
  "input_fingerprint": "39e78b499d307c47",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소파 위 쓰러진 핸드백의 끈을 부자연스럽게 꺾인 손가락으로 쥐고 있는 민정의 손 클로즈업\n\nLOCATION (lock): At the sofa seating area inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above the sofa seat and close beside 민정 (현재), dolly in obliquely across the fallen handbag rather than looking straight down. Her hand and the sharply bent fingers gripping the strap occupy the central lower area, while part of the bag and sofa remain visible as scale references as she begins to set it upright.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 민정의 손과 꺾인 손가락 in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 핸드백 (Fallen on the sofa and being set upright) — Its side and upper opening angle across the sofa seat as 민정 starts to right it; used as Provides the object whose strap reveals 민정's unusual hand movement; 소파 좌석 (Occupied by 민정 and supporting the fallen handbag) — The seat plane runs diagonally beneath the handbag and 민정's hand; used as Forms the immediate background and keeps the hand action physically grounded.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime room ambience remains subdued and neutral-to-cool, with moderate contrast sufficient to reveal the hand shape without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minjeong is righting the fallen handbag on the sofa with her unnaturally bent hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 민정 right now — 민정's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 민정 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소파 위 쓰러진 핸드백의 끈을 부자연스럽게 꺾인 손가락으로 쥐고 있는 민정의 손 클로즈업\n\nLOCATION (lock): At the sofa seating area inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above the sofa seat and close beside 민정 (현재), dolly in obliquely across the fallen handbag rather than looking straight down. Her hand and the sharply bent fingers gripping the strap occupy the central lower area, while part of the bag and sofa remain visible as scale references as she begins to set it upright.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 민정의 손과 꺾인 손가락 in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 핸드백 (Fallen on the sofa and being set upright) — Its side and upper opening angle across the sofa seat as 민정 starts to right it; used as Provides the object whose strap reveals 민정's unusual hand movement; 소파 좌석 (Occupied by 민정 and supporting the fallen handbag) — The seat plane runs diagonally beneath the handbag and 민정's hand; used as Forms the immediate background and keeps the hand action physically grounded.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime room ambience remains subdued and neutral-to-cool, with moderate contrast sufficient to reveal the hand shape without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minjeong is righting the fallen handbag on the sofa with her unnaturally bent hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 민정 right now — 민정's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 민정 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소파 위 쓰러진 핸드백의 끈을 부자연스럽게 꺾인 손가락으로 쥐고 있는 민정의 손 클로즈업\n\nLOCATION (lock): At the sofa seating area inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above the sofa seat and close beside 민정 (현재), dolly in obliquely across the fallen handbag rather than looking straight down. Her hand and the sharply bent fingers gripping the strap occupy the central lower area, while part of the bag and sofa remain visible as scale references as she begins to set it upright.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 민정의 손과 꺾인 손가락 in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 핸드백 (Fallen on the sofa and being set upright) — Its side and upper opening angle across the sofa seat as 민정 starts to right it; used as Provides the object whose strap reveals 민정's unusual hand movement; 소파 좌석 (Occupied by 민정 and supporting the fallen handbag) — The seat plane runs diagonally beneath the handbag and 민정's hand; used as Forms the immediate background and keeps the hand action physically grounded.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime room ambience remains subdued and neutral-to-cool, with moderate contrast sufficient to reveal the hand shape without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Minjeong is righting the fallen handbag on the sofa with her unnaturally bent hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 민정 right now — 민정's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 민정 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "손이 소파 위에 쓰러진 검은색 핸드백의 상단 손잡이를 자연스럽게 쥐고 있습니다.",
    "built_space": "전경에 회색 소파가 있으며, 배경에 초점이 흐려진 목재 책상과 소품들이 보입니다.",
    "entities": "민정의 의상(진회색 코트와 짙은 파란색 스웨터 소매)과 일치하는 팔, 검은색 핸드백, 회색 소파가 등장합니다.",
    "hard_violations": [],
    "physics": "핸드백은 소파 위에 놓여 물리적으로 지지되고 있으며, 손은 핸드백의 손잡이를 안정적으로 잡고 있습니다."
   },
   {
    "label": "B",
    "direction": "창백하고 뻣뻣하게 꺾인 손가락을 가진 손이 소파 위에 쓰러진 가방의 끈을 쥐고 있습니다.",
    "built_space": "전경에 회색 소파가 있고, 배경에는 이전 숏 레퍼런스에서 볼 수 있는 목재 책상과 '수사 과장' 명패가 정확히 배치되어 있습니다.",
    "entities": "민정의 의상(진회색 코트와 짙은 파란색 스웨터 소매)과 일치하는 팔, 짙은 갈색 핸드백, 회색 소파, 수사 과장 명패가 등장합니다.",
    "hard_violations": [],
    "physics": "가방은 소파 표면에 눕혀져 지지되고 있으며, 손은 가방 끈 위에 놓인 채로 무게를 지탱하고 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 10,
   "A": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 10,
    "verdict_ko": "프롬프트가 요구한 '부자연스럽게 꺾인 손가락'의 형태를 매우 잘 구현했으며, 레퍼런스의 '수사 과장' 명패까지 배경에 정확히 재현하여 장소의 연속성을 완벽하게 살렸습니다."
   },
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "구도와 의상, 장소의 기본 조건은 충족했으나, 손가락이 가방을 쥐는 모습이 너무 자연스러워 '부자연스럽게 꺾인 손가락'이라는 핵심 지시사항을 묘사하는 데 실패했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S21sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 민정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:803682>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "손이 핸드백의 끈을 쥐고 있어야 한다는 프롬프트 지시와 달리, 손이 끈이 아닌 가방 본체를 잡고 있으며 끈은 가방 위에 늘어져 있음.",
     "fix_en": "Redraw the hand and handbag strap so that the hand is firmly gripping the strap instead of resting on the bag's body. Ensure the fingers wrap around the strap. Keep the grey coat sleeve, navy cuff, the handbag body, the sofa, and the background framing unchanged.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "손가락이 부자연스럽게 꺾이지 않고 자연스럽게 굽혀 있다",
     "fix_en": "Redraw the fingers of the hand to be contorted, bent at sharp, unnatural angles as if cramped or broken, while they hold the strap. Maintain the skin tone, the coat sleeve, the handbag, the sofa, and the background framing unchanged.",
     "severity": "critical",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "손이 핸드백의 끈을 쥐고 있어야 한다는 프롬프트 지시와 달리, 손이 끈이 아닌 가방 본체를 잡고 있으며 끈은 가방 위에 늘어져 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "손이 핸드백 끈을 쥐지 않고 가방 몸통 위에 얹혀 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "손가락이 부자연스럽게 꺾이지 않고 자연스럽게 굽혀 있다",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the hand and handbag strap so that the hand is firmly gripping the strap instead of resting on the bag's body. Ensure the fingers wrap around the strap. Keep the grey coat sleeve, navy cuff, the handbag body, the sofa, and the background framing unchanged.\n- Redraw the fingers of the hand to be contorted, bent at sharp, unnatural angles as if cramped or broken, while they hold the strap. Maintain the skin tone, the coat sleeve, the handbag, the sofa, and the background framing unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지문에 명시된 '부자연스럽게 꺾인 손가락'의 묘사가 누락되어 비교적 평범하게 가방 끈을 쥐고 있는 모습입니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "소파 좌측 쿠션 표면에 미세한 텍스처 패턴 오류(수정 비용)가 발생했으나, 프롬프트의 핵심 요구사항인 '부자연스럽게 꺾인 손가락'을 기형적으로 구부러진 검지를 통해 정확히 구현하여(수정 이점) 더 우수합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라가 소파 위 핸드백과 손을 비스듬히 내려다보며, 손은 가방 끈을 향해 있음.",
      "built_space": "수사과장실 내부 소파. 배경에 '수사과장' 명패가 놓인 책상이 위치함.",
      "entities": "민정의 식별(회색 코트, 남색 소매), 갈색 가죽 핸드백, 배경의 명패 모두 기준과 일치함.",
      "hard_violations": [],
      "physics": "가방은 소파 쿠션에 의해 지탱되며, 손은 가방 끈을 물리적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "카메라가 소파 위 핸드백과 손을 비스듬히 내려다보며, 손은 가방 끈을 향해 있음.",
      "built_space": "수사과장실 내부 소파. 배경에 '수사과장' 명패가 놓인 책상이 위치함.",
      "entities": "민정의 식별 및 복장, 핸드백, 명패가 일치하며, 꺾인 손가락이 묘사됨.",
      "hard_violations": [],
      "physics": "가방은 소파에 안착해 있고, 손은 가방을 지지대 삼아 끈을 잡고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지문에 명시된 '부자연스럽게 꺾인 손가락'의 묘사가 누락되어 비교적 평범하게 가방 끈을 쥐고 있는 모습입니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "소파 좌측 쿠션 표면에 미세한 텍스처 패턴 오류(수정 비용)가 발생했으나, 프롬프트의 핵심 요구사항인 '부자연스럽게 꺾인 손가락'을 기형적으로 구부러진 검지를 통해 정확히 구현하여(수정 이점) 더 우수합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라가 소파 위 핸드백과 손을 비스듬히 내려다보며, 손은 가방 끈을 향해 있음.",
      "built_space": "수사과장실 내부 소파. 배경에 '수사과장' 명패가 놓인 책상이 위치함.",
      "entities": "민정의 식별(회색 코트, 남색 소매), 갈색 가죽 핸드백, 배경의 명패 모두 기준과 일치함.",
      "hard_violations": [],
      "physics": "가방은 소파 쿠션에 의해 지탱되며, 손은 가방 끈을 물리적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "카메라가 소파 위 핸드백과 손을 비스듬히 내려다보며, 손은 가방 끈을 향해 있음.",
      "built_space": "수사과장실 내부 소파. 배경에 '수사과장' 명패가 놓인 책상이 위치함.",
      "entities": "민정의 식별 및 복장, 핸드백, 명패가 일치하며, 꺾인 손가락이 묘사됨.",
      "hard_violations": [],
      "physics": "가방은 소파에 안착해 있고, 손은 가방을 지지대 삼아 끈을 잡고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "A의 해부학적 기형(손가락 융합)을 성공적으로 수정하였으며, 부자연스럽게 꺾인 채 가방 끈을 쥐고 있는 손의 디테일과 화면 구도를 프롬프트에 맞게 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 1194,
      "verdict_ko": "집게손가락 끝이 가방 끈에 녹아들고 관절이 기형적으로 묘사되는 물리적/해부학적 오류(하드 위반)가 발생했습니다.  ★위반: [gemini-pro] physically impossible anatomy (집게손가락 끝이 가방 끈과 융합되어 뭉개짐)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.444,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.194,
      "B": 1.714
     },
     "violations": {
      "A": [
       "[gemini-pro] physically impossible anatomy (집게손가락 끝이 가방 끈과 융합되어 뭉개짐)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "A의 해부학적 기형(손가락 융합)을 성공적으로 수정하였으며, 부자연스럽게 꺾인 채 가방 끈을 쥐고 있는 손의 디테일과 화면 구도를 프롬프트에 맞게 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 1194,
      "verdict_ko": "집게손가락 끝이 가방 끈에 녹아들고 관절이 기형적으로 묘사되는 물리적/해부학적 오류(하드 위반)가 발생했습니다.  ★위반: [gemini-pro] physically impossible anatomy (집게손가락 끝이 가방 끈과 융합되어 뭉개짐)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1720,
     "B": 1203
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S21sh5"
  }
 },
 "S27sh6::cine": {
  "applied": true,
  "fingerprint": "f715014ae34c78d3a66502c727f9eb11277998eb6030c383e2454ff9cd362f94",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S27sh6_sel.png",
  "source_sha256": "f57e4ebcd12425be57b36f7f819431e0626f596a0bd170ee4f523918cb8375d7",
  "file": "S27sh6_cine.png",
  "latency_ms": 9335
 },
 "S27sh10::signage": {
  "fp": "ed8e85aa7a90cb46",
  "inscriptions": []
 },
 "S27sh10": {
  "input_fingerprint": "bc31ce4704c1e494",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈물이 맺힌 채 벌겋게 달아오른 얼굴로 전택수를 향해 손가락을 뻗은 심옥의 일그러진 표정\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office during the family interview. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Dolly in from just behind and to one side of 전택수's shoulder, slightly below 심옥's seated eye line, holding her face and pointing hand together in an off-center close composition. 심옥 leans forward from the sofa at frame left-center, tear-filled eyes fixed on 전택수 as her extended finger crosses toward his soft shoulder silhouette at the lower-right edge.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 심옥 in the middle-left of the frame, midground, points to 전택수; 전택수's shoulder in the lower-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 면담용 소파 (Occupied by 심옥) — The sofa back runs behind 심옥 while she pitches her upper body toward 전택수; used as Anchors 심옥's seated but forceful forward movement behind her face and arm.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime room illumination holds a neutral-to-cool palette while natural facial contrast reveals 심옥's tears and flushed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈물이 맺힌 채 벌겋게 달아오른 얼굴로 전택수를 향해 손가락을 뻗은 심옥의 일그러진 표정\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office during the family interview. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Dolly in from just behind and to one side of 전택수's shoulder, slightly below 심옥's seated eye line, holding her face and pointing hand together in an off-center close composition. 심옥 leans forward from the sofa at frame left-center, tear-filled eyes fixed on 전택수 as her extended finger crosses toward his soft shoulder silhouette at the lower-right edge.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 심옥 in the middle-left of the frame, midground, points to 전택수; 전택수's shoulder in the lower-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 면담용 소파 (Occupied by 심옥) — The sofa back runs behind 심옥 while she pitches her upper body toward 전택수; used as Anchors 심옥's seated but forceful forward movement behind her face and arm.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime room illumination holds a neutral-to-cool palette while natural facial contrast reveals 심옥's tears and flushed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 눈물이 맺힌 채 벌겋게 달아오른 얼굴로 전택수를 향해 손가락을 뻗은 심옥의 일그러진 표정\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office during the family interview. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Dolly in from just behind and to one side of 전택수's shoulder, slightly below 심옥's seated eye line, holding her face and pointing hand together in an off-center close composition. 심옥 leans forward from the sofa at frame left-center, tear-filled eyes fixed on 전택수 as her extended finger crosses toward his soft shoulder silhouette at the lower-right edge.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 심옥 in the middle-left of the frame, midground, points to 전택수; 전택수's shoulder in the lower-right of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 면담용 소파 (Occupied by 심옥) — The sofa back runs behind 심옥 while she pitches her upper body toward 전택수; used as Anchors 심옥's seated but forceful forward movement behind her face and arm.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued daytime room illumination holds a neutral-to-cool palette while natural facial contrast reveals 심옥's tears and flushed expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "심옥의 시선과 뻗은 손가락이 우측 하단의 전택수 어깨 실루엣을 정확히 향함.",
    "built_space": "이전 샷과 동일한 소파에 앉아 있으며, 배경의 책상과 '수사과장' 명패가 올바르게 배치됨.",
    "entities": "심옥의 인물 레퍼런스(외모, 베이지색 상의)가 일치하며, 눈물과 상기된 표정이 사실적으로 표현됨.",
    "hard_violations": [],
    "physics": "소파에 엉덩이를 대고 상체를 앞으로 기울인 채 손을 뻗은 자세가 자연스럽게 지탱됨."
   },
   {
    "label": "B",
    "direction": "심옥의 시선과 뻗은 손가락이 우측 하단의 전택수를 향하고 있음.",
    "built_space": "소파와 배경의 책상, 명패 등 공간적 요소가 적절히 묘사됨.",
    "entities": "심옥의 얼굴이 인위적이고 과장된 붉은색으로 렌더링되었으며, 레퍼런스와 달리 이전 샷의 회색 재킷을 착용함.",
    "hard_violations": [],
    "physics": "소파에 앉아 앞으로 몸을 숙이고 팔을 뻗은 자세가 물리적으로 안정감 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 8,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "레퍼런스의 의상 지시를 정확히 따랐으며, 자연스럽게 붉어진 얼굴과 눈물 등 감정 표현 및 지정된 카메라 앵글을 훌륭하게 구현함."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "얼굴이 비현실적으로 붉게 칠해져 사실성이 떨어지며, 이전 샷의 의상을 인물에게 적용하지 말라는 지시를 위반하고 회색 재킷을 입힘."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S27sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:866898>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "캐릭터 레퍼런스에 있는 심옥의 크로스백 가방 끈이 의상에서 누락되었습니다.",
     "fix_en": "Draw a brown cross-body bag strap diagonally across the woman's chest over her sweater, preserving her face, hand, sweater, sofa, and the foreground character.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "배경 책상 위 명패의 텍스트가 이전 컷 레퍼런스의 '수사 과장'과 달리 '추사 과장'으로 잘못 생성되었습니다.",
     "fix_en": "Blur the text on the wooden nameplate on the background desk so it is completely illegible, preserving the woman, her expression, the foreground character, and the room's lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "심옥의 얼굴에 흐르는 눈물이 실제 투명한 액체가 아닌 불투명한 흰색 선과 점(그래픽 오버레이)처럼 부자연스럽게 표현되었습니다.",
     "fix_en": "Replace the opaque white streaks on the woman's face with transparent, glossy tear tracks that pool naturally on her skin, preserving her exact facial features, expression, skin texture, sweater, and the background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "클로즈업이어야 할 구도가 심옥 상반신과 소파·가방·책상까지 보이는 미디엄 샷이다",
     "fix_en": "Darken the outer edges of the frame with a subtle vignette to focus attention on the woman's face and hand, preserving all characters, clothing, and background geometry.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전택수는 오른쪽 아래 가장자리 어깨 실루엣이어야 하는데 머리가 화면 오른쪽 전경을 크게 차지한다",
     "fix_en": "Darken and slightly blur the foreground character's head on the right to reduce its prominence, preserving the woman's expression, hand, clothing, and the background.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "캐릭터 레퍼런스에 있는 심옥의 크로스백 가방 끈이 의상에서 누락되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 책상 위 명패의 텍스트가 이전 컷 레퍼런스의 '수사 과장'과 달리 '추사 과장'으로 잘못 생성되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "심옥의 얼굴에 흐르는 눈물이 실제 투명한 액체가 아닌 불투명한 흰색 선과 점(그래픽 오버레이)처럼 부자연스럽게 표현되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "클로즈업이어야 할 구도가 심옥 상반신과 소파·가방·책상까지 보이는 미디엄 샷이다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수는 오른쪽 아래 가장자리 어깨 실루엣이어야 하는데 머리가 화면 오른쪽 전경을 크게 차지한다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "캐릭터 레퍼런스에 있는 심옥의 크로스백 가방 끈이 의상에서 누락되었습니다.",
    "fix_en": "Draw a brown cross-body bag strap diagonally across the woman's chest over her sweater, preserving her face, hand, sweater, sofa, and the foreground character.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "배경 책상 위 명패의 텍스트가 이전 컷 레퍼런스의 '수사 과장'과 달리 '추사 과장'으로 잘못 생성되었습니다.",
    "fix_en": "Blur the text on the wooden nameplate on the background desk so it is completely illegible, preserving the woman, her expression, the foreground character, and the room's lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "심옥의 얼굴에 흐르는 눈물이 실제 투명한 액체가 아닌 불투명한 흰색 선과 점(그래픽 오버레이)처럼 부자연스럽게 표현되었습니다.",
    "fix_en": "Replace the opaque white streaks on the woman's face with transparent, glossy tear tracks that pool naturally on her skin, preserving her exact facial features, expression, skin texture, sweater, and the background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "클로즈업이어야 할 구도가 심옥 상반신과 소파·가방·책상까지 보이는 미디엄 샷이다",
    "fix_en": "Darken the outer edges of the frame with a subtle vignette to focus attention on the woman's face and hand, preserving all characters, clothing, and background geometry.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "전택수는 오른쪽 아래 가장자리 어깨 실루엣이어야 하는데 머리가 화면 오른쪽 전경을 크게 차지한다",
    "fix_en": "Darken and slightly blur the foreground character's head on the right to reduce its prominence, preserving the woman's expression, hand, clothing, and the background.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh6"
  }
 },
 "S27sh10::cine": {
  "applied": true,
  "fingerprint": "4258e98a9a7d2b3c96115c48df31fb6005c8fd8110e302749a15309950cbf5fc",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S27sh10_sel.png",
  "source_sha256": "40a21c05c3a2ba17bedd50c95e94a97fd17d85c5017caa26f6f392f26f28ee9c",
  "file": "S27sh10_cine.png",
  "latency_ms": 10750
 },
 "S28sh1::signage": {
  "fp": "8a8b7952742c1222",
  "inscriptions": [
   {
    "surface_native": "연구소 입구 표지판",
    "text_native": "국립과학수사연구원 광주과학수사연구소",
    "reason_ko": "설정상 광주과학수사연구소 외경의 정체성을 명확하게 드러내는 진입로 표지판이 필요합니다."
   }
  ]
 },
 "S28sh1": {
  "input_fingerprint": "b723ff4305e95a10",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밝은 햇빛을 받는 웅장한 콘크리트 건물 외벽과 그 앞에 세워진 'NFS광주과학수사연구소' 팻말의 전경\n\nLOCATION (lock): In the sunlit exterior forecourt of the forensic science institute, facing its concrete facade and roadside sign. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The crane reaches its elevated end position and holds a wide three-quarter view from farther back, placing the full identification sign in the lower-left and the building frontage across the upper and middle frame. The diagonal separation between sign and building makes the institutional identity immediately legible without allowing the sign to overwhelm the architecture.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: NFS광주과학수사연구소 identification sign in the lower-left of the frame, foreground; research institute building frontage in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 'NFS광주과학수사연구소' 팻말 (Installed in front of the building) — The text-bearing face is fully visible to camera at a three-quarter angle, with the institution name readable; used as Foreground identification anchor connecting the exterior directly to the institution; 웅장한 콘크리트 건물 외벽 (Exterior frontage fully revealed) — The front facade and one receding side are visible in a wide three-quarter view; used as Occupies the upper portion of the frame as the institutional destination revealed by the crane movement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime sunlight defines the building and sign with a restrained neutral palette and moderate, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 연구소 입구 표지판: \"국립과학수사연구원 광주과학수사연구소\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밝은 햇빛을 받는 웅장한 콘크리트 건물 외벽과 그 앞에 세워진 'NFS광주과학수사연구소' 팻말의 전경\n\nLOCATION (lock): In the sunlit exterior forecourt of the forensic science institute, facing its concrete facade and roadside sign. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The crane reaches its elevated end position and holds a wide three-quarter view from farther back, placing the full identification sign in the lower-left and the building frontage across the upper and middle frame. The diagonal separation between sign and building makes the institutional identity immediately legible without allowing the sign to overwhelm the architecture.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: NFS광주과학수사연구소 identification sign in the lower-left of the frame, foreground; research institute building frontage in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 'NFS광주과학수사연구소' 팻말 (Installed in front of the building) — The text-bearing face is fully visible to camera at a three-quarter angle, with the institution name readable; used as Foreground identification anchor connecting the exterior directly to the institution; 웅장한 콘크리트 건물 외벽 (Exterior frontage fully revealed) — The front facade and one receding side are visible in a wide three-quarter view; used as Occupies the upper portion of the frame as the institutional destination revealed by the crane movement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime sunlight defines the building and sign with a restrained neutral palette and moderate, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 연구소 입구 표지판: \"국립과학수사연구원 광주과학수사연구소\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밝은 햇빛을 받는 웅장한 콘크리트 건물 외벽과 그 앞에 세워진 'NFS광주과학수사연구소' 팻말의 전경\n\nLOCATION (lock): In the sunlit exterior forecourt of the forensic science institute, facing its concrete facade and roadside sign. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The crane reaches its elevated end position and holds a wide three-quarter view from farther back, placing the full identification sign in the lower-left and the building frontage across the upper and middle frame. The diagonal separation between sign and building makes the institutional identity immediately legible without allowing the sign to overwhelm the architecture.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: NFS광주과학수사연구소 identification sign in the lower-left of the frame, foreground; research institute building frontage in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 'NFS광주과학수사연구소' 팻말 (Installed in front of the building) — The text-bearing face is fully visible to camera at a three-quarter angle, with the institution name readable; used as Foreground identification anchor connecting the exterior directly to the institution; 웅장한 콘크리트 건물 외벽 (Exterior frontage fully revealed) — The front facade and one receding side are visible in a wide three-quarter view; used as Occupies the upper portion of the frame as the institutional destination revealed by the crane movement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime sunlight defines the building and sign with a restrained neutral palette and moderate, naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 연구소 입구 표지판: \"국립과학수사연구원 광주과학수사연구소\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 콘크리트 건물 정면부와 앞마당의 표지판을 향함.",
    "built_space": "콘크리트 건물 외벽, 입구 계단과 경사로, 앞마당이 올바르게 배치됨.",
    "entities": "지시된 위치에 건물과 표지판이 있으나, 표지판에 중복된 한글 오타 및 영문 텍스트가 있음.",
    "hard_violations": [
     "표지판 텍스트에 지시되지 않은 영문 및 중복 오타 발생(텍스트 유출/생성 오류)"
    ],
    "physics": "모든 구조물과 표지판이 바닥에 안정적으로 고정됨."
   },
   {
    "label": "B",
    "direction": "카메라가 콘크리트 건물 정면부와 앞마당의 표지판을 향함.",
    "built_space": "콘크리트 건물 외벽, 입구 계단과 경사로, 앞마당이 올바르게 배치됨.",
    "entities": "지시된 위치에 건물과 표지판이 존재하며, 요구된 텍스트가 가독성 있게 렌더링됨(단, 하단에 샷 텍스트 일부가 추가됨).",
    "hard_violations": [],
    "physics": "모든 구조물과 표지판이 바닥에 안정적으로 고정됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 6,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "지정된 앵글과 구도(좌측 하단 표지판, 우측 상단 건물)를 정확히 구현했으며, 요구된 공식 명칭을 비교적 깔끔하게 렌더링함."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "건물과 구도는 지시사항을 잘 따랐으나, 표지판 텍스트에 중복 오타와 지시되지 않은 영문이 포함되어 지침을 위반함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L28B02.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "좌측 하단 간판에 지정된 텍스트 외에 프롬프트의 지시어인 'NFS광주과학수사연구소'가 맨 아랫줄에 추가로 출력되었습니다.",
     "fix_en": "Erase the bottom text line 'NFS광주과학수사연구소' from the foreground sign, leaving plain concrete. Preserve the top two text lines, the sign's shape, the building, the planters, the framing, and the lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "팻말 글자면이 3/4 각도가 아니라 카메라에 거의 정면이다",
     "fix_en": "Rotate the foreground sign to a three-quarter angle. Preserve the background building and lighting.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "왼쪽 팻말이 프레임에 잘려 팻말 전체가 담기지 않았다",
     "fix_en": "Fit the entire left edge of the sign within the frame. Preserve the background building and lighting.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "참조 장소 우측에 있는 깃대가 보이지 않는다",
     "fix_en": "Add tall metal flagpoles on the right side. Preserve the building, sign, and planters.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "건물 좌측 긴 세로창 날개가 참조 건축과 달리 거의 없다",
     "fix_en": "Add protruding concrete fins between the left windows. Preserve the main building structure and foreground sign.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "좌측 하단 간판에 지정된 텍스트 외에 프롬프트의 지시어인 'NFS광주과학수사연구소'가 맨 아랫줄에 추가로 출력되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "전경 팻말에 지정 문구 외에 ‘NFS광주과학수사연구소’가 추가로 적혀 있다",
     "severity": "major"
    },
    {
     "issue_ko": "팻말 글자면이 3/4 각도가 아니라 카메라에 거의 정면이다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 팻말이 프레임에 잘려 팻말 전체가 담기지 않았다",
     "severity": "major"
    },
    {
     "issue_ko": "참조 장소 우측에 있는 깃대가 보이지 않는다",
     "severity": "major"
    },
    {
     "issue_ko": "건물 좌측 긴 세로창 날개가 참조 건축과 달리 거의 없다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "팻말 글자면이 3/4 각도가 아니라 카메라에 거의 정면이다",
    "fix_en": "Rotate the foreground sign to a three-quarter angle. Preserve the background building and lighting.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "왼쪽 팻말이 프레임에 잘려 팻말 전체가 담기지 않았다",
    "fix_en": "Fit the entire left edge of the sign within the frame. Preserve the background building and lighting.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "참조 장소 우측에 있는 깃대가 보이지 않는다",
    "fix_en": "Add tall metal flagpoles on the right side. Preserve the building, sign, and planters.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "건물 좌측 긴 세로창 날개가 참조 건축과 달리 거의 없다",
    "fix_en": "Add protruding concrete fins between the left windows. Preserve the main building structure and foreground sign.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Erase the bottom text line 'NFS광주과학수사연구소' from the foreground sign, leaving plain concrete. Preserve the top two text lines, the sign's shape, the building, the planters, the framing, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 표지판 텍스트만을 정확하게 렌더링하고 불필요한 추가 텍스트 없이 지시사항을 완벽히 준수하여 우수함."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 텍스트 외에 금지된 추가 문구('NFS광주과학수사연구소')를 표지판에 렌더링하여 감점됨."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "카메라는 건물을 향해 3/4 각도로 조준됨.",
      "built_space": "참조 이미지의 콘크리트 외벽, 계단, 입구 구조 및 화단 배치가 정확히 일치함.",
      "entities": "좌측 하단 팻말에 지정된 텍스트만 렌더링됨. 인물 없음.",
      "hard_violations": [],
      "physics": "건물과 팻말 등 모든 구조물이 지면에 안정적으로 고정됨."
     },
     {
      "label": "A",
      "direction": "카메라는 건물을 향해 3/4 각도로 조준됨.",
      "built_space": "참조 이미지의 콘크리트 외벽, 계단, 입구 구조 및 화단 배치가 일치함.",
      "entities": "좌측 하단 팻말에 지정된 텍스트 외에 추가적인 줄이 렌더링됨. 인물 없음.",
      "hard_violations": [],
      "physics": "건물과 팻말 등 모든 구조물이 지면에 안정적으로 고정됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 표지판 텍스트만을 정확하게 렌더링하고 불필요한 추가 텍스트 없이 지시사항을 완벽히 준수하여 우수함."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 텍스트 외에 금지된 추가 문구('NFS광주과학수사연구소')를 표지판에 렌더링하여 감점됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "카메라는 건물을 향해 3/4 각도로 조준됨.",
      "built_space": "참조 이미지의 콘크리트 외벽, 계단, 입구 구조 및 화단 배치가 정확히 일치함.",
      "entities": "좌측 하단 팻말에 지정된 텍스트만 렌더링됨. 인물 없음.",
      "hard_violations": [],
      "physics": "건물과 팻말 등 모든 구조물이 지면에 안정적으로 고정됨."
     },
     {
      "label": "A",
      "direction": "카메라는 건물을 향해 3/4 각도로 조준됨.",
      "built_space": "참조 이미지의 콘크리트 외벽, 계단, 입구 구조 및 화단 배치가 일치함.",
      "entities": "좌측 하단 팻말에 지정된 텍스트 외에 추가적인 줄이 렌더링됨. 인물 없음.",
      "hard_violations": [],
      "physics": "건물과 팻말 등 모든 구조물이 지면에 안정적으로 고정됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 표지판 텍스트만 정확하게 렌더링하였으며, 건물과 햇빛의 묘사가 프롬프트의 지시를 잘 따름."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "지정된 텍스트 외에 샷 텍스트에 있던 문구를 표지판에 추가로 출력하여 '다른 텍스트 추가 금지' 규칙을 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "방향성을 지닌 인물이나 사물의 조준선이 존재하지 않음.",
      "built_space": "콘크리트 건물 외벽, 출입구, 계단 및 경사로가 참조 이미지와 일치하게 구성되었으며, 프레임 좌측 하단에 표지판이 올바르게 위치함.",
      "entities": "지시된 건물과 표지판이 존재하며, 표지판의 텍스트가 요청된 내용과 정확히 일치함.",
      "hard_violations": [],
      "physics": "건물과 표지판 등 모든 구조물이 지면에 안정적으로 배치되어 있으며, 햇빛에 의한 그림자가 자연스러움."
     },
     {
      "label": "B",
      "direction": "방향성을 지닌 인물이나 사물의 조준선이 존재하지 않음.",
      "built_space": "건물의 구조, 출입구, 계단, 좌측 하단의 표지판 배치가 A와 동일하게 올바름.",
      "entities": "건물과 표지판이 존재하나, 표지판에 지시되지 않은 'NFS광주과학수사연구소'라는 문구가 추가됨.",
      "hard_violations": [
       "지시되지 않은 샷 텍스트를 화면에 그대로 출력함 (표지판의 세 번째 줄 텍스트)"
      ],
      "physics": "모든 구조물이 지면에 안정적으로 고정되어 물리적 오류가 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지정된 표지판 텍스트만 정확하게 렌더링하였으며, 건물과 햇빛의 묘사가 프롬프트의 지시를 잘 따름."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 텍스트 외에 샷 텍스트에 있던 문구를 표지판에 추가로 출력하여 '다른 텍스트 추가 금지' 규칙을 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "방향성을 지닌 인물이나 사물의 조준선이 존재하지 않음.",
      "built_space": "콘크리트 건물 외벽, 출입구, 계단 및 경사로가 참조 이미지와 일치하게 구성되었으며, 프레임 좌측 하단에 표지판이 올바르게 위치함.",
      "entities": "지시된 건물과 표지판이 존재하며, 표지판의 텍스트가 요청된 내용과 정확히 일치함.",
      "hard_violations": [],
      "physics": "건물과 표지판 등 모든 구조물이 지면에 안정적으로 배치되어 있으며, 햇빛에 의한 그림자가 자연스러움."
     },
     {
      "label": "A",
      "direction": "방향성을 지닌 인물이나 사물의 조준선이 존재하지 않음.",
      "built_space": "건물의 구조, 출입구, 계단, 좌측 하단의 표지판 배치가 A와 동일하게 올바름.",
      "entities": "건물과 표지판이 존재하나, 표지판에 지시되지 않은 'NFS광주과학수사연구소'라는 문구가 추가됨.",
      "hard_violations": [
       "지시되지 않은 샷 텍스트를 화면에 그대로 출력함 (표지판의 세 번째 줄 텍스트)"
      ],
      "physics": "모든 구조물이 지면에 안정적으로 고정되어 물리적 오류가 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S28sh1::cine": {
  "applied": true,
  "fingerprint": "b678f69d8647d1397c9338a057c5b1190af13657306c838705125a89c1aae130",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S28sh1_sel.png",
  "source_sha256": "5f873dab21ed19cd1aea3c876dba6baae7ca4dfbb0562d4882490b29f2df82d1",
  "file": "S28sh1_cine.png",
  "latency_ms": 12470
 },
 "S29sh1::signage": {
  "fp": "1a204f4096faaf4a",
  "inscriptions": [
   {
    "surface_native": "부검소견서 서류 표지",
    "text_native": "부검소견서",
    "reason_ko": "테이블 위에 펼쳐진 핵심 소품인 부검소견서의 제목을 사실적으로 표시하여 장면에 리얼리티를 더하기 위함."
   },
   {
    "surface_native": "사건기록 파일 표지",
    "text_native": "사건기록",
    "reason_ko": "두 인물이 검토 중인 서류들이 사건 수사 기록임을 시각적으로 전달하기 위함."
   }
  ]
 },
 "era_assess::7bcca9169cadb7fe": {
  "subjects": [
   {
    "subject_native": "국립과학수사연구원 문서철 및 감정서 (2000년대-2010년대 대한민국)",
    "search_terms_native": [
     "국과수 감정서",
     "경찰 수사기록",
     "정부 문서철 황화일",
     "부검감정서"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 작성되어야 하며 타 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "2000년대 한국 관공서 및 국과수에서 사용하던 보고서 양식과 문서철(황화일)은 일반적인 서양식 마닐라 폴더와 형태, 직인, 텍스트 배치 면에서 크게 다릅니다."
   }
  ]
 },
 "era_fail::534c68ac1df7ac8c": {
  "stage": "research",
  "subject": "국립과학수사연구원 문서철 및 감정서 (2000년대-2010년대 대한민국)",
  "attempts": 4
 },
 "S29sh1::bgfirst_bg": {
  "input_fingerprint": "82ce961d3c02fa33",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블 위 펼쳐진 부검소견서와 사건기록들을 사이에 두고 마주 앉은 전택수와 조일남의 전경\n\nLOCATION (lock): At the sofa table inside a forensic institute research office, with reports and case files spread across it.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From one short end of the sofa table, hold a slightly elevated wide diagonal view that places 전택수 on the near-left side and 조일남 across at the far-right, both seated and facing each other. The autopsy report and case records form a central band between them, leading the eye along the table toward their measured exchange and leaving a clear route for the following dolly toward 전택수.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 조일남 in the middle-right of the frame, midground; sofa table with autopsy report and case records in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 소파 테이블 (Covered with the autopsy report and case records) — Its short end faces the camera and its long axis recedes diagonally between 전택수 and 조일남; used as Creates the central procedural barrier and visual connection between the two men; 부검소견서 (Spread open on the table) — Its text-bearing face lies upward toward both men and remains visible at an oblique angle to camera; used as Provides the evidentiary focus at the center of the conversation; 사건기록들 (Spread across the table) — Their covers and pages face upward in a loose spread between the two seated men; used as Extends the procedural context across the middle of the table; 마주 놓인 소파 좌석 (Occupied by 전택수 and 조일남) — The seats face one another across the table on opposing sides of the frame; used as Keeps both men grounded in the face-to-face consultation layout.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination is graded in subdued neutral-to-cool tones with moderate contrast and lightly documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블 위 펼쳐진 부검소견서와 사건기록들을 사이에 두고 마주 앉은 전택수와 조일남의 전경\n\nLOCATION (lock): At the sofa table inside a forensic institute research office, with reports and case files spread across it.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From one short end of the sofa table, hold a slightly elevated wide diagonal view that places 전택수 on the near-left side and 조일남 across at the far-right, both seated and facing each other. The autopsy report and case records form a central band between them, leading the eye along the table toward their measured exchange and leaving a clear route for the following dolly toward 전택수.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 조일남 in the middle-right of the frame, midground; sofa table with autopsy report and case records in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 소파 테이블 (Covered with the autopsy report and case records) — Its short end faces the camera and its long axis recedes diagonally between 전택수 and 조일남; used as Creates the central procedural barrier and visual connection between the two men; 부검소견서 (Spread open on the table) — Its text-bearing face lies upward toward both men and remains visible at an oblique angle to camera; used as Provides the evidentiary focus at the center of the conversation; 사건기록들 (Spread across the table) — Their covers and pages face upward in a loose spread between the two seated men; used as Extends the procedural context across the middle of the table; 마주 놓인 소파 좌석 (Occupied by 전택수 and 조일남) — The seats face one another across the table on opposing sides of the frame; used as Keeps both men grounded in the face-to-face consultation layout.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination is graded in subdued neutral-to-cool tones with moderate contrast and lightly documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S29sh1__bgfirst_bg.png",
  "asset_id": "f163a2f4-af15-4bab-bc33-537210d6f54f",
  "input_asset_ids": [
   "867fb129-3a06-4fb7-ac20-95f9e810b085",
   "45039ddf-84bc-4860-ba90-052abd9c4d84"
  ]
 },
 "S29sh1": {
  "input_fingerprint": "730091b77084f9d7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 펼쳐진 부검소견서와 사건기록들을 사이에 두고 마주 앉은 전택수와 조일남의 전경\n\nLOCATION (lock): At the sofa table inside a forensic institute research office, with reports and case files spread across it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From one short end of the sofa table, hold a slightly elevated wide diagonal view that places 전택수 on the near-left side and 조일남 across at the far-right, both seated and facing each other. The autopsy report and case records form a central band between them, leading the eye along the table toward their measured exchange and leaving a clear route for the following dolly toward 전택수.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 조일남 in the middle-right of the frame, midground; sofa table with autopsy report and case records in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 소파 테이블 (Covered with the autopsy report and case records) — Its short end faces the camera and its long axis recedes diagonally between 전택수 and 조일남; used as Creates the central procedural barrier and visual connection between the two men; 부검소견서 (Spread open on the table) — Its text-bearing face lies upward toward both men and remains visible at an oblique angle to camera; used as Provides the evidentiary focus at the center of the conversation; 사건기록들 (Spread across the table) — Their covers and pages face upward in a loose spread between the two seated men; used as Extends the procedural context across the middle of the table; 마주 놓인 소파 좌석 (Occupied by 전택수 and 조일남) — The seats face one another across the table on opposing sides of the frame; used as Keeps both men grounded in the face-to-face consultation layout.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination is graded in subdued neutral-to-cool tones with moderate contrast and lightly documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy report and case records remain spread across the table between Taeksu and Jo Ilnam.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 부검소견서 서류 표지: \"부검소견서\"\n- 사건기록 파일 표지: \"사건기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 펼쳐진 부검소견서와 사건기록들을 사이에 두고 마주 앉은 전택수와 조일남의 전경\n\nLOCATION (lock): At the sofa table inside a forensic institute research office, with reports and case files spread across it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From one short end of the sofa table, hold a slightly elevated wide diagonal view that places 전택수 on the near-left side and 조일남 across at the far-right, both seated and facing each other. The autopsy report and case records form a central band between them, leading the eye along the table toward their measured exchange and leaving a clear route for the following dolly toward 전택수.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 조일남 in the middle-right of the frame, midground; sofa table with autopsy report and case records in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 소파 테이블 (Covered with the autopsy report and case records) — Its short end faces the camera and its long axis recedes diagonally between 전택수 and 조일남; used as Creates the central procedural barrier and visual connection between the two men; 부검소견서 (Spread open on the table) — Its text-bearing face lies upward toward both men and remains visible at an oblique angle to camera; used as Provides the evidentiary focus at the center of the conversation; 사건기록들 (Spread across the table) — Their covers and pages face upward in a loose spread between the two seated men; used as Extends the procedural context across the middle of the table; 마주 놓인 소파 좌석 (Occupied by 전택수 and 조일남) — The seats face one another across the table on opposing sides of the frame; used as Keeps both men grounded in the face-to-face consultation layout.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination is graded in subdued neutral-to-cool tones with moderate contrast and lightly documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy report and case records remain spread across the table between Taeksu and Jo Ilnam.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 부검소견서 서류 표지: \"부검소견서\"\n- 사건기록 파일 표지: \"사건기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 펼쳐진 부검소견서와 사건기록들을 사이에 두고 마주 앉은 전택수와 조일남의 전경\n\nLOCATION (lock): At the sofa table inside a forensic institute research office, with reports and case files spread across it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From one short end of the sofa table, hold a slightly elevated wide diagonal view that places 전택수 on the near-left side and 조일남 across at the far-right, both seated and facing each other. The autopsy report and case records form a central band between them, leading the eye along the table toward their measured exchange and leaving a clear route for the following dolly toward 전택수.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 조일남 in the middle-right of the frame, midground; sofa table with autopsy report and case records in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 소파 테이블 (Covered with the autopsy report and case records) — Its short end faces the camera and its long axis recedes diagonally between 전택수 and 조일남; used as Creates the central procedural barrier and visual connection between the two men; 부검소견서 (Spread open on the table) — Its text-bearing face lies upward toward both men and remains visible at an oblique angle to camera; used as Provides the evidentiary focus at the center of the conversation; 사건기록들 (Spread across the table) — Their covers and pages face upward in a loose spread between the two seated men; used as Extends the procedural context across the middle of the table; 마주 놓인 소파 좌석 (Occupied by 전택수 and 조일남) — The seats face one another across the table on opposing sides of the frame; used as Keeps both men grounded in the face-to-face consultation layout.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient research-room illumination is graded in subdued neutral-to-cool tones with moderate contrast and lightly documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy report and case records remain spread across the table between Taeksu and Jo Ilnam.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 부검소견서 서류 표지: \"부검소견서\"\n- 사건기록 파일 표지: \"사건기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S29sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S29sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 조일남: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878922>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L28B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 조일남: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:878922>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 공간 레퍼런스의 가구와 환경을 완벽하게 재현했으며, 인물의 배치와 한글 텍스트 요구사항을 모두 충실히 이행함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물과 텍스트는 훌륭하게 구현되었으나, 공간 레퍼런스에 명시된 가구(소파, 테이블)의 재질과 색상을 완전히 다르게 묘사함."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "전택수와 조일남은 서로를 정면으로 마주보며 시선을 맞추고 있음.",
      "built_space": "레퍼런스와 일치하는 갈색 패브릭 소파와 나무 테이블이 배치된 연구실 내부. 전택수는 왼쪽, 조일남은 오른쪽 소파에 착석함.",
      "entities": "두 인물의 외형, 헤어스타일, 의상 모두 레퍼런스와 일치함. 테이블 위 서류에 '부검소견서', '사건기록' 문구가 정확하게 적혀 있음.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파에 체중을 싣고 안정적으로 앉아 있음. 테이블 위 서류들도 정상적으로 놓여 있음."
     },
     {
      "label": "A",
      "direction": "전택수와 조일남은 중앙 테이블을 사이에 두고 서로를 마주봄.",
      "built_space": "연구실 배경이나, 레퍼런스와 어긋나는 밝은 베이지색 소파와 회색 계열의 테이블이 배치됨.",
      "entities": "인물들의 외모와 복장은 레퍼런스와 잘 일치함. 서류 표지의 '부검소견서', '사건기록' 텍스트가 정확함.",
      "hard_violations": [
       "지정된 장소 레퍼런스의 가구(소파와 테이블) 재질 및 색상을 임의로 변경함 (공간 잠금 위반)"
      ],
      "physics": "두 인물은 소파에 정상적으로 착석해 있으며, 물리적으로 어색한 부분은 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 공간 레퍼런스의 가구와 환경을 완벽하게 재현했으며, 인물의 배치와 한글 텍스트 요구사항을 모두 충실히 이행함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물과 텍스트는 훌륭하게 구현되었으나, 공간 레퍼런스에 명시된 가구(소파, 테이블)의 재질과 색상을 완전히 다르게 묘사함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수와 조일남은 서로를 정면으로 마주보며 시선을 맞추고 있음.",
      "built_space": "레퍼런스와 일치하는 갈색 패브릭 소파와 나무 테이블이 배치된 연구실 내부. 전택수는 왼쪽, 조일남은 오른쪽 소파에 착석함.",
      "entities": "두 인물의 외형, 헤어스타일, 의상 모두 레퍼런스와 일치함. 테이블 위 서류에 '부검소견서', '사건기록' 문구가 정확하게 적혀 있음.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파에 체중을 싣고 안정적으로 앉아 있음. 테이블 위 서류들도 정상적으로 놓여 있음."
     },
     {
      "label": "A",
      "direction": "전택수와 조일남은 중앙 테이블을 사이에 두고 서로를 마주봄.",
      "built_space": "연구실 배경이나, 레퍼런스와 어긋나는 밝은 베이지색 소파와 회색 계열의 테이블이 배치됨.",
      "entities": "인물들의 외모와 복장은 레퍼런스와 잘 일치함. 서류 표지의 '부검소견서', '사건기록' 텍스트가 정확함.",
      "hard_violations": [
       "지정된 장소 레퍼런스의 가구(소파와 테이블) 재질 및 색상을 임의로 변경함 (공간 잠금 위반)"
      ],
      "physics": "두 인물은 소파에 정상적으로 착석해 있으며, 물리적으로 어색한 부분은 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 로케이션의 특징(나무 책장, 창문 위치 등)을 충실히 반영했으며, 인물의 외형과 서류의 텍스트 모두 정확하게 렌더링되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물과 서류 텍스트는 훌륭하게 구현되었으나, 지정된 목재 인테리어 대신 완전히 다른 회색톤의 모던한 배경이 생성되어 로케이션을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수와 조일남이 테이블을 사이에 두고 서로의 얼굴을 향해 시선을 맞추고 있습니다.",
      "built_space": "카메라가 테이블의 짧은 모서리 쪽에 위치해 있으며, 참조 이미지에 등장하는 왼쪽 창문과 오른쪽의 나무 소재 책장, 어두운 목재 테이블과 베이지색 소파가 정확히 배치되어 있습니다.",
      "entities": "전택수의 복장(모자, 남색 재킷)과 조일남의 복장(의사 가운)이 참조와 일치하며, 테이블 위 서류에 '부검소견서'와 '사건기록' 글씨가 정확히 적혀 있습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파에 안정적으로 앉아 있으며, 서류와 파일들은 테이블 표면에 정상적으로 얹혀 있습니다."
     },
     {
      "label": "B",
      "direction": "전택수와 조일남이 서로 마주보며 시선을 교환하고 있습니다.",
      "built_space": "테이블 짧은 면에서 바라보는 구도는 맞으나, 회색 철제 책장과 금속/유리 소재의 테이블 등 지정된 로케이션 사진과 완전히 다른 재질과 구조의 공간이 나타납니다.",
      "entities": "인물들의 얼굴과 복장이 참조 이미지와 잘 맞고, 서류 표지의 '부검소견서', '사건기록' 텍스트가 뚜렷하게 출력되었습니다.",
      "hard_violations": [],
      "physics": "인물들이 소파 위에 바르게 착석해 있고, 모든 서류는 테이블 위에 잘 지지되어 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 로케이션의 특징(나무 책장, 창문 위치 등)을 충실히 반영했으며, 인물의 외형과 서류의 텍스트 모두 정확하게 렌더링되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물과 서류 텍스트는 훌륭하게 구현되었으나, 지정된 목재 인테리어 대신 완전히 다른 회색톤의 모던한 배경이 생성되어 로케이션을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수와 조일남이 테이블을 사이에 두고 서로의 얼굴을 향해 시선을 맞추고 있습니다.",
      "built_space": "카메라가 테이블의 짧은 모서리 쪽에 위치해 있으며, 참조 이미지에 등장하는 왼쪽 창문과 오른쪽의 나무 소재 책장, 어두운 목재 테이블과 베이지색 소파가 정확히 배치되어 있습니다.",
      "entities": "전택수의 복장(모자, 남색 재킷)과 조일남의 복장(의사 가운)이 참조와 일치하며, 테이블 위 서류에 '부검소견서'와 '사건기록' 글씨가 정확히 적혀 있습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파에 안정적으로 앉아 있으며, 서류와 파일들은 테이블 표면에 정상적으로 얹혀 있습니다."
     },
     {
      "label": "A",
      "direction": "전택수와 조일남이 서로 마주보며 시선을 교환하고 있습니다.",
      "built_space": "테이블 짧은 면에서 바라보는 구도는 맞으나, 회색 철제 책장과 금속/유리 소재의 테이블 등 지정된 로케이션 사진과 완전히 다른 재질과 구조의 공간이 나타납니다.",
      "entities": "인물들의 얼굴과 복장이 참조 이미지와 잘 맞고, 서류 표지의 '부검소견서', '사건기록' 텍스트가 뚜렷하게 출력되었습니다.",
      "hard_violations": [],
      "physics": "인물들이 소파 위에 바르게 착석해 있고, 모든 서류는 테이블 위에 잘 지지되어 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "B",
    "direction": "전택수와 조일남은 서로를 정면으로 마주보며 시선을 맞추고 있음.",
    "built_space": "레퍼런스와 일치하는 갈색 패브릭 소파와 나무 테이블이 배치된 연구실 내부. 전택수는 왼쪽, 조일남은 오른쪽 소파에 착석함.",
    "entities": "두 인물의 외형, 헤어스타일, 의상 모두 레퍼런스와 일치함. 테이블 위 서류에 '부검소견서', '사건기록' 문구가 정확하게 적혀 있음.",
    "hard_violations": [],
    "physics": "두 인물 모두 소파에 체중을 싣고 안정적으로 앉아 있음. 테이블 위 서류들도 정상적으로 놓여 있음."
   },
   {
    "label": "A",
    "direction": "전택수와 조일남은 중앙 테이블을 사이에 두고 서로를 마주봄.",
    "built_space": "연구실 배경이나, 레퍼런스와 어긋나는 밝은 베이지색 소파와 회색 계열의 테이블이 배치됨.",
    "entities": "인물들의 외모와 복장은 레퍼런스와 잘 일치함. 서류 표지의 '부검소견서', '사건기록' 텍스트가 정확함.",
    "hard_violations": [
     "지정된 장소 레퍼런스의 가구(소파와 테이블) 재질 및 색상을 임의로 변경함 (공간 잠금 위반)"
    ],
    "physics": "두 인물은 소파에 정상적으로 착석해 있으며, 물리적으로 어색한 부분은 없음."
   }
  ],
  "totals": {
   "A": 8,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 공간 레퍼런스의 가구와 환경을 완벽하게 재현했으며, 인물의 배치와 한글 텍스트 요구사항을 모두 충실히 이행함."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "인물과 텍스트는 훌륭하게 구현되었으나, 공간 레퍼런스에 명시된 가구(소파, 테이블)의 재질과 색상을 완전히 다르게 묘사함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L28B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 조일남: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878922>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 최전경에 인물들이 공유하는 소파 테이블과 분리된 별도의 나무 표면이 임의로 생성되어, 그 위에 '부검소견서'와 '사건기록' 서류가 놓여 있습니다.",
     "fix_en": "Merge the foreground wooden surface into the main sofa table. Preserve the characters, their poses, clothing, room background, and lighting.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "최전경에 배치된 '부검소견서'와 '사건기록' 서류가 두 인물이 읽을 수 있는 방향이 아니라 카메라 렌즈를 향해 정면(가로)으로 놓여 있습니다.",
     "fix_en": "Rotate the foreground documents so their text faces the seated characters. Preserve the characters, their poses, clothing, room background, and lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "왼쪽 인물(전택수)의 재킷 우측 가슴 부위에 새겨진 영문 패치 글자가 뭉개져 알아볼 수 없습니다.",
     "fix_en": "Fix the blurred text on the left character's right chest patch. Preserve the characters, their poses, clothing, room background, and lighting.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "왼쪽 인물(전택수)이 무릎 위에 모으고 있는 두 손의 손가락 마디가 부자연스럽게 융합되어 있습니다.",
     "fix_en": "Redraw the left character's clasped hands to resolve the fused fingers. Preserve the characters, their poses, clothing, room background, and lighting.",
     "severity": "minor",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 최전경에 인물들이 공유하는 소파 테이블과 분리된 별도의 나무 표면이 임의로 생성되어, 그 위에 '부검소견서'와 '사건기록' 서류가 놓여 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "최전경에 배치된 '부검소견서'와 '사건기록' 서류가 두 인물이 읽을 수 있는 방향이 아니라 카메라 렌즈를 향해 정면(가로)으로 놓여 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 인물(전택수)의 재킷 우측 가슴 부위에 새겨진 영문 패치 글자가 뭉개져 알아볼 수 없습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "왼쪽 인물(전택수)이 무릎 위에 모으고 있는 두 손의 손가락 마디가 부자연스럽게 융합되어 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "전택수가 모자·수사 점퍼를 입고 있어 레퍼런스의 머리·복장 정체성과 맞지 않는다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 최전경에 인물들이 공유하는 소파 테이블과 분리된 별도의 나무 표면이 임의로 생성되어, 그 위에 '부검소견서'와 '사건기록' 서류가 놓여 있습니다.",
    "fix_en": "Merge the foreground wooden surface into the main sofa table. Preserve the characters, their poses, clothing, room background, and lighting.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "최전경에 배치된 '부검소견서'와 '사건기록' 서류가 두 인물이 읽을 수 있는 방향이 아니라 카메라 렌즈를 향해 정면(가로)으로 놓여 있습니다.",
    "fix_en": "Rotate the foreground documents so their text faces the seated characters. Preserve the characters, their poses, clothing, room background, and lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "왼쪽 인물(전택수)의 재킷 우측 가슴 부위에 새겨진 영문 패치 글자가 뭉개져 알아볼 수 없습니다.",
    "fix_en": "Fix the blurred text on the left character's right chest patch. Preserve the characters, their poses, clothing, room background, and lighting.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "왼쪽 인물(전택수)이 무릎 위에 모으고 있는 두 손의 손가락 마디가 부자연스럽게 융합되어 있습니다.",
    "fix_en": "Redraw the left character's clasped hands to resolve the fused fingers. Preserve the characters, their poses, clothing, room background, and lighting.",
    "severity": "minor",
    "observation_index": 3
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S29sh1__bgfirst_bg.png",
   "bg_asset_id": "f163a2f4-af15-4bab-bc33-537210d6f54f",
   "bg_record_key": "S29sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S29sh1::cine": {
  "applied": true,
  "fingerprint": "375f00045fb6620bc337824dfb42c7e82ffc69b3d506b57b7d66b230020b9d94",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S29sh1_sel.png",
  "source_sha256": "32da260cc053667d66275a612b50d4acd8463b6fd9d9af6be6b1ec92c94529f1",
  "file": "S29sh1_cine.png",
  "latency_ms": 10219
 },
 "S29sh4::signage": {
  "fp": "bf8cbfcd91b62022",
  "inscriptions": []
 },
 "S29sh4": {
  "input_fingerprint": "2d38e44ee3542712",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소파 등받이에 몸을 푹 기댄 채 시선을 허공으로 돌린 전택수의 상체\n\nLOCATION (lock): In the sofa seating area of the forensic institute research office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close beside 전택수's open side and slightly below his seated eye line, the tracking move settles into an off-center three-quarter upper-body frame as he sinks deeply against the sofa back. He occupies the left half while restrained negative space remains to the right in the direction of his distant, averted gaze; the change from the preceding moment is concentrated solely in the tighter camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 소파 등받이 (전택수가 몸을 깊이 기대고 있음) — Its inner seating side and upper edge face the camera behind 전택수; used as Supports 전택수's reclined posture and forms a firm line behind his shoulders; 부검소견서와 사건기록 (테이블 위에 펼쳐져 있음) — The printed page faces are visible obliquely at the lower edge of frame; used as A small lower-frame reminder of the investigation continuing while his attention drifts.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the laboratory, held in a subdued neutral-to-cool palette with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same research-office sofas, document-covered table, daylight, and institutional materials. Exclude the seated medical expert from the tighter composition and show the officer leaning back.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy report and case records remain spread on the table while Taeksu reclines on the sofa.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소파 등받이에 몸을 푹 기댄 채 시선을 허공으로 돌린 전택수의 상체\n\nLOCATION (lock): In the sofa seating area of the forensic institute research office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close beside 전택수's open side and slightly below his seated eye line, the tracking move settles into an off-center three-quarter upper-body frame as he sinks deeply against the sofa back. He occupies the left half while restrained negative space remains to the right in the direction of his distant, averted gaze; the change from the preceding moment is concentrated solely in the tighter camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 소파 등받이 (전택수가 몸을 깊이 기대고 있음) — Its inner seating side and upper edge face the camera behind 전택수; used as Supports 전택수's reclined posture and forms a firm line behind his shoulders; 부검소견서와 사건기록 (테이블 위에 펼쳐져 있음) — The printed page faces are visible obliquely at the lower edge of frame; used as A small lower-frame reminder of the investigation continuing while his attention drifts.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the laboratory, held in a subdued neutral-to-cool palette with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same research-office sofas, document-covered table, daylight, and institutional materials. Exclude the seated medical expert from the tighter composition and show the officer leaning back.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy report and case records remain spread on the table while Taeksu reclines on the sofa.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 소파 등받이에 몸을 푹 기댄 채 시선을 허공으로 돌린 전택수의 상체\n\nLOCATION (lock): In the sofa seating area of the forensic institute research office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close beside 전택수's open side and slightly below his seated eye line, the tracking move settles into an off-center three-quarter upper-body frame as he sinks deeply against the sofa back. He occupies the left half while restrained negative space remains to the right in the direction of his distant, averted gaze; the change from the preceding moment is concentrated solely in the tighter camera distance.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 소파 등받이 (전택수가 몸을 깊이 기대고 있음) — Its inner seating side and upper edge face the camera behind 전택수; used as Supports 전택수's reclined posture and forms a firm line behind his shoulders; 부검소견서와 사건기록 (테이블 위에 펼쳐져 있음) — The printed page faces are visible obliquely at the lower edge of frame; used as A small lower-frame reminder of the investigation continuing while his attention drifts.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient light appropriate to the laboratory, held in a subdued neutral-to-cool palette with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same research-office sofas, document-covered table, daylight, and institutional materials. Exclude the seated medical expert from the tighter composition and show the officer leaning back.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy report and case records remain spread on the table while Taeksu reclines on the sofa.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수의 시선이 화면 우측의 여백(허공)을 자연스럽게 향하고 있음.",
    "built_space": "기준 이미지와 동일한 창문, 책장, 소파, 테이블의 위치와 구조가 정확히 유지됨.",
    "entities": "전택수의 복장(모자, 재킷 등)이 일치하며, 하단 테이블에 '부검소견서', '사건기록' 텍스트가 명확히 보임.",
    "hard_violations": [],
    "physics": "소파 등받이에 상체의 체중을 싣고 기대어 있는 자세가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "전택수의 시선이 우측 방향의 허공을 향하고 있음.",
    "built_space": "소파, 뒤편 책장, 앞쪽 테이블 등 기준 이미지의 공간 구성 요소를 올바르게 유지함.",
    "entities": "기준 이미지와 동일한 인물 복장과 테이블 위 서류들의 텍스트가 정확하게 나타남.",
    "hard_violations": [],
    "physics": "엉덩이와 등을 소파에 지지한 채 앉아 있는 자세가 안정적으로 묘사됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 '상체 위주의 미디엄 샷'과 '더 가까워진 카메라 거리'를 정확히 구현하여 프레이밍 지침을 가장 잘 따름."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "상체 프레임 지시와 달리 다리까지 화면에 포함된 넓은 앵글로 촬영되어 샷 크기 지침을 충족하지 못함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S29sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "재킷 우측 가슴의 영문 텍스트가 레퍼런스와 다르게 'SALLKS'로 왜곡되어 표기됨.",
     "fix_en": "Change the distorted text on the right chest to 'SAUKS'. Preserve the character's face, jacket, lighting, and background.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "재킷 우측 소매 패치 하단의 글자가 의미를 알 수 없는 형태로 뭉개져 있음.",
     "fix_en": "Blur or remove the gibberish text on the sleeve patch. Keep the character's face, jacket, lighting, and background unchanged.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "전택수가 소파 등받이에 깊이 기대지 않고 등받이와 등 사이가 떠 있다.",
     "fix_en": "Shift the character's upper body back to rest firmly against the sofa backrest. Preserve his identity, clothing, the documents, and lighting.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "부검소견서·사건기록이 프레임 하단 가장자리에 비스듬히 보이지 않고 전경에 크게 정면으로 놓여 있다.",
     "fix_en": "Scale down the foreground documents to sit obliquely at the bottom edge, revealing the table. Keep the character, sofa, and background intact.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "재킷 우측 가슴의 영문 텍스트가 레퍼런스와 다르게 'SALLKS'로 왜곡되어 표기됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "재킷 우측 소매 패치 하단의 글자가 의미를 알 수 없는 형태로 뭉개져 있음.",
     "severity": "minor"
    },
    {
     "issue_ko": "전택수가 소파 등받이에 깊이 기대지 않고 등받이와 등 사이가 떠 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 열린 옆구리 쪽 아래가 아니라 거의 정면 옆에서 잡아 상체가 프레임 왼쪽 반을 차지하지 않는다.",
     "severity": "major"
    },
    {
     "issue_ko": "부검소견서·사건기록이 프레임 하단 가장자리에 비스듬히 보이지 않고 전경에 크게 정면으로 놓여 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "재킷 우측 가슴의 영문 텍스트가 레퍼런스와 다르게 'SALLKS'로 왜곡되어 표기됨.",
    "fix_en": "Change the distorted text on the right chest to 'SAUKS'. Preserve the character's face, jacket, lighting, and background.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "재킷 우측 소매 패치 하단의 글자가 의미를 알 수 없는 형태로 뭉개져 있음.",
    "fix_en": "Blur or remove the gibberish text on the sleeve patch. Keep the character's face, jacket, lighting, and background unchanged.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "전택수가 소파 등받이에 깊이 기대지 않고 등받이와 등 사이가 떠 있다.",
    "fix_en": "Shift the character's upper body back to rest firmly against the sofa backrest. Preserve his identity, clothing, the documents, and lighting.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "부검소견서·사건기록이 프레임 하단 가장자리에 비스듬히 보이지 않고 전경에 크게 정면으로 놓여 있다.",
    "fix_en": "Scale down the foreground documents to sit obliquely at the bottom edge, revealing the table. Keep the character, sofa, and background intact.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S29sh1"
  }
 },
 "S29sh4::cine": {
  "applied": true,
  "fingerprint": "f289f0c792f2d8b7d2480d20c9cfa6389f6679482fc2d9ba8d0b9d51585f815a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S29sh4_sel.png",
  "source_sha256": "3584db9d83ba6938a1a41d0d7eff74d40740c5d3b3dda42e9feeffc2399bce34",
  "file": "S29sh4_cine.png",
  "latency_ms": 9703
 },
 "S30sh2::signage": {
  "fp": "a2d57722c92902a7",
  "inscriptions": [
   {
    "surface_native": "사진 액자 속 인화지 하단",
    "text_native": "사랑하는 우리 가족",
    "reason_ko": "2000년대 한국 스튜디오 가족사진 하단에 흔히 인쇄되는 문구로, 과거의 행복했던 순간과 현재의 텅 빈 거실의 쓸쓸한 대비를 강조하기 위함입니다."
   }
  ]
 },
 "S30sh2": {
  "input_fingerprint": "543beb07626a2b97",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 거실 탁자 위에 놓인 단란한 가족사진 액자 클로즈업\n\nLOCATION (lock): Inside the now-lit but empty living room, at the small table holding the family photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At tabletop height beside the living-room table, the dolly-in ends on a shallow diagonal rather than square to the frame, keeping the photograph just under two-fifths of the image while the empty room remains legible around it. Within the picture, 은성 leans toward 전택수 and the adjacent spouse, with the family members' shoulders and gazes subtly turned toward one another rather than toward our lens; the tighter distance is the single emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 가족사진 액자 (거실 탁자 위에 놓여 있음) — The photograph-bearing front face is visible on a shallow diagonal, showing 은성 and 전택수 부부 together; used as Primary focal object approached along the table surface; 거실 탁자 (가족사진 액자가 놓여 있음) — Its top plane recedes diagonally from the lower foreground; used as Provides scale around the close photograph and preserves the surrounding absence; 거실 (텅 비어 있음); used as Remains softly readable beyond the photograph as the emotional counterpoint.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Switched-on interior illumination reveals the previously dark living room in a restrained neutral-to-cool nighttime palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The framed family photograph of Eunseong with Taeksu and his wife remains standing on the otherwise empty living-room table.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사진 액자 속 인화지 하단: \"사랑하는 우리 가족\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 거실 탁자 위에 놓인 단란한 가족사진 액자 클로즈업\n\nLOCATION (lock): Inside the now-lit but empty living room, at the small table holding the family photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At tabletop height beside the living-room table, the dolly-in ends on a shallow diagonal rather than square to the frame, keeping the photograph just under two-fifths of the image while the empty room remains legible around it. Within the picture, 은성 leans toward 전택수 and the adjacent spouse, with the family members' shoulders and gazes subtly turned toward one another rather than toward our lens; the tighter distance is the single emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 가족사진 액자 (거실 탁자 위에 놓여 있음) — The photograph-bearing front face is visible on a shallow diagonal, showing 은성 and 전택수 부부 together; used as Primary focal object approached along the table surface; 거실 탁자 (가족사진 액자가 놓여 있음) — Its top plane recedes diagonally from the lower foreground; used as Provides scale around the close photograph and preserves the surrounding absence; 거실 (텅 비어 있음); used as Remains softly readable beyond the photograph as the emotional counterpoint.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Switched-on interior illumination reveals the previously dark living room in a restrained neutral-to-cool nighttime palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The framed family photograph of Eunseong with Taeksu and his wife remains standing on the otherwise empty living-room table.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사진 액자 속 인화지 하단: \"사랑하는 우리 가족\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 텅 빈 거실 탁자 위에 놓인 단란한 가족사진 액자 클로즈업\n\nLOCATION (lock): Inside the now-lit but empty living room, at the small table holding the family photograph. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At tabletop height beside the living-room table, the dolly-in ends on a shallow diagonal rather than square to the frame, keeping the photograph just under two-fifths of the image while the empty room remains legible around it. Within the picture, 은성 leans toward 전택수 and the adjacent spouse, with the family members' shoulders and gazes subtly turned toward one another rather than toward our lens; the tighter distance is the single emphasized change.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 가족사진 액자 (거실 탁자 위에 놓여 있음) — The photograph-bearing front face is visible on a shallow diagonal, showing 은성 and 전택수 부부 together; used as Primary focal object approached along the table surface; 거실 탁자 (가족사진 액자가 놓여 있음) — Its top plane recedes diagonally from the lower foreground; used as Provides scale around the close photograph and preserves the surrounding absence; 거실 (텅 비어 있음); used as Remains softly readable beyond the photograph as the emotional counterpoint.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Switched-on interior illumination reveals the previously dark living room in a restrained neutral-to-cool nighttime palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The framed family photograph of Eunseong with Taeksu and his wife remains standing on the otherwise empty living-room table.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사진 액자 속 인화지 하단: \"사랑하는 우리 가족\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "사진 속 세 인물의 시선이 서로를 향하고 있으며, 렌즈를 바라보지 않음",
    "built_space": "텅 빈 거실 공간, 소파와 창문이 배경에 배치됨, 천장 조명이 켜져 있음",
    "entities": "탁자 위의 가족사진 액자, 액자 속 젊은 여성, 중년 남성, 중년 여성, 텍스트 '사랑하는 우리 가족' 일치함",
    "hard_violations": [],
    "physics": "액자가 탁자 위에 안정적으로 세워져 있음"
   },
   {
    "label": "B",
    "direction": "사진 속 세 인물의 시선이 서로를 향하고 있음",
    "built_space": "텅 빈 거실 공간, 배경에 소파와 켜진 스탠드 조명, 밤을 보여주는 창문",
    "entities": "탁자 위의 가족사진 액자, 액자 속 세 인물, 텍스트 '사랑하는 우리 가족'이 존재하나 글자 형태가 미세하게 어색함",
    "hard_violations": [],
    "physics": "액자가 탁자 위에 안정적으로 세워져 있음"
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 8,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "요구된 텍스트('사랑하는 우리 가족')가 인화지 하단에 정확히 렌더링되었으며, 차갑고 중립적인 밤 시간대의 거실 조명(한국식 천장등)과 대각선 구도가 지시사항을 훌륭하게 충족합니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "구도와 인물들의 시선 처리는 좋으나 텍스트의 형태가 약간 불안정하고, 배경 조명이 요구된 '차갑고 중립적인' 톤보다 다소 따뜻하게 표현되었습니다."
   }
  ],
  "refs": [],
  "critique": {
   "issues": [],
   "observer_observations": [],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 0,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_skipped": true,
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S30sh2::cine": {
  "applied": true,
  "fingerprint": "4d456199b74944b133087017dd32019a9a897e51c224403268c00bd18c78d158",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S30sh2_sel.png",
  "source_sha256": "bc152acdcf2a9e1e592493111d53247ba824c3a5922200b56522a50d2114c398",
  "file": "S30sh2_cine.png",
  "latency_ms": 11093
 },
 "S30sh4::signage": {
  "fp": "e099536515525f33",
  "inscriptions": []
 },
 "S30sh4": {
  "input_fingerprint": "d5b3cc3b46df14f1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): TV 앞 거실 테이블에 앉아 김치통을 앞에 둔 채 라면 면발을 젓가락으로 높이 들어 올린 전택수\n\nLOCATION (lock): At the low table in front of the television inside the living room, under the home's switched-on lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the living-room table at seated chest height, the dolly-in settles at a close-medium three-quarter angle to 전택수 and the television exactly as the suspended noodles reach their highest point. 전택수 sits off-center with his shoulders folded toward the meal, chopsticks raised above the table and his attention lowered to the noodles, while the kimchi container remains clearly separated in front of him; only the camera distance tightens from the preceding approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 거실 테이블 (전택수가 앞에 앉아 식사 중) — The near edge crosses the lower foreground and the top plane recedes toward 전택수; used as Forms the lower compositional plane beneath 전택수 and his solitary meal; 커다란 김치통 (전택수 앞에 놓여 있음) — Its upper and near side face the camera from the tabletop; used as Placed in front of 전택수 as a plain, practical accompaniment to the meal; 젓가락과 라면 면발 (면발을 가장 높이 들어 올린 순간) — The chopsticks angle upward from 전택수's hand, with the noodles hanging vertically over the table; used as Holds the lifted noodles at the peak of the action without obscuring 전택수's face; TV (거실 테이블 앞에 위치함) — Its front display face is directed toward 전택수; the display content is not specified; used as Establishes the direction in which the table and 전택수 are oriented.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The room's switched-on interior illumination maintains a subdued neutral-to-cool nighttime tone without glamorizing the solitary meal.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large kimchi container remains on the living-room table in front of Taeksu while he eats ramen.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): TV 앞 거실 테이블에 앉아 김치통을 앞에 둔 채 라면 면발을 젓가락으로 높이 들어 올린 전택수\n\nLOCATION (lock): At the low table in front of the television inside the living room, under the home's switched-on lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the living-room table at seated chest height, the dolly-in settles at a close-medium three-quarter angle to 전택수 and the television exactly as the suspended noodles reach their highest point. 전택수 sits off-center with his shoulders folded toward the meal, chopsticks raised above the table and his attention lowered to the noodles, while the kimchi container remains clearly separated in front of him; only the camera distance tightens from the preceding approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 거실 테이블 (전택수가 앞에 앉아 식사 중) — The near edge crosses the lower foreground and the top plane recedes toward 전택수; used as Forms the lower compositional plane beneath 전택수 and his solitary meal; 커다란 김치통 (전택수 앞에 놓여 있음) — Its upper and near side face the camera from the tabletop; used as Placed in front of 전택수 as a plain, practical accompaniment to the meal; 젓가락과 라면 면발 (면발을 가장 높이 들어 올린 순간) — The chopsticks angle upward from 전택수's hand, with the noodles hanging vertically over the table; used as Holds the lifted noodles at the peak of the action without obscuring 전택수's face; TV (거실 테이블 앞에 위치함) — Its front display face is directed toward 전택수; the display content is not specified; used as Establishes the direction in which the table and 전택수 are oriented.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The room's switched-on interior illumination maintains a subdued neutral-to-cool nighttime tone without glamorizing the solitary meal.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large kimchi container remains on the living-room table in front of Taeksu while he eats ramen.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): TV 앞 거실 테이블에 앉아 김치통을 앞에 둔 채 라면 면발을 젓가락으로 높이 들어 올린 전택수\n\nLOCATION (lock): At the low table in front of the television inside the living room, under the home's switched-on lighting. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the living-room table at seated chest height, the dolly-in settles at a close-medium three-quarter angle to 전택수 and the television exactly as the suspended noodles reach their highest point. 전택수 sits off-center with his shoulders folded toward the meal, chopsticks raised above the table and his attention lowered to the noodles, while the kimchi container remains clearly separated in front of him; only the camera distance tightens from the preceding approach.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 거실 테이블 (전택수가 앞에 앉아 식사 중) — The near edge crosses the lower foreground and the top plane recedes toward 전택수; used as Forms the lower compositional plane beneath 전택수 and his solitary meal; 커다란 김치통 (전택수 앞에 놓여 있음) — Its upper and near side face the camera from the tabletop; used as Placed in front of 전택수 as a plain, practical accompaniment to the meal; 젓가락과 라면 면발 (면발을 가장 높이 들어 올린 순간) — The chopsticks angle upward from 전택수's hand, with the noodles hanging vertically over the table; used as Holds the lifted noodles at the peak of the action without obscuring 전택수's face; TV (거실 테이블 앞에 위치함) — Its front display face is directed toward 전택수; the display content is not specified; used as Establishes the direction in which the table and 전택수 are oriented.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The room's switched-on interior illumination maintains a subdued neutral-to-cool nighttime tone without glamorizing the solitary meal.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The large kimchi container remains on the living-room table in front of Taeksu while he eats ramen.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.071,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gemini-pro] physically impossible anatomy (손가락과 젓가락 융합)",
     "[gemini-pro] leaked markers/diagrams/text (김치통 라벨에 '수사' 표기)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1667,
   "A": 1071
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "손의 해부학적 구조가 안정적이고 모자 디테일을 잘 살렸으며, 구도와 동작을 프롬프트에 맞게 충실히 구현함."
   },
   {
    "label": "A",
    "score": 1071,
    "verdict_ko": "젓가락을 쥔 손의 구조적 오류(치명적 결함)가 있으며, 김치통에 의도치 않은 텍스트가 유출됨.  ★위반: [gemini-pro] physically impossible anatomy (손가락과 젓가락 융합) / [gemini-pro] leaked markers/diagrams/text (김치통 라벨에 '수사' 표기)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S30sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "라면 면발이 젓가락 끝에 집혀 있지 않고 젓가락 중간 부분을 관통함.",
     "fix_en": "Redraw the chopsticks so their tips properly pinch the top of the suspended noodles, replacing the middle section where noodles incorrectly pass through with solid wooden chopstick shafts. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "젓가락이 위를 향해 들려 있어야 한다는 지시와 달리 손에서 수평 방향으로 뻗어 있음.",
     "fix_en": "Redraw the man's right hand and the chopsticks so the chopsticks angle firmly upward toward the suspended noodles rather than extending horizontally. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "김치통 라벨과 재킷 소매 패치의 텍스트가 무의미한 글자로 뭉개짐.",
     "fix_en": "Replace the garbled text on the container label and the jacket's sleeve patch with generic, illegible textural markings or a blank surface. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "라면이 김치통과 분리되지 않고 같은 통에서 들어 올려져 김치통이 앞에 따로 놓인 식사가 아니다.",
     "fix_en": "Redraw the single container into a steaming metal ramen pot holding the noodles, and place a separate rectangular plastic kimchi container on the table nearby. Preserve the people present and their positions, their clothing, the suspended noodles, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "전택수가 50대 중반·각진 얼굴의 지정 인물이 아니라 훨씬 젊고 날씬한 다른 얼굴이다.",
     "fix_en": "Replace the man's face to match the reference actor, giving him a wider, angular jaw and mature 50s features while maintaining his downward gaze. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 4
    },
    {
     "issue_ko": "이전 샷에 테이블 전면에 있던 가족 액자가 사라지고 대신 라면·김치 통이 그 자리에 있다.",
     "fix_en": "Draw the framed family photo from the reference onto the table to the left, facing the camera. Preserve the people present and their positions, their clothing, the meal, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "라면 면발이 젓가락 끝에 집혀 있지 않고 젓가락 중간 부분을 관통함.",
     "severity": "major"
    },
    {
     "issue_ko": "젓가락이 위를 향해 들려 있어야 한다는 지시와 달리 손에서 수평 방향으로 뻗어 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "김치통 라벨과 재킷 소매 패치의 텍스트가 무의미한 글자로 뭉개짐.",
     "severity": "major"
    },
    {
     "issue_ko": "라면이 김치통과 분리되지 않고 같은 통에서 들어 올려져 김치통이 앞에 따로 놓인 식사가 아니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수가 50대 중반·각진 얼굴의 지정 인물이 아니라 훨씬 젊고 날씬한 다른 얼굴이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 샷에 테이블 전면에 있던 가족 액자가 사라지고 대신 라면·김치 통이 그 자리에 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "라면 면발이 젓가락 끝에 집혀 있지 않고 젓가락 중간 부분을 관통함.",
    "fix_en": "Redraw the chopsticks so their tips properly pinch the top of the suspended noodles, replacing the middle section where noodles incorrectly pass through with solid wooden chopstick shafts. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "젓가락이 위를 향해 들려 있어야 한다는 지시와 달리 손에서 수평 방향으로 뻗어 있음.",
    "fix_en": "Redraw the man's right hand and the chopsticks so the chopsticks angle firmly upward toward the suspended noodles rather than extending horizontally. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "김치통 라벨과 재킷 소매 패치의 텍스트가 무의미한 글자로 뭉개짐.",
    "fix_en": "Replace the garbled text on the container label and the jacket's sleeve patch with generic, illegible textural markings or a blank surface. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "이전 샷에 테이블 전면에 있던 가족 액자가 사라지고 대신 라면·김치 통이 그 자리에 있다.",
    "fix_en": "Draw the framed family photo from the reference onto the table to the left, facing the camera. Preserve the people present and their positions, their clothing, the meal, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the single container into a steaming metal ramen pot holding the noodles, and place a separate rectangular plastic kimchi container on the table nearby. Preserve the people present and their positions, their clothing, the suspended noodles, the set, the light, and the framing.\n- Replace the man's face to match the reference actor, giving him a wider, angular jaw and mature 50s features while maintaining his downward gaze. Preserve the people present and their positions, their clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "전택수의 외모와 배경을 레퍼런스에 맞게 잘 재현했으며, 라면 냄비와 별도로 놓인 김치통을 명확히 분리하여 프롬프트의 세부 지시사항을 훌륭히 이행함."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "인물과 배경의 일치도는 높으나, 김치통이 식사와 명확히 분리되어야 한다는 지시를 어기고 플라스틱 김치통 안에서 라면을 들어 올리는 형태로 연출됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 젓가락으로 들어 올린 라면 면발을 정확히 향하고 있음.",
      "built_space": "거실의 조명, 소파, TV, 창문 등 배경 요소들이 레퍼런스 이미지와 동일한 위치와 구도로 배치됨.",
      "entities": "전택수의 얼굴, 모자, 점퍼 등은 레퍼런스와 일치하나, 테이블 위에는 별도의 식사 용기 없이 플라스틱 김치통에 라면이 담겨 있음.",
      "hard_violations": [],
      "physics": "손이 젓가락을 잡고 있으며, 젓가락이 면발을 들어 올린 상태로 공중에 잘 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선이 들어 올린 라면 면발을 자연스럽게 응시하고 있음.",
      "built_space": "TV, 수납장, 소파가 놓인 거실 배경이 레퍼런스와 정확히 일치하며 올바른 원근감을 가짐.",
      "entities": "전택수의 인상과 복장이 레퍼런스와 완벽히 일치하며, 지시된 대로 라면이 담긴 금속 냄비와 플라스틱 김치통이 명확히 분리되어 배치됨.",
      "hard_violations": [],
      "physics": "젓가락을 쥔 손의 파지가 자연스럽고, 들어 올려진 라면 면발의 무게감과 형태가 물리적으로 타당함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "전택수의 외모와 배경을 레퍼런스에 맞게 잘 재현했으며, 라면 냄비와 별도로 놓인 김치통을 명확히 분리하여 프롬프트의 세부 지시사항을 훌륭히 이행함."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "인물과 배경의 일치도는 높으나, 김치통이 식사와 명확히 분리되어야 한다는 지시를 어기고 플라스틱 김치통 안에서 라면을 들어 올리는 형태로 연출됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 젓가락으로 들어 올린 라면 면발을 정확히 향하고 있음.",
      "built_space": "거실의 조명, 소파, TV, 창문 등 배경 요소들이 레퍼런스 이미지와 동일한 위치와 구도로 배치됨.",
      "entities": "전택수의 얼굴, 모자, 점퍼 등은 레퍼런스와 일치하나, 테이블 위에는 별도의 식사 용기 없이 플라스틱 김치통에 라면이 담겨 있음.",
      "hard_violations": [],
      "physics": "손이 젓가락을 잡고 있으며, 젓가락이 면발을 들어 올린 상태로 공중에 잘 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선이 들어 올린 라면 면발을 자연스럽게 응시하고 있음.",
      "built_space": "TV, 수납장, 소파가 놓인 거실 배경이 레퍼런스와 정확히 일치하며 올바른 원근감을 가짐.",
      "entities": "전택수의 인상과 복장이 레퍼런스와 완벽히 일치하며, 지시된 대로 라면이 담긴 금속 냄비와 플라스틱 김치통이 명확히 분리되어 배치됨.",
      "hard_violations": [],
      "physics": "젓가락을 쥔 손의 파지가 자연스럽고, 들어 올려진 라면 면발의 무게감과 형태가 물리적으로 타당함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "라면 식사와 김치통을 명확히 분리하여 배치하라는 지시를 잘 이행했으며, 캐릭터와 배경의 일치도가 매우 높습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "라면 식사와 김치통이 하나의 용기로 합쳐져 있어, 두 요소가 분리되어야 한다는 프롬프트의 지시를 어겼습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 젓가락으로 들어 올린 라면 면발을 정확히 향하고 있음.",
      "built_space": "기준 이미지의 거실 배경(테이블, TV, 소파, 창문, 조명)을 정확하게 재현함.",
      "entities": "전택수의 외모, 연령대, 복장(수사 모자, 재킷)이 기준과 일치함. 라면이 담긴 냄비와 별도의 김치통이 테이블 위에 올바르게 구현됨.",
      "hard_violations": [],
      "physics": "오른손이 젓가락을 자연스럽게 쥐고 있으며, 젓가락이 중력을 거슬러 면발을 들어 올리는 물리적 상태가 타당함."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 젓가락으로 들어 올린 라면 면발을 향하고 있음.",
      "built_space": "기준 이미지의 거실 배경(테이블, TV, 소파, 창문)을 잘 반영함.",
      "entities": "전택수의 인물 특징과 복장은 기준에 부합하나, 라면과 김치통이 별개로 존재하지 않고 하나의 플라스틱 용기로 묘사됨.",
      "hard_violations": [],
      "physics": "손의 형태와 젓가락을 쥐는 동작, 면발이 들어 올려진 형태의 물리적 지지가 정상적임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "라면 식사와 김치통을 명확히 분리하여 배치하라는 지시를 잘 이행했으며, 캐릭터와 배경의 일치도가 매우 높습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "라면 식사와 김치통이 하나의 용기로 합쳐져 있어, 두 요소가 분리되어야 한다는 프롬프트의 지시를 어겼습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수의 시선은 젓가락으로 들어 올린 라면 면발을 정확히 향하고 있음.",
      "built_space": "기준 이미지의 거실 배경(테이블, TV, 소파, 창문, 조명)을 정확하게 재현함.",
      "entities": "전택수의 외모, 연령대, 복장(수사 모자, 재킷)이 기준과 일치함. 라면이 담긴 냄비와 별도의 김치통이 테이블 위에 올바르게 구현됨.",
      "hard_violations": [],
      "physics": "오른손이 젓가락을 자연스럽게 쥐고 있으며, 젓가락이 중력을 거슬러 면발을 들어 올리는 물리적 상태가 타당함."
     },
     {
      "label": "A",
      "direction": "전택수의 시선은 젓가락으로 들어 올린 라면 면발을 향하고 있음.",
      "built_space": "기준 이미지의 거실 배경(테이블, TV, 소파, 창문)을 잘 반영함.",
      "entities": "전택수의 인물 특징과 복장은 기준에 부합하나, 라면과 김치통이 별개로 존재하지 않고 하나의 플라스틱 용기로 묘사됨.",
      "hard_violations": [],
      "physics": "손의 형태와 젓가락을 쥐는 동작, 면발이 들어 올려진 형태의 물리적 지지가 정상적임."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S30sh2"
  }
 },
 "S30sh4::cine": {
  "applied": true,
  "fingerprint": "72f9be5a904c6502e13c859c6ee7a2d3f6fa5cfcca364e9b120c973e29876798",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S30sh4_sel.png",
  "source_sha256": "272f1ddb12c021c4291fb23419515f0283ba2ee1b57074b0e8f8d0cbc52bc258",
  "file": "S30sh4_cine.png",
  "latency_ms": 11070
 },
 "S31sh2::signage": {
  "fp": "835996f688988602",
  "inscriptions": []
 },
 "era_assess::00dfd7a62bbe62bc": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 원룸 자취방 인테리어",
    "search_terms_native": [
     "2000년대 원룸",
     "한국 자취방 노란장판",
     "2010년대 원룸 인테리어",
     "옛날 원룸 내부"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며, 영어 등 타 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "2000년대와 2010년대 한국의 전형적인 원룸(자취방)은 특유의 노란 장판, 벽지 패턴, 창문 구조, 소형 가전 배치 등이 있어 서구식 스튜디오 아파트나 현대적인 오피스텔 디자인과 확연히 구분됩니다."
   }
  ]
 },
 "era_ref::1bd9089f148125cb": {
  "subject": "2000년대~2010년대 한국의 원룸 자취방 인테리어",
  "terms": [
   "2000년대 원룸",
   "한국 자취방 노란장판",
   "2010년대 원룸 인테리어",
   "옛날 원룸 내부"
  ],
  "queries": [
   [
    "2000년대 한국 원룸 자취방 노란장판 옛날 원룸 내부",
    "2010년대 한국 원룸 인테리어 자취방 내부"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://contents-cdn.viewus.co.kr/image/230131/5bdaa3cd-91fc-4c1a-8303-53111ae13463.jpeg",
  "picked_reason_ko": "3번은 침대·수납장·낮은 탁자와 생활용품이 한 공간에 놓인 2000~2010년대 한국 원룸 자취방의 일상적인 구성과 마감이 가장 넓고 선명하게 읽힌다.",
  "sha256": "8188de6e71bbbc9203d56d0a257081f9be13d2ccaeeac5f4074e7ca470dddcda",
  "file": "eraref_1bd9089f148125cb.png"
 },
 "S31sh2::bgfirst_bg": {
  "input_fingerprint": "a23f7dcd61f3c47f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): TV 화면 불빛이 비치는 가운데 동공이 크게 확장된 채 딱딱하게 굳어버린 이미경의 얼굴 클로즈업\n\nLOCATION (lock): In the main living area of the one-room apartment, illuminated by the switched-on television.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 이미경's face height on the television-facing side, the dolly-in ends in an extreme facial close-up from a three-quarter angle, keeping both enlarged pupils and the rigid set of her mouth legible. Her face occupies most of the frame as her shoulders lock in the interrupted turn toward the off-screen television; the sole emphasized change is the severe tightening of camera distance.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원룸 실내 (밤의 실내); used as Remains only as a minimal soft boundary around the isolated face; 젖은 머리카락 (샤워 후 젖어 있음); used as Visible as damp strands along the edge of the facial close-up, carrying continuity from the shower.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Television light falls across 이미경's rigid face at night, rendered with restrained contrast and a subdued neutral-to-cool cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 원룸 자취방 인테리어: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): TV 화면 불빛이 비치는 가운데 동공이 크게 확장된 채 딱딱하게 굳어버린 이미경의 얼굴 클로즈업\n\nLOCATION (lock): In the main living area of the one-room apartment, illuminated by the switched-on television.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 이미경's face height on the television-facing side, the dolly-in ends in an extreme facial close-up from a three-quarter angle, keeping both enlarged pupils and the rigid set of her mouth legible. Her face occupies most of the frame as her shoulders lock in the interrupted turn toward the off-screen television; the sole emphasized change is the severe tightening of camera distance.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원룸 실내 (밤의 실내); used as Remains only as a minimal soft boundary around the isolated face; 젖은 머리카락 (샤워 후 젖어 있음); used as Visible as damp strands along the edge of the facial close-up, carrying continuity from the shower.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Television light falls across 이미경's rigid face at night, rendered with restrained contrast and a subdued neutral-to-cool cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 원룸 자취방 인테리어: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S31sh2__bgfirst_bg.png",
  "asset_id": "76773662-c6de-4a4e-99bb-ea7879de2f20",
  "input_asset_ids": [
   "71deedeb-7041-46d1-9934-a24874a52a0b",
   "b370b24f-a2b7-413a-9983-0f2a592c1447"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국의 원룸 자취방 인테리어",
   "queries": [
    [
     "2000년대 한국 원룸 자취방 노란장판 옛날 원룸 내부",
     "2010년대 한국 원룸 인테리어 자취방 내부"
    ]
   ],
   "picked_url": "https://contents-cdn.viewus.co.kr/image/230131/5bdaa3cd-91fc-4c1a-8303-53111ae13463.jpeg",
   "sha256": "8188de6e71bbbc9203d56d0a257081f9be13d2ccaeeac5f4074e7ca470dddcda",
   "file": "eraref_1bd9089f148125cb.png"
  }
 },
 "S31sh2": {
  "input_fingerprint": "c7927c8d59c1445a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): TV 화면 불빛이 비치는 가운데 동공이 크게 확장된 채 딱딱하게 굳어버린 이미경의 얼굴 클로즈업\n\nLOCATION (lock): In the main living area of the one-room apartment, illuminated by the switched-on television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 이미경's face height on the television-facing side, the dolly-in ends in an extreme facial close-up from a three-quarter angle, keeping both enlarged pupils and the rigid set of her mouth legible. Her face occupies most of the frame as her shoulders lock in the interrupted turn toward the off-screen television; the sole emphasized change is the severe tightening of camera distance.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원룸 실내 (밤의 실내); used as Remains only as a minimal soft boundary around the isolated face; 젖은 머리카락 (샤워 후 젖어 있음); used as Visible as damp strands along the edge of the facial close-up, carrying continuity from the shower.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Television light falls across 이미경's rigid face at night, rendered with restrained contrast and a subdued neutral-to-cool cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): TV 화면 불빛이 비치는 가운데 동공이 크게 확장된 채 딱딱하게 굳어버린 이미경의 얼굴 클로즈업\n\nLOCATION (lock): In the main living area of the one-room apartment, illuminated by the switched-on television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 이미경's face height on the television-facing side, the dolly-in ends in an extreme facial close-up from a three-quarter angle, keeping both enlarged pupils and the rigid set of her mouth legible. Her face occupies most of the frame as her shoulders lock in the interrupted turn toward the off-screen television; the sole emphasized change is the severe tightening of camera distance.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원룸 실내 (밤의 실내); used as Remains only as a minimal soft boundary around the isolated face; 젖은 머리카락 (샤워 후 젖어 있음); used as Visible as damp strands along the edge of the facial close-up, carrying continuity from the shower.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Television light falls across 이미경's rigid face at night, rendered with restrained contrast and a subdued neutral-to-cool cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): TV 화면 불빛이 비치는 가운데 동공이 크게 확장된 채 딱딱하게 굳어버린 이미경의 얼굴 클로즈업\n\nLOCATION (lock): In the main living area of the one-room apartment, illuminated by the switched-on television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 이미경's face height on the television-facing side, the dolly-in ends in an extreme facial close-up from a three-quarter angle, keeping both enlarged pupils and the rigid set of her mouth legible. Her face occupies most of the frame as her shoulders lock in the interrupted turn toward the off-screen television; the sole emphasized change is the severe tightening of camera distance.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원룸 실내 (밤의 실내); used as Remains only as a minimal soft boundary around the isolated face; 젖은 머리카락 (샤워 후 젖어 있음); used as Visible as damp strands along the edge of the facial close-up, carrying continuity from the shower.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Television light falls across 이미경's rigid face at night, rendered with restrained contrast and a subdued neutral-to-cool cast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S31sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S31sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L30B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "오프스크린 TV 방향과 카메라 위치 지시를 정확히 따랐으며 크게 확장된 동공을 잘 표현했으나, 익스트림 클로즈업 척도에는 조금 미치지 못함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "TV가 배경에 노출되어 오프스크린 및 카메라 위치 지시를 위반했으며, 동공 확장이 상대적으로 미약함."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선은 화면 우측 밖을 향함.",
      "built_space": "소파, 창문, 냉장고가 보이는 원룸 실내. TV 쪽에서 인물을 바라보는 올바른 앵글.",
      "entities": "이미경과 일치하는 외모와 복장(분홍색 가디건, 흰 블라우스). 젖은 머리카락과 크게 확장된 동공이 명확함.",
      "hard_violations": [],
      "physics": "보이지 않는 바닥/의자에 안정적으로 지탱됨."
     },
     {
      "label": "A",
      "direction": "시선은 화면 우측 밖을 향함.",
      "built_space": "소파, 창문, TV가 보이는 원룸 실내. 카메라가 TV 방향을 바라봄.",
      "entities": "이미경과 일치하는 외모와 복장. 젖은 머리카락이 확인되나 동공 확장이 두드러지지 않음.",
      "hard_violations": [],
      "physics": "보이지 않는 바닥/의자에 안정적으로 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "오프스크린 TV 방향과 카메라 위치 지시를 정확히 따랐으며 크게 확장된 동공을 잘 표현했으나, 익스트림 클로즈업 척도에는 조금 미치지 못함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "TV가 배경에 노출되어 오프스크린 및 카메라 위치 지시를 위반했으며, 동공 확장이 상대적으로 미약함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 화면 우측 밖을 향함.",
      "built_space": "소파, 창문, 냉장고가 보이는 원룸 실내. TV 쪽에서 인물을 바라보는 올바른 앵글.",
      "entities": "이미경과 일치하는 외모와 복장(분홍색 가디건, 흰 블라우스). 젖은 머리카락과 크게 확장된 동공이 명확함.",
      "hard_violations": [],
      "physics": "보이지 않는 바닥/의자에 안정적으로 지탱됨."
     },
     {
      "label": "A",
      "direction": "시선은 화면 우측 밖을 향함.",
      "built_space": "소파, 창문, TV가 보이는 원룸 실내. 카메라가 TV 방향을 바라봄.",
      "entities": "이미경과 일치하는 외모와 복장. 젖은 머리카락이 확인되나 동공 확장이 두드러지지 않음.",
      "hard_violations": [],
      "physics": "보이지 않는 바닥/의자에 안정적으로 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 '극단적인 얼굴 클로즈업' 프레이밍을 정확히 준수하여 화면 대부분을 얼굴로 채웠으며, 동공 확장과 화면 밖 TV 조명 묘사가 우수합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "화면 밖이어야 할 TV가 배경에 포함될 정도로 프레임이 넓어져 '극단적인 얼굴 클로즈업' 지시를 위반했으며, 배경의 TV 위치와 얼굴에 맺힌 조명 방향이 일치하지 않습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 화면 밖 우측(TV가 있는 방향)을 향함.",
      "built_space": "배경은 흐릿하게 처리되어 창문과 냉장고 등 원룸의 형태만 부드럽게 암시함.",
      "entities": "이미경(타원형 얼굴, 핑크색 카디건, 젖은 머리카락, 확장된 동공)이 정확히 묘사됨.",
      "hard_violations": [],
      "physics": "자연스러운 자세로 지탱됨."
     },
     {
      "label": "B",
      "direction": "시선은 화면 밖 우측을 향함.",
      "built_space": "소파, 창문, 냉장고, 켜진 TV가 명확히 보이며, TV가 인물 뒤 우측에 배치됨.",
      "entities": "이미경(타원형 얼굴, 핑크색 카디건, 젖은 머리카락, 확장된 동공)이 묘사됨.",
      "hard_violations": [],
      "physics": "자연스러운 자세로 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 '극단적인 얼굴 클로즈업' 프레이밍을 정확히 준수하여 화면 대부분을 얼굴로 채웠으며, 동공 확장과 화면 밖 TV 조명 묘사가 우수합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "화면 밖이어야 할 TV가 배경에 포함될 정도로 프레임이 넓어져 '극단적인 얼굴 클로즈업' 지시를 위반했으며, 배경의 TV 위치와 얼굴에 맺힌 조명 방향이 일치하지 않습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 화면 밖 우측(TV가 있는 방향)을 향함.",
      "built_space": "배경은 흐릿하게 처리되어 창문과 냉장고 등 원룸의 형태만 부드럽게 암시함.",
      "entities": "이미경(타원형 얼굴, 핑크색 카디건, 젖은 머리카락, 확장된 동공)이 정확히 묘사됨.",
      "hard_violations": [],
      "physics": "자연스러운 자세로 지탱됨."
     },
     {
      "label": "A",
      "direction": "시선은 화면 밖 우측을 향함.",
      "built_space": "소파, 창문, 냉장고, 켜진 TV가 명확히 보이며, TV가 인물 뒤 우측에 배치됨.",
      "entities": "이미경(타원형 얼굴, 핑크색 카디건, 젖은 머리카락, 확장된 동공)이 묘사됨.",
      "hard_violations": [],
      "physics": "자연스러운 자세로 지탱됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "B",
    "direction": "시선은 화면 우측 밖을 향함.",
    "built_space": "소파, 창문, 냉장고가 보이는 원룸 실내. TV 쪽에서 인물을 바라보는 올바른 앵글.",
    "entities": "이미경과 일치하는 외모와 복장(분홍색 가디건, 흰 블라우스). 젖은 머리카락과 크게 확장된 동공이 명확함.",
    "hard_violations": [],
    "physics": "보이지 않는 바닥/의자에 안정적으로 지탱됨."
   },
   {
    "label": "A",
    "direction": "시선은 화면 우측 밖을 향함.",
    "built_space": "소파, 창문, TV가 보이는 원룸 실내. 카메라가 TV 방향을 바라봄.",
    "entities": "이미경과 일치하는 외모와 복장. 젖은 머리카락이 확인되나 동공 확장이 두드러지지 않음.",
    "hard_violations": [],
    "physics": "보이지 않는 바닥/의자에 안정적으로 지탱됨."
   }
  ],
  "totals": {
   "A": 8,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "오프스크린 TV 방향과 카메라 위치 지시를 정확히 따랐으며 크게 확장된 동공을 잘 표현했으나, 익스트림 클로즈업 척도에는 조금 미치지 못함."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "TV가 배경에 노출되어 오프스크린 및 카메라 위치 지시를 위반했으며, 동공 확장이 상대적으로 미약함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L30B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "양쪽 눈 동공이 크게 확장되지 않고 홍채가 많이 보인다.",
     "fix_en": "Widen the black central pupils so they are extremely dilated and cover most of the irises, replacing the small pupils, while preserving the woman, her position, her clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "입이 딱딱하게 굳지 않고 살짝 벌어진 부드러운 표정이다.",
     "fix_en": "Close the mouth into a rigid, tightly sealed line, replacing the slightly parted lips, while preserving the woman, her position, her clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "양쪽 안구가 해부학적으로 비정상적으로 크고 둥글다.",
     "fix_en": "Scale down the overly large eyeballs to normal human anatomical proportions, filling the excess space with normal eyelids and skin, while preserving the woman, her position, her clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "샷 텍스트와 카메라 지시문에서 '크게 확장된 동공'을 명시했으나, 인물의 눈동자 중심(동공)이 일반적이거나 오히려 작게 수축된 상태로 묘사됨.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 이목구비가 참조 인물과 일치하지 않는다.",
     "severity": "major"
    },
    {
     "issue_ko": "양쪽 눈 동공이 크게 확장되지 않고 홍채가 많이 보인다.",
     "severity": "major"
    },
    {
     "issue_ko": "입이 딱딱하게 굳지 않고 살짝 벌어진 부드러운 표정이다.",
     "severity": "major"
    },
    {
     "issue_ko": "양쪽 안구가 해부학적으로 비정상적으로 크고 둥글다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "양쪽 눈 동공이 크게 확장되지 않고 홍채가 많이 보인다.",
    "fix_en": "Widen the black central pupils so they are extremely dilated and cover most of the irises, replacing the small pupils, while preserving the woman, her position, her clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "입이 딱딱하게 굳지 않고 살짝 벌어진 부드러운 표정이다.",
    "fix_en": "Close the mouth into a rigid, tightly sealed line, replacing the slightly parted lips, while preserving the woman, her position, her clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "양쪽 안구가 해부학적으로 비정상적으로 크고 둥글다.",
    "fix_en": "Scale down the overly large eyeballs to normal human anatomical proportions, filling the excess space with normal eyelids and skin, while preserving the woman, her position, her clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S31sh2__bgfirst_bg.png",
   "bg_asset_id": "76773662-c6de-4a4e-99bb-ea7879de2f20",
   "bg_record_key": "S31sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S31sh2::cine": {
  "applied": true,
  "fingerprint": "6a10621f1a88a115a0b8a140e311776982effe57c581c29a9c26f3a1ba20b213",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S31sh2_sel.png",
  "source_sha256": "015ed9e15a3044d535fc18ad89f104f4103793ab1062ee71766d01dae3911ce1",
  "file": "S31sh2_cine.png",
  "latency_ms": 10940
 },
 "S32sh1::signage": {
  "fp": "25118bc729daf60e",
  "inscriptions": [
   {
    "surface_native": "TV 화면 하단 자막 바",
    "text_native": "심층보도: 강가의 의문사",
    "reason_ko": "2000년대 한국 시사 프로그램 특유의 현장감과 고발 분위기를 연출하기 위해 화면 하단에 표시되는 보도 주제 자막이 필요합니다."
   }
  ]
 },
 "S32sh1": {
  "input_fingerprint": "65495bf7e0dbdf9a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 실내 공간의 켜진 TV 화면 속, 모자이크 처리된 유가족이 강가에 선 모습이 띄워진 시사 프로그램 화면 클로즈업\n\nLOCATION (lock): Inside the deputy chief prosecutor's dark office, focused on the switched-on television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera begins static directly before the television at screen height and extreme close distance, presenting the broadcast image nearly edge to edge with only a narrow trace of the display boundary. The mosaicked 유가족 stand by the river in a naturally uneven grouping—different weight shifts, head angles, and spacing—while remaining collectively subdued; no room context is introduced at this starting stage.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: TV 화면 (시사 프로그램이 재생 중) — The active front display face is shown directly, carrying the current-affairs program image of the mosaicked 유가족 by the river; used as Functions as the mediated field through which the entire shot is perceived; 강가 사건 현장 (방송 화면에 표시됨); used as Appears inside the televised image as the location surrounding the family; 모자이크 처리 (유가족 얼굴에 적용됨) — The broadcast-facing image layer covers the visible faces of the 유가족; used as Obscures identifying facial information within the broadcast.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The television display supplies the visible broadcast luminance, presented with subdued neutral-to-cool tonality and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- TV 화면 하단 자막 바: \"심층보도: 강가의 의문사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 실내 공간의 켜진 TV 화면 속, 모자이크 처리된 유가족이 강가에 선 모습이 띄워진 시사 프로그램 화면 클로즈업\n\nLOCATION (lock): Inside the deputy chief prosecutor's dark office, focused on the switched-on television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera begins static directly before the television at screen height and extreme close distance, presenting the broadcast image nearly edge to edge with only a narrow trace of the display boundary. The mosaicked 유가족 stand by the river in a naturally uneven grouping—different weight shifts, head angles, and spacing—while remaining collectively subdued; no room context is introduced at this starting stage.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: TV 화면 (시사 프로그램이 재생 중) — The active front display face is shown directly, carrying the current-affairs program image of the mosaicked 유가족 by the river; used as Functions as the mediated field through which the entire shot is perceived; 강가 사건 현장 (방송 화면에 표시됨); used as Appears inside the televised image as the location surrounding the family; 모자이크 처리 (유가족 얼굴에 적용됨) — The broadcast-facing image layer covers the visible faces of the 유가족; used as Obscures identifying facial information within the broadcast.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The television display supplies the visible broadcast luminance, presented with subdued neutral-to-cool tonality and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- TV 화면 하단 자막 바: \"심층보도: 강가의 의문사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 실내 공간의 켜진 TV 화면 속, 모자이크 처리된 유가족이 강가에 선 모습이 띄워진 시사 프로그램 화면 클로즈업\n\nLOCATION (lock): Inside the deputy chief prosecutor's dark office, focused on the switched-on television. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera begins static directly before the television at screen height and extreme close distance, presenting the broadcast image nearly edge to edge with only a narrow trace of the display boundary. The mosaicked 유가족 stand by the river in a naturally uneven grouping—different weight shifts, head angles, and spacing—while remaining collectively subdued; no room context is introduced at this starting stage.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: TV 화면 (시사 프로그램이 재생 중) — The active front display face is shown directly, carrying the current-affairs program image of the mosaicked 유가족 by the river; used as Functions as the mediated field through which the entire shot is perceived; 강가 사건 현장 (방송 화면에 표시됨); used as Appears inside the televised image as the location surrounding the family; 모자이크 처리 (유가족 얼굴에 적용됨) — The broadcast-facing image layer covers the visible faces of the 유가족; used as Obscures identifying facial information within the broadcast.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The television display supplies the visible broadcast luminance, presented with subdued neutral-to-cool tonality and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- TV 화면 하단 자막 바: \"심층보도: 강가의 의문사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 TV 화면을 정면에서 직각으로 바라보고 있습니다.",
    "built_space": "TV의 얇은 베젤만 화면 가장자리에 보이며, 방의 다른 구조물이나 벽면은 보이지 않습니다.",
    "entities": "화면 속 강가에 선 인물들의 얼굴이 모자이크 처리되어 있으며, 하단 자막 '심층보도: 강가의 의문사'가 정확히 표기되었습니다.",
    "hard_violations": [],
    "physics": "화면 속 인물들은 땅에 안정적으로 서 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라가 TV 화면을 약간 왼쪽 아래 측면에서 비스듬히 바라보고 있습니다.",
    "built_space": "TV의 두꺼운 측면 베젤과 화면 너머의 벽면이 함께 화면에 잡히고 있습니다.",
    "entities": "모자이크 처리된 인물들과 강가 배경이 보이며, 하단 자막 텍스트는 정확하게 표시되어 있습니다.",
    "hard_violations": [],
    "physics": "화면 속 인물들은 땅에 안정적으로 서 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "정면에서 TV 화면을 가득 채우는 프레이밍과 자막 텍스트, 모자이크 처리 등 프롬프트의 지시사항을 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "카메라가 정면이 아닌 측면에서 촬영되어 벽면이 노출되었으며, 이는 주변 공간을 배제하라는 프레이밍 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L31B02.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우측에서 두 번째 아이의 복부 앞에 얼굴과 무관한 사각형 모자이크 픽셀 덩어리가 허공에 떠 있습니다.",
     "fix_en": "Remove the floating mosaic block from the chest of the second person from the right, restoring the continuous dark grey fabric of their shirt. Preserve all people, their current positions, their facial mosaics, the river background, the TV frame, and the subtitle text.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 속 유가족들이 정면을 향해 두 줄로 가지런히 서서 양발에 체중을 둔 단체 사진처럼 보인다",
     "fix_en": "Adjust the postures, head tilts, and shoulder angles of the family members to show uneven weight shifts and a less posed, asymmetrical arrangement. Preserve the identities, the TV frame, the lighting, the facial mosaics, the background river, and the subtitle text.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우측에서 두 번째 아이의 복부 앞에 얼굴과 무관한 사각형 모자이크 픽셀 덩어리가 허공에 떠 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 속 유가족들이 정면을 향해 두 줄로 가지런히 서서 양발에 체중을 둔 단체 사진처럼 보인다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 속 유가족들이 정면을 향해 두 줄로 가지런히 서서 양발에 체중을 둔 단체 사진처럼 보인다",
    "fix_en": "Adjust the postures, head tilts, and shoulder angles of the family members to show uneven weight shifts and a less posed, asymmetrical arrangement. Preserve the identities, the TV frame, the lighting, the facial mosaics, the background river, and the subtitle text.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the floating mosaic block from the chest of the second person from the right, restoring the continuous dark grey fabric of their shirt. Preserve all people, their current positions, their facial mosaics, the river background, the TV frame, and the subtitle text.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 화면 구도와 자막을 정확히 구현했으며, A에 있던 불필요한 모자이크 오류를 깔끔하게 수정함."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지시사항을 잘 따랐으나 우측 소년의 복부 부근에 불필요한 모자이크 덩어리가 남아 있음."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 TV 화면을 정면으로 주시하며, 화면 속 유가족은 앞을 향해 서 있음.",
      "built_space": "TV 베젤과 실내 벽면의 일부가 가장자리에 좁게 보임.",
      "entities": "TV 화면, 강가의 유가족, 지정된 텍스트 자막이 확인됨. 다만 우측 소년의 옷 부근에 원인 모를 모자이크가 추가로 존재함.",
      "hard_violations": [],
      "physics": "화면 속 인물들은 지면 위에 안정적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "카메라는 TV 화면을 정면으로 주시하며, 화면 속 유가족은 앞을 향해 서 있음.",
      "built_space": "TV 베젤과 실내 벽면의 일부가 가장자리에 좁게 보임.",
      "entities": "TV 화면, 얼굴이 모자이크 처리된 강가의 유가족, 지정된 텍스트 자막이 모두 정확하게 반영됨.",
      "hard_violations": [],
      "physics": "화면 속 인물들은 지면 위에 안정적으로 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 화면 구도와 자막을 정확히 구현했으며, A에 있던 불필요한 모자이크 오류를 깔끔하게 수정함."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지시사항을 잘 따랐으나 우측 소년의 복부 부근에 불필요한 모자이크 덩어리가 남아 있음."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 TV 화면을 정면으로 주시하며, 화면 속 유가족은 앞을 향해 서 있음.",
      "built_space": "TV 베젤과 실내 벽면의 일부가 가장자리에 좁게 보임.",
      "entities": "TV 화면, 강가의 유가족, 지정된 텍스트 자막이 확인됨. 다만 우측 소년의 옷 부근에 원인 모를 모자이크가 추가로 존재함.",
      "hard_violations": [],
      "physics": "화면 속 인물들은 지면 위에 안정적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "카메라는 TV 화면을 정면으로 주시하며, 화면 속 유가족은 앞을 향해 서 있음.",
      "built_space": "TV 베젤과 실내 벽면의 일부가 가장자리에 좁게 보임.",
      "entities": "TV 화면, 얼굴이 모자이크 처리된 강가의 유가족, 지정된 텍스트 자막이 모두 정확하게 반영됨.",
      "hard_violations": [],
      "physics": "화면 속 인물들은 지면 위에 안정적으로 서 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 TV 화면 클로즈업, 모자이크 처리된 유가족, 지정된 자막 텍스트를 정확히 구현하여 프롬프트의 지시사항을 훌륭하게 따름."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "우측 소년의 옷 위에 정체불명의 모자이크된 얼굴이 허공에 떠 있는 치명적인 생성 오류가 발생함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "TV 화면 속 인물들이 대체로 중앙이나 아래를 향해 시선을 두고 있음.",
      "built_space": "TV 베젤이 화면 외곽을 감싸고 있으며 지시된 대로 실내 배경은 배제된 클로즈업 구도임.",
      "entities": "모자이크 처리된 유가족 무리, 강가 배경, 하단 자막 '심층보도: 강가의 의문사' 모두 프롬프트와 일치함.",
      "hard_violations": [],
      "physics": "인물들이 화면 속 땅(풀밭) 위에 자연스럽게 서 있음."
     },
     {
      "label": "B",
      "direction": "TV 화면 속 인물들이 대체로 중앙이나 아래를 향해 시선을 두고 있음.",
      "built_space": "TV 베젤이 화면 외곽을 감싸고 있으며 실내 배경은 배제됨.",
      "entities": "모자이크 처리된 유가족 무리와 강가 배경, 하단 자막이 존재함.",
      "hard_violations": [
       "화면 우측 후드티를 입은 소년의 옷 위에 프롬프트에 없는 정체불명의 작은 모자이크 얼굴이 떠 있음 (불가능한 신체/객체 추가)."
      ],
      "physics": "인물들은 땅에 서 있으나, 우측 소년의 몸통에 부자연스러운 모자이크 조각이 지지대 없이 생성됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 TV 화면 클로즈업, 모자이크 처리된 유가족, 지정된 자막 텍스트를 정확히 구현하여 프롬프트의 지시사항을 훌륭하게 따름."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "우측 소년의 옷 위에 정체불명의 모자이크된 얼굴이 허공에 떠 있는 치명적인 생성 오류가 발생함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "TV 화면 속 인물들이 대체로 중앙이나 아래를 향해 시선을 두고 있음.",
      "built_space": "TV 베젤이 화면 외곽을 감싸고 있으며 지시된 대로 실내 배경은 배제된 클로즈업 구도임.",
      "entities": "모자이크 처리된 유가족 무리, 강가 배경, 하단 자막 '심층보도: 강가의 의문사' 모두 프롬프트와 일치함.",
      "hard_violations": [],
      "physics": "인물들이 화면 속 땅(풀밭) 위에 자연스럽게 서 있음."
     },
     {
      "label": "A",
      "direction": "TV 화면 속 인물들이 대체로 중앙이나 아래를 향해 시선을 두고 있음.",
      "built_space": "TV 베젤이 화면 외곽을 감싸고 있으며 실내 배경은 배제됨.",
      "entities": "모자이크 처리된 유가족 무리와 강가 배경, 하단 자막이 존재함.",
      "hard_violations": [
       "화면 우측 후드티를 입은 소년의 옷 위에 프롬프트에 없는 정체불명의 작은 모자이크 얼굴이 떠 있음 (불가능한 신체/객체 추가)."
      ],
      "physics": "인물들은 땅에 서 있으나, 우측 소년의 몸통에 부자연스러운 모자이크 조각이 지지대 없이 생성됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "플레이트만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S32sh1::cine": {
  "applied": true,
  "fingerprint": "4ee8c90934688cd2d7f4890e45758a279e0fc67adf1d7cc68fc4d3d419626f29",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S32sh1_sel.png",
  "source_sha256": "62d8026ec4ca1ff80eace1a506064db78a1a218cc9283d772eb0352865d6b561",
  "file": "S32sh1_cine.png",
  "latency_ms": 10953
 },
 "S32sh2::signage": {
  "fp": "e4333e6f34afeb29",
  "inscriptions": [
   {
    "surface_native": "서예 액자",
    "text_native": "正道",
    "reason_ko": "차장검사 집무실 응접 공간 뒷벽에 걸려 있는 서예 액자로, 한국 검찰 간부 집무실 특유의 권위 있고 엄숙한 분위기를 연출하기 위해 필요한 시각적 요소입니다."
   }
  ]
 },
 "S32sh2": {
  "input_fingerprint": "4ec63f43acbcf69a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 장원섭을 향해 미간을 잔뜩 찌푸린 채 쏘아보는 차장검사의 상체와 굳은 표정의 장원섭의 어깨\n\nLOCATION (lock): In the television-facing seating area of the deputy chief prosecutor's office, with the screen providing visible light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just behind 장원섭's near shoulder and slightly below the 차장검사's eye line, the dolly-in settles into an over-the-shoulder medium close frame that compresses the reprimand onto one axis. 장원섭's rigid shoulder occupies the near right edge while the 차장검사 leans into the left-center, brow clenched and eyes fixed on 장원섭; only the camera distance tightens as the confrontation peaks.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 차장검사실 내부 (밤의 실내); used as Provides restrained office context behind the confrontation without interrupting the eyeline.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime office ambience with restrained spill from the active television, kept neutral-to-cool and procedurally sober.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서예 액자: \"正道\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 장원섭을 향해 미간을 잔뜩 찌푸린 채 쏘아보는 차장검사의 상체와 굳은 표정의 장원섭의 어깨\n\nLOCATION (lock): In the television-facing seating area of the deputy chief prosecutor's office, with the screen providing visible light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just behind 장원섭's near shoulder and slightly below the 차장검사's eye line, the dolly-in settles into an over-the-shoulder medium close frame that compresses the reprimand onto one axis. 장원섭's rigid shoulder occupies the near right edge while the 차장검사 leans into the left-center, brow clenched and eyes fixed on 장원섭; only the camera distance tightens as the confrontation peaks.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 차장검사실 내부 (밤의 실내); used as Provides restrained office context behind the confrontation without interrupting the eyeline.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime office ambience with restrained spill from the active television, kept neutral-to-cool and procedurally sober.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서예 액자: \"正道\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 장원섭을 향해 미간을 잔뜩 찌푸린 채 쏘아보는 차장검사의 상체와 굳은 표정의 장원섭의 어깨\n\nLOCATION (lock): In the television-facing seating area of the deputy chief prosecutor's office, with the screen providing visible light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just behind 장원섭's near shoulder and slightly below the 차장검사's eye line, the dolly-in settles into an over-the-shoulder medium close frame that compresses the reprimand onto one axis. 장원섭's rigid shoulder occupies the near right edge while the 차장검사 leans into the left-center, brow clenched and eyes fixed on 장원섭; only the camera distance tightens as the confrontation peaks.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 차장검사실 내부 (밤의 실내); used as Provides restrained office context behind the confrontation without interrupting the eyeline.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime office ambience with restrained spill from the active television, kept neutral-to-cool and procedurally sober.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서예 액자: \"正道\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "차장검사의 시선은 화면 우측의 장원섭을 향하고 있으며, 장원섭 역시 차장검사를 마주보고 있습니다.",
    "built_space": "야간의 사무실 내부로, 벽면에는 '正道' 서예 액자가 걸려 있고 배경에는 레퍼런스와 동일한 화면을 재생 중인 TV가 배치되어 있습니다.",
    "entities": "차장검사(좌측)는 레퍼런스의 더블 브레스티드 정장과 넥타이를 착용했고 찌푸린 표정을 짓고 있습니다. 장원섭(우측)은 회색 정장을 입고 있으며 측면 얼굴이 상당 부분 노출되었습니다. TV 화면 속 픽셀화된 인물들과 자막 텍스트, 서예 액자 모두 지시와 일치합니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥을 지지하고 서 있는 자세이며, TV 등 사물들도 가구 위에 안정적으로 놓여 있어 물리적 오류가 없습니다."
   },
   {
    "label": "B",
    "direction": "차장검사의 시선과 몸의 방향이 화면 우측 전경에 있는 장원섭을 정확히 향하고 있습니다.",
    "built_space": "야간의 사무실 환경이며, 뒤편 장식장 위 TV와 좌측 벽면의 서예 액자가 제 위치에 있습니다. 차장검사는 가죽 의자에 앉아 있는 구도입니다.",
    "entities": "차장검사(좌측)는 지정된 정장과 넥타이, 찌푸린 표정을 정확히 보여줍니다. 장원섭(우측 전경)은 등과 어깨, 뒤통수만 노출되어 오버더숄더 구도에 부합합니다. TV 화면의 디테일(모자이크 처리된 인물들, 자막)과 '正道' 서예 액자의 텍스트가 일치합니다.",
    "hard_violations": [],
    "physics": "차장검사는 의자에 몸을 지지하고 있으며, 화면 내 떠 있거나 지지되지 않은 사물은 없습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 오버더숄더(OTS) 앵글을 정확히 구현하여 장원섭의 어깨를 우측에 걸고 차장검사에게 시선을 집중시켰으며, 레퍼런스의 디테일(TV 화면, 서예 액자, 인물 외양)을 충실히 반영했습니다."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "인물들의 외양과 배경 요소는 레퍼런스와 잘 일치하나, 카메라가 측면으로 너무 이동하여 오버더숄더 앵글보다는 두 인물의 측면 프로필 샷에 가깝게 연출되어 프레이밍 지시에서 감점되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S32sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   },
   {
    "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869633>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "TV 화면 안의 뉴스 이미지와 자막('심층보도: 강가의 의문사')이 모니터 베젤을 벗어나 뒷배경 및 차장검사의 옷 위에 평면적인 오버레이로 떠 있음.",
     "fix_en": "Remove the flat graphic overlay of the TV broadcast image and subtitle from the deputy chief prosecutor's suit and the background wall. Confine the broadcast image and subtitle strictly within the physical bezel of the television screen in the background. Maintain the current characters, their poses and clothing, the lighting, framing, and office environment.",
     "severity": "critical",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "TV 화면 안의 뉴스 이미지와 자막('심층보도: 강가의 의문사')이 모니터 베젤을 벗어나 뒷배경 및 차장검사의 옷 위에 평면적인 오버레이로 떠 있음.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the flat graphic overlay of the TV broadcast image and subtitle from the deputy chief prosecutor's suit and the background wall. Confine the broadcast image and subtitle strictly within the physical bezel of the television screen in the background. Maintain the current characters, their poses and clothing, the lighting, framing, and office environment.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 오버더숄더 구도, 차장검사의 표정과 시선 처리, 그리고 이전 샷의 TV 화면과 '正道' 액자까지 프롬프트의 요구사항을 매우 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "장원섭의 어깨가 프레임에 포함되어야 하는 핵심 구도를 완전히 무시했으며, 차장검사의 표정과 자세도 프롬프트와 맞지 않습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "차장검사의 시선은 화면 우측에 걸쳐 있는 장원섭을 향해 고정되어 있으며 미간을 찌푸리고 있습니다.",
      "built_space": "이전 샷과 동일한 화면을 띄운 TV가 배경 중앙에 위치하며, 좌측 벽에는 요구된 한자 액자가 걸려 있습니다. 공간의 깊이와 가구 배치가 자연스럽습니다.",
      "entities": "차장검사는 레퍼런스와 동일한 스트라이프 정장과 넥타이를 착용하고 있으며 인상착의가 일치합니다. 우측에는 장원섭으로 보이는 인물의 회색 정장 어깨가 자리 잡고 있습니다. '正道' 한자도 정확히 표시되었습니다.",
      "hard_violations": [],
      "physics": "차장검사는 앉은 상태로 몸을 약간 기울인 자세를 안정적으로 유지하고 있으며, 장원섭의 어깨도 프레임에 자연스럽게 걸쳐져 있습니다."
     },
     {
      "label": "B",
      "direction": "차장검사가 정면을 응시하고 있으며, 표정은 찌푸리지 않고 평온합니다. 시선이 닿는 구체적인 대상이 프레임 내에 없습니다.",
      "built_space": "배경에 TV가 걸려 있으나 이전 샷의 구도와는 다른 넓은 공간으로 연출되었습니다.",
      "entities": "차장검사의 복장과 얼굴은 레퍼런스와 일치하나, 샷에 반드시 포함되어야 할 장원섭의 어깨가 전혀 보이지 않습니다.",
      "hard_violations": [
       "장원섭 결누락 (오버더숄더 구도 무시)",
       "프롬프트에 명시된 샷 크기(미디엄 샷) 및 앵글 위반"
      ],
      "physics": "차장검사는 바닥에 두 발을 딛고 서 있으며, 물리적으로 불가능한 부분은 없습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 오버더숄더 구도, 차장검사의 표정과 시선 처리, 그리고 이전 샷의 TV 화면과 '正道' 액자까지 프롬프트의 요구사항을 매우 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "장원섭의 어깨가 프레임에 포함되어야 하는 핵심 구도를 완전히 무시했으며, 차장검사의 표정과 자세도 프롬프트와 맞지 않습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "차장검사의 시선은 화면 우측에 걸쳐 있는 장원섭을 향해 고정되어 있으며 미간을 찌푸리고 있습니다.",
      "built_space": "이전 샷과 동일한 화면을 띄운 TV가 배경 중앙에 위치하며, 좌측 벽에는 요구된 한자 액자가 걸려 있습니다. 공간의 깊이와 가구 배치가 자연스럽습니다.",
      "entities": "차장검사는 레퍼런스와 동일한 스트라이프 정장과 넥타이를 착용하고 있으며 인상착의가 일치합니다. 우측에는 장원섭으로 보이는 인물의 회색 정장 어깨가 자리 잡고 있습니다. '正道' 한자도 정확히 표시되었습니다.",
      "hard_violations": [],
      "physics": "차장검사는 앉은 상태로 몸을 약간 기울인 자세를 안정적으로 유지하고 있으며, 장원섭의 어깨도 프레임에 자연스럽게 걸쳐져 있습니다."
     },
     {
      "label": "B",
      "direction": "차장검사가 정면을 응시하고 있으며, 표정은 찌푸리지 않고 평온합니다. 시선이 닿는 구체적인 대상이 프레임 내에 없습니다.",
      "built_space": "배경에 TV가 걸려 있으나 이전 샷의 구도와는 다른 넓은 공간으로 연출되었습니다.",
      "entities": "차장검사의 복장과 얼굴은 레퍼런스와 일치하나, 샷에 반드시 포함되어야 할 장원섭의 어깨가 전혀 보이지 않습니다.",
      "hard_violations": [
       "장원섭 결누락 (오버더숄더 구도 무시)",
       "프롬프트에 명시된 샷 크기(미디엄 샷) 및 앵글 위반"
      ],
      "physics": "차장검사는 바닥에 두 발을 딛고 서 있으며, 물리적으로 불가능한 부분은 없습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지정된 오버더숄더 앵글과 두 인물의 배치, 차장검사의 표정, 배경의 TV 화면 및 붓글씨 액자까지 프롬프트의 요구사항을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트가 요구한 장원섭의 어깨 및 오버더숄더 구도가 누락되었고, 인물이 정면을 응시하여 연출 의도를 전혀 살리지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "차장검사의 시선이 화면 우측 전경에 있는 장원섭을 향하고 있음.",
      "built_space": "사무실 배경으로 뒤쪽에 TV가 놓여 있고 좌측 벽에 서예 액자가 걸려 있음.",
      "entities": "차장검사(레퍼런스와 일치하는 외모 및 복장), 장원섭(어깨와 뒷모습), 지정된 레퍼런스 화면이 띄워진 TV, '正道'가 적힌 서예 액자가 모두 존재함.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥 또는 의자에 지탱되어 자연스럽게 서 있거나 앉아 있음."
     },
     {
      "label": "A",
      "direction": "인물이 카메라 정면을 응시하고 있음.",
      "built_space": "사무실 배경으로 뒤쪽 벽에 TV가 걸려 있고 앞에 책상과 의자가 있음.",
      "entities": "차장검사는 존재하나 장원섭이 화면에 없으며, 서예 액자도 보이지 않음. TV 화면은 레퍼런스와 일치함.",
      "hard_violations": [
       "지정된 인물(장원섭) 누락",
       "명시된 오버더숄더 구도 및 카메라 앵글 위반"
      ],
      "physics": "인물이 두 발로 바닥을 딛고 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지정된 오버더숄더 앵글과 두 인물의 배치, 차장검사의 표정, 배경의 TV 화면 및 붓글씨 액자까지 프롬프트의 요구사항을 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 요구한 장원섭의 어깨 및 오버더숄더 구도가 누락되었고, 인물이 정면을 응시하여 연출 의도를 전혀 살리지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "차장검사의 시선이 화면 우측 전경에 있는 장원섭을 향하고 있음.",
      "built_space": "사무실 배경으로 뒤쪽에 TV가 놓여 있고 좌측 벽에 서예 액자가 걸려 있음.",
      "entities": "차장검사(레퍼런스와 일치하는 외모 및 복장), 장원섭(어깨와 뒷모습), 지정된 레퍼런스 화면이 띄워진 TV, '正道'가 적힌 서예 액자가 모두 존재함.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥 또는 의자에 지탱되어 자연스럽게 서 있거나 앉아 있음."
     },
     {
      "label": "B",
      "direction": "인물이 카메라 정면을 응시하고 있음.",
      "built_space": "사무실 배경으로 뒤쪽 벽에 TV가 걸려 있고 앞에 책상과 의자가 있음.",
      "entities": "차장검사는 존재하나 장원섭이 화면에 없으며, 서예 액자도 보이지 않음. TV 화면은 레퍼런스와 일치함.",
      "hard_violations": [
       "지정된 인물(장원섭) 누락",
       "명시된 오버더숄더 구도 및 카메라 앵글 위반"
      ],
      "physics": "인물이 두 발로 바닥을 딛고 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh1"
  }
 },
 "S32sh2::cine": {
  "applied": true,
  "fingerprint": "2c9a5dee3dc768327c23f27a7e6a57075bd4ee3f22094d4943e842f7b2635556",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S32sh2_sel.png",
  "source_sha256": "7b56acffcfcd0e4ccb446b588372c4f02d878940fff84119ab93d22a14a999e7",
  "file": "S32sh2_cine.png",
  "latency_ms": 12566
 },
 "S33sh2::signage": {
  "fp": "080620c3eeaa9b7b",
  "inscriptions": [
   {
    "surface_native": "책상 위 명패",
    "text_native": "나주경찰서장",
    "reason_ko": "나주경찰서장실이라는 극 중 배경과 인물의 직책을 명확하게 드러내기 위해 서장 책상 위의 명패 표기가 필요합니다."
   }
  ]
 },
 "era_assess::cf0126f783a96655": {
  "subjects": [
   {
    "subject_native": "대한민국 경찰서장 집무실 (2000년대-2010년대)",
    "search_terms_native": [
     "경찰서장실 내부",
     "경찰서 집무실",
     "경찰 서장실"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색되어야 하며 영어나 다른 언어로 번역하거나 혼용하지 마십시오.",
    "reason_ko": "한국의 경찰서장 집무실은 태극기와 경찰기(참수리 마크), 특유의 권위 있는 가구 배치 및 가죽 소파 등 고유의 구성 요소가 있어 서구식 집무실과 뚜렷하게 구분됩니다."
   }
  ]
 },
 "era_ref::028b1970e4fe56d5": {
  "subject": "대한민국 경찰서장 집무실 (2000년대-2010년대)",
  "terms": [
   "경찰서장실 내부",
   "경찰서 집무실",
   "경찰 서장실"
  ],
  "queries": [
   [
    "대한민국 경찰서장실 내부 경찰서장 집무실 2000년대 2010년대",
    "경찰서 서장실 내부 집무실 대한민국 경찰서"
   ],
   [
    "경찰서장 집무실 사진 서장실 내부",
    "경찰서장실 내부 사진 취임 간담회"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://cdn.imweb.me/upload/S20240716bf5c580042e3c/189d52a362ded.jpg",
  "picked_reason_ko": "2번은 내부가 거의 보이지 않는 약점이 있지만, 대한민국 경찰서의 ‘경찰서장실’ 출입구와 문·표지·마감재를 명확히 보여 주는 유일한 후보이다.",
  "sha256": "e8b5e05296510853cac19f2a35462be5f8d4c586b692b446e9ab6634897ca5a6",
  "file": "eraref_028b1970e4fe56d5.png"
 },
 "S33sh2::bgfirst_bg": {
  "input_fingerprint": "5e59a44cf23bd71f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노트북 모니터 화면에 시선을 고정한 나주경찰서장과 그 맞은편에 앉아 허공을 응시하는 전택수의 모습\n\nLOCATION (lock): At the meeting table inside the police chief's office, facing an open laptop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the open corner of the desk and slightly above seated eye level, a static diagonal wide frame looks across the laptop to both men on opposite sides. The 나주경찰서장 leans toward the screen on the left while 전택수 sits deeper on the right with his weight withdrawn and gaze displaced into empty space; the defining change from the prior flow position is the characters' opposed placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 노트북 모니터 (방송 영상이 재생 중) — The active display face is angled primarily toward 나주경찰서장 and partially readable from the camera, showing the current-affairs footage; used as Creates the central visual bridge between the two seated men; 서장실 책상 (두 사람이 맞은편에 앉아 있음) — Its open corner faces the camera, with the top plane extending diagonally between both men; used as Divides the men spatially while guiding the subsequent diagonal camera move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination appropriate to the office is rendered in restrained neutral-to-cool tones with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 경찰서장 집무실 (2000년대-2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노트북 모니터 화면에 시선을 고정한 나주경찰서장과 그 맞은편에 앉아 허공을 응시하는 전택수의 모습\n\nLOCATION (lock): At the meeting table inside the police chief's office, facing an open laptop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the open corner of the desk and slightly above seated eye level, a static diagonal wide frame looks across the laptop to both men on opposite sides. The 나주경찰서장 leans toward the screen on the left while 전택수 sits deeper on the right with his weight withdrawn and gaze displaced into empty space; the defining change from the prior flow position is the characters' opposed placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 노트북 모니터 (방송 영상이 재생 중) — The active display face is angled primarily toward 나주경찰서장 and partially readable from the camera, showing the current-affairs footage; used as Creates the central visual bridge between the two seated men; 서장실 책상 (두 사람이 맞은편에 앉아 있음) — Its open corner faces the camera, with the top plane extending diagonally between both men; used as Divides the men spatially while guiding the subsequent diagonal camera move.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination appropriate to the office is rendered in restrained neutral-to-cool tones with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 경찰서장 집무실 (2000년대-2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S33sh2__bgfirst_bg.png",
  "asset_id": "b1596680-3983-4d2d-844e-25cc232f7328",
  "input_asset_ids": [
   "efcb856f-f486-4313-91b4-7febe280149b",
   "9304d4ab-470e-4715-a4e0-15f612633e70"
  ],
  "era_research": {
   "subject": "대한민국 경찰서장 집무실 (2000년대-2010년대)",
   "queries": [
    [
     "대한민국 경찰서장실 내부 경찰서장 집무실 2000년대 2010년대",
     "경찰서 서장실 내부 집무실 대한민국 경찰서"
    ],
    [
     "경찰서장 집무실 사진 서장실 내부",
     "경찰서장실 내부 사진 취임 간담회"
    ]
   ],
   "picked_url": "https://cdn.imweb.me/upload/S20240716bf5c580042e3c/189d52a362ded.jpg",
   "sha256": "e8b5e05296510853cac19f2a35462be5f8d4c586b692b446e9ab6634897ca5a6",
   "file": "eraref_028b1970e4fe56d5.png"
  }
 },
 "S33sh2": {
  "input_fingerprint": "ae5a5732fd462cea",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 노트북 모니터 화면에 시선을 고정한 나주경찰서장과 그 맞은편에 앉아 허공을 응시하는 전택수의 모습\n\nLOCATION (lock): At the meeting table inside the police chief's office, facing an open laptop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the open corner of the desk and slightly above seated eye level, a static diagonal wide frame looks across the laptop to both men on opposite sides. The 나주경찰서장 leans toward the screen on the left while 전택수 sits deeper on the right with his weight withdrawn and gaze displaced into empty space; the defining change from the prior flow position is the characters' opposed placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 노트북 모니터 (방송 영상이 재생 중) — The active display face is angled primarily toward 나주경찰서장 and partially readable from the camera, showing the current-affairs footage; used as Creates the central visual bridge between the two seated men; 서장실 책상 (두 사람이 맞은편에 앉아 있음) — Its open corner faces the camera, with the top plane extending diagonally between both men; used as Divides the men spatially while guiding the subsequent diagonal camera move.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination appropriate to the office is rendered in restrained neutral-to-cool tones with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"나주경찰서장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 노트북 모니터 화면에 시선을 고정한 나주경찰서장과 그 맞은편에 앉아 허공을 응시하는 전택수의 모습\n\nLOCATION (lock): At the meeting table inside the police chief's office, facing an open laptop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the open corner of the desk and slightly above seated eye level, a static diagonal wide frame looks across the laptop to both men on opposite sides. The 나주경찰서장 leans toward the screen on the left while 전택수 sits deeper on the right with his weight withdrawn and gaze displaced into empty space; the defining change from the prior flow position is the characters' opposed placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 노트북 모니터 (방송 영상이 재생 중) — The active display face is angled primarily toward 나주경찰서장 and partially readable from the camera, showing the current-affairs footage; used as Creates the central visual bridge between the two seated men; 서장실 책상 (두 사람이 맞은편에 앉아 있음) — Its open corner faces the camera, with the top plane extending diagonally between both men; used as Divides the men spatially while guiding the subsequent diagonal camera move.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination appropriate to the office is rendered in restrained neutral-to-cool tones with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"나주경찰서장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 노트북 모니터 화면에 시선을 고정한 나주경찰서장과 그 맞은편에 앉아 허공을 응시하는 전택수의 모습\n\nLOCATION (lock): At the meeting table inside the police chief's office, facing an open laptop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the open corner of the desk and slightly above seated eye level, a static diagonal wide frame looks across the laptop to both men on opposite sides. The 나주경찰서장 leans toward the screen on the left while 전택수 sits deeper on the right with his weight withdrawn and gaze displaced into empty space; the defining change from the prior flow position is the characters' opposed placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 노트북 모니터 (방송 영상이 재생 중) — The active display face is angled primarily toward 나주경찰서장 and partially readable from the camera, showing the current-affairs footage; used as Creates the central visual bridge between the two seated men; 서장실 책상 (두 사람이 맞은편에 앉아 있음) — Its open corner faces the camera, with the top plane extending diagonally between both men; used as Divides the men spatially while guiding the subsequent diagonal camera move.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime ambient illumination appropriate to the office is rendered in restrained neutral-to-cool tones with natural contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"나주경찰서장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S33sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S33sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832602>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L32B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832602>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "제시된 앵글과 공간 구조를 정확히 구현하며, 탁자를 사이에 둔 두 인물의 상반된 자세와 시선을 지시대로 완벽하게 포착했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "두 인물의 묘사와 자세는 훌륭하나, 카메라 앞 전경에 존재하지 않는 메인 책상을 추가로 생성해 공간 구조를 왜곡한 점이 치명적입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "서장은 몸을 기울여 탁자 위 노트북 화면을 주시하고 있으며, 전택수는 등받이에 기대어 시선을 서장 너머의 허공에 두고 있음.",
      "built_space": "소파와 의자, 회의용 탁자의 배치는 레퍼런스와 일치하나, 카메라 바로 앞 전경에 명패가 놓인 거대한 사무용 책상이 추가로 생성됨. 배경에 원래의 서장 책상이 이미 존재하므로 명백한 공간 구조 오류임.",
      "entities": "나주경찰서장(경찰 제복)과 전택수(수사 모자, 재킷)의 외형이 레퍼런스와 일치함. 명패 텍스트('나주경찰서장')도 정확함.",
      "hard_violations": [
       "invented/duplicated furniture: 배경에 메인 책상이 있음에도 전경에 명패가 놓인 메인 책상을 임의로 중복 생성함"
      ],
      "physics": "모든 인물이 의자에 안정적으로 앉아 있으며, 소품들 역시 책상과 탁자 위에 자연스럽게 놓여 있음."
     },
     {
      "label": "B",
      "direction": "서장은 화면 좌측에서 탁자 쪽으로 몸을 기울여 노트북을 응시하고, 전택수는 우측에서 몸을 뒤로 뺀 채 허공을 바라봄.",
      "built_space": "회의용 탁자 모서리에서 대각선으로 두 사람을 담아내는 구도를 정확히 구현함. 배경의 창문, 소파, 뒤편의 서장 책상과 깃발 등 레퍼런스의 공간을 그대로 재현함.",
      "entities": "나주경찰서장과 전택수의 인상착의 및 복장이 일치함. '나주경찰서장' 명패와 노트북이 회의용 탁자 위에 올바르게 위치함.",
      "hard_violations": [],
      "physics": "인물들의 체중이 소파와 의자에 올바르게 지지되어 있으며, 탁자 위 노트북의 반사광 등 물리적 표현이 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "제시된 앵글과 공간 구조를 정확히 구현하며, 탁자를 사이에 둔 두 인물의 상반된 자세와 시선을 지시대로 완벽하게 포착했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "두 인물의 묘사와 자세는 훌륭하나, 카메라 앞 전경에 존재하지 않는 메인 책상을 추가로 생성해 공간 구조를 왜곡한 점이 치명적입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서장은 몸을 기울여 탁자 위 노트북 화면을 주시하고 있으며, 전택수는 등받이에 기대어 시선을 서장 너머의 허공에 두고 있음.",
      "built_space": "소파와 의자, 회의용 탁자의 배치는 레퍼런스와 일치하나, 카메라 바로 앞 전경에 명패가 놓인 거대한 사무용 책상이 추가로 생성됨. 배경에 원래의 서장 책상이 이미 존재하므로 명백한 공간 구조 오류임.",
      "entities": "나주경찰서장(경찰 제복)과 전택수(수사 모자, 재킷)의 외형이 레퍼런스와 일치함. 명패 텍스트('나주경찰서장')도 정확함.",
      "hard_violations": [
       "invented/duplicated furniture: 배경에 메인 책상이 있음에도 전경에 명패가 놓인 메인 책상을 임의로 중복 생성함"
      ],
      "physics": "모든 인물이 의자에 안정적으로 앉아 있으며, 소품들 역시 책상과 탁자 위에 자연스럽게 놓여 있음."
     },
     {
      "label": "B",
      "direction": "서장은 화면 좌측에서 탁자 쪽으로 몸을 기울여 노트북을 응시하고, 전택수는 우측에서 몸을 뒤로 뺀 채 허공을 바라봄.",
      "built_space": "회의용 탁자 모서리에서 대각선으로 두 사람을 담아내는 구도를 정확히 구현함. 배경의 창문, 소파, 뒤편의 서장 책상과 깃발 등 레퍼런스의 공간을 그대로 재현함.",
      "entities": "나주경찰서장과 전택수의 인상착의 및 복장이 일치함. '나주경찰서장' 명패와 노트북이 회의용 탁자 위에 올바르게 위치함.",
      "hard_violations": [],
      "physics": "인물들의 체중이 소파와 의자에 올바르게 지지되어 있으며, 탁자 위 노트북의 반사광 등 물리적 표현이 자연스러움."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 로케이션 내 인물의 배치와 시선 처리, 명패 텍스트까지 프롬프트의 요구사항을 물리적 오류 없이 충실히 구현함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물의 포즈와 텍스트는 양호하나, 원본 공간 구조를 무시하고 전경에 책상을 중복으로 생성하여 실격됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "좌측 인물(나주경찰서장)은 중앙의 노트북 화면을 향해 시선을 고정하고, 우측 인물(전택수)은 정면 허공을 응시함.",
      "built_space": "제공된 로케이션 사진과 동일한 구조의 서장실. 좌측 소파와 우측 의자, 그 사이의 탁자가 정확한 비례와 구조로 배치됨.",
      "entities": "좌측 인물은 경찰 제복을 입은 나주경찰서장 레퍼런스와 일치하며, 우측 인물은 모자와 점퍼를 착용한 전택수 레퍼런스와 일치함. '나주경찰서장' 명패의 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파와 의자에 체중을 싣고 자연스럽게 착석해 있으며, 바닥과 가구에 의한 지지가 확실함."
     },
     {
      "label": "B",
      "direction": "좌측 인물은 노트북 화면을, 우측 인물은 허공을 응시하고 있음.",
      "built_space": "배경에 이미 서장실 메인 책상이 존재함에도, 카메라 바로 앞 전경에 또 다른 서장실 책상이 불필요하게 생성되어 공간 구조가 파괴됨.",
      "entities": "인물들의 외형과 복장은 레퍼런스와 일치하며 명패 텍스트도 올바르나, 우측 인물의 의상 디테일(로고 등)이 다소 뭉개짐.",
      "hard_violations": [
       "중복된 사물 및 발명된 구조 (배경에 이미 책상이 있음에도 전경에 서장실 책상을 중복으로 배치하여 불가능한 공간 구조를 만듦)"
      ],
      "physics": "두 인물 모두 지정된 가구에 의해 체중이 정상적으로 지지된 채 앉아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 로케이션 내 인물의 배치와 시선 처리, 명패 텍스트까지 프롬프트의 요구사항을 물리적 오류 없이 충실히 구현함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물의 포즈와 텍스트는 양호하나, 원본 공간 구조를 무시하고 전경에 책상을 중복으로 생성하여 실격됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "좌측 인물(나주경찰서장)은 중앙의 노트북 화면을 향해 시선을 고정하고, 우측 인물(전택수)은 정면 허공을 응시함.",
      "built_space": "제공된 로케이션 사진과 동일한 구조의 서장실. 좌측 소파와 우측 의자, 그 사이의 탁자가 정확한 비례와 구조로 배치됨.",
      "entities": "좌측 인물은 경찰 제복을 입은 나주경찰서장 레퍼런스와 일치하며, 우측 인물은 모자와 점퍼를 착용한 전택수 레퍼런스와 일치함. '나주경찰서장' 명패의 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파와 의자에 체중을 싣고 자연스럽게 착석해 있으며, 바닥과 가구에 의한 지지가 확실함."
     },
     {
      "label": "A",
      "direction": "좌측 인물은 노트북 화면을, 우측 인물은 허공을 응시하고 있음.",
      "built_space": "배경에 이미 서장실 메인 책상이 존재함에도, 카메라 바로 앞 전경에 또 다른 서장실 책상이 불필요하게 생성되어 공간 구조가 파괴됨.",
      "entities": "인물들의 외형과 복장은 레퍼런스와 일치하며 명패 텍스트도 올바르나, 우측 인물의 의상 디테일(로고 등)이 다소 뭉개짐.",
      "hard_violations": [
       "중복된 사물 및 발명된 구조 (배경에 이미 책상이 있음에도 전경에 서장실 책상을 중복으로 배치하여 불가능한 공간 구조를 만듦)"
      ],
      "physics": "두 인물 모두 지정된 가구에 의해 체중이 정상적으로 지지된 채 앉아 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "서장은 몸을 기울여 탁자 위 노트북 화면을 주시하고 있으며, 전택수는 등받이에 기대어 시선을 서장 너머의 허공에 두고 있음.",
    "built_space": "소파와 의자, 회의용 탁자의 배치는 레퍼런스와 일치하나, 카메라 바로 앞 전경에 명패가 놓인 거대한 사무용 책상이 추가로 생성됨. 배경에 원래의 서장 책상이 이미 존재하므로 명백한 공간 구조 오류임.",
    "entities": "나주경찰서장(경찰 제복)과 전택수(수사 모자, 재킷)의 외형이 레퍼런스와 일치함. 명패 텍스트('나주경찰서장')도 정확함.",
    "hard_violations": [
     "invented/duplicated furniture: 배경에 메인 책상이 있음에도 전경에 명패가 놓인 메인 책상을 임의로 중복 생성함"
    ],
    "physics": "모든 인물이 의자에 안정적으로 앉아 있으며, 소품들 역시 책상과 탁자 위에 자연스럽게 놓여 있음."
   },
   {
    "label": "B",
    "direction": "서장은 화면 좌측에서 탁자 쪽으로 몸을 기울여 노트북을 응시하고, 전택수는 우측에서 몸을 뒤로 뺀 채 허공을 바라봄.",
    "built_space": "회의용 탁자 모서리에서 대각선으로 두 사람을 담아내는 구도를 정확히 구현함. 배경의 창문, 소파, 뒤편의 서장 책상과 깃발 등 레퍼런스의 공간을 그대로 재현함.",
    "entities": "나주경찰서장과 전택수의 인상착의 및 복장이 일치함. '나주경찰서장' 명패와 노트북이 회의용 탁자 위에 올바르게 위치함.",
    "hard_violations": [],
    "physics": "인물들의 체중이 소파와 의자에 올바르게 지지되어 있으며, 탁자 위 노트북의 반사광 등 물리적 표현이 자연스러움."
   }
  ],
  "totals": {
   "A": 8,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "제시된 앵글과 공간 구조를 정확히 구현하며, 탁자를 사이에 둔 두 인물의 상반된 자세와 시선을 지시대로 완벽하게 포착했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "두 인물의 묘사와 자세는 훌륭하나, 카메라 앞 전경에 존재하지 않는 메인 책상을 추가로 생성해 공간 구조를 왜곡한 점이 치명적입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L32B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832602>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 탁자 모서리가 카메라를 향하는 대각선 구도를 지시했으나, 탁자의 측면이 카메라와 평행하게 배치된 측면 정면 구도로 렌더링되었습니다.",
     "fix_en": "Requires regeneration to correct camera angle. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "노트북 화면에 재생 중인 방송 영상 하단의 자막 텍스트가 해독할 수 없는 기호로 뭉개져 있습니다.",
     "fix_en": "Remove the garbled subtitles on the laptop screen. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "우측 전택수가 착용한 모자의 '수사' 글자 앞에 정체불명의 뭉개진 문양이 추가되어 있습니다.",
     "fix_en": "Erase the yellow mark on the right man's cap. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "좌측 나주경찰서장의 근무복 흉장이 레퍼런스(좌측 가슴)와 반대로 우측 가슴에 부착되어 있습니다.",
     "fix_en": "Move the badge to the left man's left chest. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "좌측 나주경찰서장이 무릎 위에 얹고 있는 왼손의 손가락들이 서로 융합되어 형태가 어색합니다.",
     "fix_en": "Fix the fused fingers on the left man's left hand. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "minor",
     "observation_index": 4
    },
    {
     "issue_ko": "배경 우측에 위치한 녹색 깃발과 벽면 액자 안의 텍스트가 알아볼 수 없게 뭉개져 있습니다.",
     "fix_en": "Blur the garbled text on the background flag and frame. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "minor",
     "observation_index": 5
    },
    {
     "issue_ko": "노트북 화면이 서장 쪽이 아니라 카메라를 주로 향해 열려 있다",
     "fix_en": "Angle the laptop screen toward the left man. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 탁자 모서리가 카메라를 향하는 대각선 구도를 지시했으나, 탁자의 측면이 카메라와 평행하게 배치된 측면 정면 구도로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "노트북 화면에 재생 중인 방송 영상 하단의 자막 텍스트가 해독할 수 없는 기호로 뭉개져 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 전택수가 착용한 모자의 '수사' 글자 앞에 정체불명의 뭉개진 문양이 추가되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "좌측 나주경찰서장의 근무복 흉장이 레퍼런스(좌측 가슴)와 반대로 우측 가슴에 부착되어 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "좌측 나주경찰서장이 무릎 위에 얹고 있는 왼손의 손가락들이 서로 융합되어 형태가 어색합니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "배경 우측에 위치한 녹색 깃발과 벽면 액자 안의 텍스트가 알아볼 수 없게 뭉개져 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "노트북 화면이 서장 쪽이 아니라 카메라를 주로 향해 열려 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 6,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 7,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 탁자 모서리가 카메라를 향하는 대각선 구도를 지시했으나, 탁자의 측면이 카메라와 평행하게 배치된 측면 정면 구도로 렌더링되었습니다.",
    "fix_en": "Requires regeneration to correct camera angle. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "노트북 화면에 재생 중인 방송 영상 하단의 자막 텍스트가 해독할 수 없는 기호로 뭉개져 있습니다.",
    "fix_en": "Remove the garbled subtitles on the laptop screen. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "우측 전택수가 착용한 모자의 '수사' 글자 앞에 정체불명의 뭉개진 문양이 추가되어 있습니다.",
    "fix_en": "Erase the yellow mark on the right man's cap. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "좌측 나주경찰서장의 근무복 흉장이 레퍼런스(좌측 가슴)와 반대로 우측 가슴에 부착되어 있습니다.",
    "fix_en": "Move the badge to the left man's left chest. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "minor",
    "observation_index": 3
   },
   {
    "issue_ko": "좌측 나주경찰서장이 무릎 위에 얹고 있는 왼손의 손가락들이 서로 융합되어 형태가 어색합니다.",
    "fix_en": "Fix the fused fingers on the left man's left hand. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "minor",
    "observation_index": 4
   },
   {
    "issue_ko": "배경 우측에 위치한 녹색 깃발과 벽면 액자 안의 텍스트가 알아볼 수 없게 뭉개져 있습니다.",
    "fix_en": "Blur the garbled text on the background flag and frame. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "minor",
    "observation_index": 5
   },
   {
    "issue_ko": "노트북 화면이 서장 쪽이 아니라 카메라를 주로 향해 열려 있다",
    "fix_en": "Angle the laptop screen toward the left man. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S33sh2__bgfirst_bg.png",
   "bg_asset_id": "b1596680-3983-4d2d-844e-25cc232f7328",
   "bg_record_key": "S33sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S33sh2::cine": {
  "applied": true,
  "fingerprint": "9cf8930d90a865abc77fdab177f766336dccc27c1c8c2852c0781666c43ffd1c",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S33sh2_sel.png",
  "source_sha256": "a5b5a0e81d739f2c97e52be3b0c856ff4d8eba82cb7c1e77d0b8b078a98cdd24",
  "file": "S33sh2_cine.png",
  "latency_ms": 12678
 },
 "S33sh7::signage": {
  "fp": "f4c2a2eca456ef70",
  "inscriptions": [
   {
    "surface_native": "벽면 액자",
    "text_native": "신뢰받는 경찰, 안전한 사회",
    "reason_ko": "대한민국 경찰서장실 응접 공간의 벽면에 걸려 있는 전형적인 경찰 슬로건 액자로, 극중 배경이 경찰서 내부임을 명확히 보여줍니다."
   }
  ]
 },
 "S33sh7": {
  "input_fingerprint": "786a04f05acf8c02",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 맞은편을 향해 굳은 의지가 담긴 눈빛으로 입술을 뗀 전택수의 상체\n\nLOCATION (lock): In the seated meeting area of the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Tracking along 전택수's near flank from slightly above his seated eye line, the camera settles off his frontal axis in a tight upper-body frame as he lifts his head and parts his lips to answer. His shoulders advance from the chair and his eyes lock onto the 나주경찰서장 beyond the frame, with the sole emphasized shift carried by the newly direct eyeline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground, looks toward 나주경찰서장 opposite the desk.\n- KEY BACKGROUND ELEMENTS: 서장실 책상 가장자리 (두 사람 사이에 놓여 있음) — The near edge enters the lower frame and recedes toward the off-screen 나주경찰서장; used as Its edge remains as a low spatial cue connecting 전택수 to the unseen official opposite him; 열린 노트북 (스페이스바 입력 후 방송 영상이 멈춰 있음) — The display face is angled away from the camera toward the 나주경찰서장; its paused program image is only partly visible; used as A peripheral reminder of the paused broadcast that prompted the confrontation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience remains subdued and neutral-to-cool, preserving natural facial detail without dramatic exaggeration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station-chief office, laptop setup, sofa seating, and daytime light. Exclude the chief's foreground emphasis and frame the opposing officer speaking firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 액자: \"신뢰받는 경찰, 안전한 사회\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 맞은편을 향해 굳은 의지가 담긴 눈빛으로 입술을 뗀 전택수의 상체\n\nLOCATION (lock): In the seated meeting area of the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Tracking along 전택수's near flank from slightly above his seated eye line, the camera settles off his frontal axis in a tight upper-body frame as he lifts his head and parts his lips to answer. His shoulders advance from the chair and his eyes lock onto the 나주경찰서장 beyond the frame, with the sole emphasized shift carried by the newly direct eyeline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground, looks toward 나주경찰서장 opposite the desk.\n- KEY BACKGROUND ELEMENTS: 서장실 책상 가장자리 (두 사람 사이에 놓여 있음) — The near edge enters the lower frame and recedes toward the off-screen 나주경찰서장; used as Its edge remains as a low spatial cue connecting 전택수 to the unseen official opposite him; 열린 노트북 (스페이스바 입력 후 방송 영상이 멈춰 있음) — The display face is angled away from the camera toward the 나주경찰서장; its paused program image is only partly visible; used as A peripheral reminder of the paused broadcast that prompted the confrontation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience remains subdued and neutral-to-cool, preserving natural facial detail without dramatic exaggeration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station-chief office, laptop setup, sofa seating, and daytime light. Exclude the chief's foreground emphasis and frame the opposing officer speaking firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 액자: \"신뢰받는 경찰, 안전한 사회\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 맞은편을 향해 굳은 의지가 담긴 눈빛으로 입술을 뗀 전택수의 상체\n\nLOCATION (lock): In the seated meeting area of the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Tracking along 전택수's near flank from slightly above his seated eye line, the camera settles off his frontal axis in a tight upper-body frame as he lifts his head and parts his lips to answer. His shoulders advance from the chair and his eyes lock onto the 나주경찰서장 beyond the frame, with the sole emphasized shift carried by the newly direct eyeline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground, looks toward 나주경찰서장 opposite the desk.\n- KEY BACKGROUND ELEMENTS: 서장실 책상 가장자리 (두 사람 사이에 놓여 있음) — The near edge enters the lower frame and recedes toward the off-screen 나주경찰서장; used as Its edge remains as a low spatial cue connecting 전택수 to the unseen official opposite him; 열린 노트북 (스페이스바 입력 후 방송 영상이 멈춰 있음) — The display face is angled away from the camera toward the 나주경찰서장; its paused program image is only partly visible; used as A peripheral reminder of the paused broadcast that prompted the confrontation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience remains subdued and neutral-to-cool, preserving natural facial detail without dramatic exaggeration.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station-chief office, laptop setup, sofa seating, and daytime light. Exclude the chief's foreground emphasis and frame the opposing officer speaking firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 액자: \"신뢰받는 경찰, 안전한 사회\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.429,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.429
   },
   "adjusted": {
    "A": 1.571,
    "B": 0.929
   },
   "violations": {
    "B": [
     "[gemini-pro] 배경 요소여야 할 노트북이 전경에서 과도하게 크게 렌더링되어 'never foreground-sized' 지침을 위반함.",
     "[gemini-pro] 기준 이미지에서 서장을 향해 있던 노트북 화면이 반대로 전택수를 향하도록 뒤집혀 소품 방향 지침을 위반함."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1571,
   "B": 929
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "지시된 프레이밍과 피사체 비율을 잘 준수하고 텍스트를 완벽히 구현했으나, 노트북 화면이 지시와 달리 카메라를 향한 점이 아쉬움."
   },
   {
    "label": "B",
    "score": 929,
    "verdict_ko": "배경 요소인 노트북을 전경에 과도하게 크게 배치하고 그 방향마저 반대로 뒤집어 주요 지침을 심각하게 위반함.  ★위반: [gemini-pro] 배경 요소여야 할 노트북이 전경에서 과도하게 크게 렌더링되어 'never foreground-sized' 지침을 위반함. / [gemini-pro] 기준 이미지에서 서장을 향해 있던 노트북 화면이 반대로 전택수를 향하도록 뒤집혀 소품 방향 지침을 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S33sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 좌측 하단의 노트북 화면과 명패가 프롬프트 설정(카메라를 등지고 서장을 향함)과 반대로 카메라와 전택수 방향을 향해 있음.",
     "fix_en": "Redraw the laptop and nameplate on the desk to face away from the camera; render the dark back cover of the laptop lid and the blank back of the nameplate facing forward, showing only a shallow-focus sliver of the paused screen at an oblique angle, preserving the man, his pose, his clothing, the desk, and the office background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "숏 텍스트에 '입술을 뗀'이라고 명시되어 있으나 인물의 입이 굳게 닫혀 있음.",
     "fix_en": "Redraw the man's mouth to have slightly parted lips, preserving his facial structure, determined expression, clothing, and the background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "지정된 타이트 상체 클로즈업이 아니라 상반신과 사무실 배경이 넓게 잡힌 미디엄에 가깝다.",
     "fix_en": "Reframe the shot to an upper-body close-up, preserving the man's appearance and the lighting.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전택수가 프레임 정중앙이 아니라 오른쪽에 치우쳐 있다.",
     "fix_en": "Shift the man into the center of the frame, preserving his pose, clothing, and the office setting.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 좌측 하단의 노트북 화면과 명패가 프롬프트 설정(카메라를 등지고 서장을 향함)과 반대로 카메라와 전택수 방향을 향해 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "숏 텍스트에 '입술을 뗀'이라고 명시되어 있으나 인물의 입이 굳게 닫혀 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "지정된 타이트 상체 클로즈업이 아니라 상반신과 사무실 배경이 넓게 잡힌 미디엄에 가깝다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 프레임 정중앙이 아니라 오른쪽에 치우쳐 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "노트북 화면이 서장 쪽으로 외면되어 일부만 보여야 하는데 카메라를 향해 열려 방송 영상이 거의 전부 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "숏 텍스트에 '입술을 뗀'이라고 명시되어 있으나 인물의 입이 굳게 닫혀 있음.",
    "fix_en": "Redraw the man's mouth to have slightly parted lips, preserving his facial structure, determined expression, clothing, and the background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "지정된 타이트 상체 클로즈업이 아니라 상반신과 사무실 배경이 넓게 잡힌 미디엄에 가깝다.",
    "fix_en": "Reframe the shot to an upper-body close-up, preserving the man's appearance and the lighting.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "전택수가 프레임 정중앙이 아니라 오른쪽에 치우쳐 있다.",
    "fix_en": "Shift the man into the center of the frame, preserving his pose, clothing, and the office setting.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the laptop and nameplate on the desk to face away from the camera; render the dark back cover of the laptop lid and the blank back of the nameplate facing forward, showing only a shallow-focus sliver of the paused screen at an oblique angle, preserving the man, his pose, his clothing, the desk, and the office background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "프롬프트의 지시대로 입술을 뗀 전택수의 표정과 결연한 시선이 잘 표현되었으며, 액자 텍스트, 명패, 방송 화면이 켜진 노트북 등 모든 요소가 제자리에 정확히 구현되었습니다."
     },
     {
      "label": "B",
      "score": 1194,
      "verdict_ko": "전택수의 입술이 닫혀 있어 연기 지시를 누락했으며, 책상 위에 있어야 할 노트북이 왼쪽 소파 방석 위에 잘못 배치되는 배치 오류가 발생했습니다.  ★위반: [gemini-pro] 소파 위에 잘못된 위치의 노트북이 생성됨 (invented/misplaced object)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 1.444
     },
     "adjusted": {
      "A": 1.714,
      "B": 1.194
     },
     "violations": {
      "B": [
       "[gemini-pro] 소파 위에 잘못된 위치의 노트북이 생성됨 (invented/misplaced object)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "프롬프트의 지시대로 입술을 뗀 전택수의 표정과 결연한 시선이 잘 표현되었으며, 액자 텍스트, 명패, 방송 화면이 켜진 노트북 등 모든 요소가 제자리에 정확히 구현되었습니다."
     },
     {
      "label": "B",
      "score": 1194,
      "verdict_ko": "전택수의 입술이 닫혀 있어 연기 지시를 누락했으며, 책상 위에 있어야 할 노트북이 왼쪽 소파 방석 위에 잘못 배치되는 배치 오류가 발생했습니다.  ★위반: [gemini-pro] 소파 위에 잘못된 위치의 노트북이 생성됨 (invented/misplaced object)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "시점이 바뀌면서 명패와 노트북의 뒷면이 보이도록 공간 방향을 올바르게 구현했으나, 지시된 노트북 화면의 일부가 보이지 않는 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "이전 샷의 명패 정면과 노트북 화면을 그대로 복사해 붙여넣어 변경된 카메라 앵글의 3D 공간 및 방향성과 완전히 모순되는 심각한 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수가 고개를 들고 화면 밖 맞은편을 응시하고 있습니다.",
      "built_space": "경찰서장실의 배경과 책상, 소파가 제대로 배치되어 있으며, 변경된 카메라 시점에 맞게 노트북 뒷면과 명패의 뒷면이 보입니다.",
      "entities": "전택수의 얼굴, 모자, 복장이 레퍼런스와 일치하며 벽면 액자의 글귀('신뢰받는 경찰, 안전한 사회')도 정확히 구현되었습니다.",
      "hard_violations": [],
      "physics": "소파에 앉아 있는 자세가 자연스러우며, 책상 위 사물들도 안정적으로 놓여 있습니다."
     },
     {
      "label": "B",
      "direction": "전택수가 화면 밖 맞은편을 향해 시선을 두고 있습니다.",
      "built_space": "배경은 일치하나, 책상 위 노트북 화면과 명패 정면이 카메라를 향해 있어 시점상 불가능한 공간 배치를 보입니다.",
      "entities": "인물의 외형과 복장, 액자의 텍스트는 지시사항과 잘 일치합니다.",
      "hard_violations": [
       "변경된 카메라 시점과 무관하게 이전 샷의 사물(노트북 화면, 명패 정면)을 그대로 붙여넣어 물리적으로 불가능한 스테이징 발생"
      ],
      "physics": "인물의 앉은 자세는 자연스럽게 지탱되고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "시점이 바뀌면서 명패와 노트북의 뒷면이 보이도록 공간 방향을 올바르게 구현했으나, 지시된 노트북 화면의 일부가 보이지 않는 점이 아쉽습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "이전 샷의 명패 정면과 노트북 화면을 그대로 복사해 붙여넣어 변경된 카메라 앵글의 3D 공간 및 방향성과 완전히 모순되는 심각한 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수가 고개를 들고 화면 밖 맞은편을 응시하고 있습니다.",
      "built_space": "경찰서장실의 배경과 책상, 소파가 제대로 배치되어 있으며, 변경된 카메라 시점에 맞게 노트북 뒷면과 명패의 뒷면이 보입니다.",
      "entities": "전택수의 얼굴, 모자, 복장이 레퍼런스와 일치하며 벽면 액자의 글귀('신뢰받는 경찰, 안전한 사회')도 정확히 구현되었습니다.",
      "hard_violations": [],
      "physics": "소파에 앉아 있는 자세가 자연스러우며, 책상 위 사물들도 안정적으로 놓여 있습니다."
     },
     {
      "label": "A",
      "direction": "전택수가 화면 밖 맞은편을 향해 시선을 두고 있습니다.",
      "built_space": "배경은 일치하나, 책상 위 노트북 화면과 명패 정면이 카메라를 향해 있어 시점상 불가능한 공간 배치를 보입니다.",
      "entities": "인물의 외형과 복장, 액자의 텍스트는 지시사항과 잘 일치합니다.",
      "hard_violations": [
       "변경된 카메라 시점과 무관하게 이전 샷의 사물(노트북 화면, 명패 정면)을 그대로 붙여넣어 물리적으로 불가능한 스테이징 발생"
      ],
      "physics": "인물의 앉은 자세는 자연스럽게 지탱되고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1717,
     "B": 1201
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S33sh2"
  }
 },
 "S33sh7::cine": {
  "applied": true,
  "fingerprint": "2372ddad36f2395d2488b6241f3ac1274a3343d81613369d022f4dfe29fe7387",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S33sh7_sel.png",
  "source_sha256": "ed2948c2595559df75a738678b46e5e17058de5f99d6e7a355459e8fc877e663",
  "file": "S33sh7_cine.png",
  "latency_ms": 10572
 },
 "S34sh1::confined_fp_apt": {
  "applies": false,
  "reason_ko": "카메라가 차량 앞유리 너머에서 운전대를 잡은 인물을 바라보는 외부 시점의 샷이므로 차량 내부 공간에 카메라가 위치하지 않으며, 일반적인 승용차의 단일 운전자 위치는 복잡한 좌석 배치를 요구하지 않아 평면도 레이아웃 보조가 불필요합니다.",
  "input_fingerprint": "0461c0a4f1158f7a"
 },
 "S34sh1::signage": {
  "fp": "dfcc14ec9c2315cd",
  "inscriptions": [
   {
    "surface_native": "정문 현판",
    "text_native": "광주지방검찰청",
    "reason_ko": "주인공이 진입하는 장소가 광주지방검찰청임을 명확히 보여주기 위해 정문 기둥의 기관 명판에 한글 표기가 필요합니다."
   }
  ]
 },
 "era_assess::72bfd36da0eab2c1": {
  "subjects": [
   {
    "subject_native": "대한민국 검찰청 정문 (2001년–2017년)",
    "search_terms_native": [
     "검찰청 정문",
     "지방검찰청 정문",
     "검찰청 입구 차단기"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색되어야 하며 다른 언어의 단어를 혼용해서는 안 됩니다.",
    "reason_ko": "대한민국 검찰청 특유의 정문 구조, 보안 초소, 로고 표지판 및 차량 차단기의 형태는 일반적인 해외 관공서나 현대식 정문과 뚜렷이 구분되므로 고증이 필요합니다."
   }
  ]
 },
 "era_ref::3d7b8b2145d2c932": {
  "subject": "대한민국 검찰청 정문 (2001년–2017년)",
  "terms": [
   "검찰청 정문",
   "지방검찰청 정문",
   "검찰청 입구 차단기"
  ],
  "queries": [
   [
    "대한민국 검찰청 정문 지방검찰청 정문 2001년 2017년",
    "검찰청 입구 차단기 정문"
   ],
   [
    "검찰청 정문 사진 2001년",
    "검찰청 정문 사진 2017년",
    "지방검찰청 정문 차단기 사진"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://img.cztv.com/platform/tpnews/cms_1762571102_XZbw92GFe7.png",
  "picked_reason_ko": "검찰청 정문의 석재 문주, 경비 공간, 금속 차단문과 출입 통제 설비가 정면에서 선명하게 보여 형태와 재료를 가장 충실히 읽을 수 있다.",
  "sha256": "0e4fa54b3b3d3a103f128950d9acda54349672550fa7cd664e69aef5ae596968",
  "file": "eraref_3d7b8b2145d2c932.png"
 },
 "S34sh1::bgfirst_bg": {
  "input_fingerprint": "49267a367c97927f",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 밝은 대낮, 광주지방검찰청 정문을 반쯤 통과한 차량 앞유리 너머로 굳은 표정의 전택수가 운전대를 쥔 모습이 보이는 전경\n\nLOCATION (lock): Inside the car at the driver's seat as it passes through the prosecution office's exterior main gate; the windshield, wheel, mirrors, and dashboard closely frame the driver.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just inside the main gate at chest height, the camera tracks backward along the car's front three-quarter axis, holding enough medium-long distance to show the vehicle halfway across the boundary. Through the windshield, 전택수 sits near center with both hands firm on the wheel and his gaze fixed along the car's path; the emphasized change is the vehicle's position advancing from outside to inside the gate.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 inside the entering car in the middle-center of the frame, midground, moves toward inside the prosecution compound; main gate boundary in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방검찰청 정문 (차량이 반쯤 통과 중) — The gate sides flank the car while the passage opens inward behind the camera; used as Frames the vehicle at the exact threshold and makes the institutional crossing legible; 차량 앞유리 (전택수의 얼굴과 운전대가 보임) — The windshield faces the camera on a front three-quarter angle, with 전택수 visible beyond it; used as Acts as the transparent layer through which 전택수's expression and grip remain visible; 전택수의 차 (정문으로 진입 중) — Its front and one side face the retreating camera; used as Carries the forward movement through the gate while preserving realistic scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime natural light is held to moderate contrast and a restrained neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.\n\nPERIOD REFERENCE — 대한민국 검찰청 정문 (2001년–2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 밝은 대낮, 광주지방검찰청 정문을 반쯤 통과한 차량 앞유리 너머로 굳은 표정의 전택수가 운전대를 쥔 모습이 보이는 전경\n\nLOCATION (lock): Inside the car at the driver's seat as it passes through the prosecution office's exterior main gate; the windshield, wheel, mirrors, and dashboard closely frame the driver.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just inside the main gate at chest height, the camera tracks backward along the car's front three-quarter axis, holding enough medium-long distance to show the vehicle halfway across the boundary. Through the windshield, 전택수 sits near center with both hands firm on the wheel and his gaze fixed along the car's path; the emphasized change is the vehicle's position advancing from outside to inside the gate.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 inside the entering car in the middle-center of the frame, midground, moves toward inside the prosecution compound; main gate boundary in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방검찰청 정문 (차량이 반쯤 통과 중) — The gate sides flank the car while the passage opens inward behind the camera; used as Frames the vehicle at the exact threshold and makes the institutional crossing legible; 차량 앞유리 (전택수의 얼굴과 운전대가 보임) — The windshield faces the camera on a front three-quarter angle, with 전택수 visible beyond it; used as Acts as the transparent layer through which 전택수's expression and grip remain visible; 전택수의 차 (정문으로 진입 중) — Its front and one side face the retreating camera; used as Carries the forward movement through the gate while preserving realistic scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime natural light is held to moderate contrast and a restrained neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.\n\nPERIOD REFERENCE — 대한민국 검찰청 정문 (2001년–2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S34sh1__bgfirst_bg.png",
  "asset_id": "fc19036e-3169-44cc-8535-53f033eef6fa",
  "input_asset_ids": [
   "86cee38b-89da-4e95-bf67-5a7acf620533",
   "342cd5b6-5822-46d2-8ddd-7cc796442c66",
   "07e080a5-2e85-45ba-8a25-a23a69d53eb2"
  ],
  "era_research": {
   "subject": "대한민국 검찰청 정문 (2001년–2017년)",
   "queries": [
    [
     "대한민국 검찰청 정문 지방검찰청 정문 2001년 2017년",
     "검찰청 입구 차단기 정문"
    ],
    [
     "검찰청 정문 사진 2001년",
     "검찰청 정문 사진 2017년",
     "지방검찰청 정문 차단기 사진"
    ]
   ],
   "picked_url": "https://img.cztv.com/platform/tpnews/cms_1762571102_XZbw92GFe7.png",
   "sha256": "0e4fa54b3b3d3a103f128950d9acda54349672550fa7cd664e69aef5ae596968",
   "file": "eraref_3d7b8b2145d2c932.png"
  }
 },
 "S34sh1": {
  "input_fingerprint": "37c1f9fa43db961d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밝은 대낮, 광주지방검찰청 정문을 반쯤 통과한 차량 앞유리 너머로 굳은 표정의 전택수가 운전대를 쥔 모습이 보이는 전경\n\nLOCATION (lock): Inside the car at the driver's seat as it passes through the prosecution office's exterior main gate; the windshield, wheel, mirrors, and dashboard closely frame the driver. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just inside the main gate at chest height, the camera tracks backward along the car's front three-quarter axis, holding enough medium-long distance to show the vehicle halfway across the boundary. Through the windshield, 전택수 sits near center with both hands firm on the wheel and his gaze fixed along the car's path; the emphasized change is the vehicle's position advancing from outside to inside the gate.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 inside the entering car in the middle-center of the frame, midground, moves toward inside the prosecution compound; main gate boundary in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방검찰청 정문 (차량이 반쯤 통과 중) — The gate sides flank the car while the passage opens inward behind the camera; used as Frames the vehicle at the exact threshold and makes the institutional crossing legible; 차량 앞유리 (전택수의 얼굴과 운전대가 보임) — The windshield faces the camera on a front three-quarter angle, with 전택수 visible beyond it; used as Acts as the transparent layer through which 전택수's expression and grip remain visible; 전택수의 차 (정문으로 진입 중) — Its front and one side face the retreating camera; used as Carries the forward movement through the gate while preserving realistic scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime natural light is held to moderate contrast and a restrained neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 정문 현판: \"광주지방검찰청\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective, every panel, control, seat, mirror, window and fixture stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: how many steering wheels or control surfaces, how many seats and which way each faces, where each mirror sits and what it could reflect from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second rim, sliding a seat, turning a mirror or growing a new panel is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: which seat the body occupies, which side of the wheel the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밝은 대낮, 광주지방검찰청 정문을 반쯤 통과한 차량 앞유리 너머로 굳은 표정의 전택수가 운전대를 쥔 모습이 보이는 전경\n\nLOCATION (lock): Inside the car at the driver's seat as it passes through the prosecution office's exterior main gate; the windshield, wheel, mirrors, and dashboard closely frame the driver. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just inside the main gate at chest height, the camera tracks backward along the car's front three-quarter axis, holding enough medium-long distance to show the vehicle halfway across the boundary. Through the windshield, 전택수 sits near center with both hands firm on the wheel and his gaze fixed along the car's path; the emphasized change is the vehicle's position advancing from outside to inside the gate.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 inside the entering car in the middle-center of the frame, midground, moves toward inside the prosecution compound; main gate boundary in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방검찰청 정문 (차량이 반쯤 통과 중) — The gate sides flank the car while the passage opens inward behind the camera; used as Frames the vehicle at the exact threshold and makes the institutional crossing legible; 차량 앞유리 (전택수의 얼굴과 운전대가 보임) — The windshield faces the camera on a front three-quarter angle, with 전택수 visible beyond it; used as Acts as the transparent layer through which 전택수's expression and grip remain visible; 전택수의 차 (정문으로 진입 중) — Its front and one side face the retreating camera; used as Carries the forward movement through the gate while preserving realistic scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime natural light is held to moderate contrast and a restrained neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 정문 현판: \"광주지방검찰청\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밝은 대낮, 광주지방검찰청 정문을 반쯤 통과한 차량 앞유리 너머로 굳은 표정의 전택수가 운전대를 쥔 모습이 보이는 전경\n\nLOCATION (lock): Inside the car at the driver's seat as it passes through the prosecution office's exterior main gate; the windshield, wheel, mirrors, and dashboard closely frame the driver. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just inside the main gate at chest height, the camera tracks backward along the car's front three-quarter axis, holding enough medium-long distance to show the vehicle halfway across the boundary. Through the windshield, 전택수 sits near center with both hands firm on the wheel and his gaze fixed along the car's path; the emphasized change is the vehicle's position advancing from outside to inside the gate.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 inside the entering car in the middle-center of the frame, midground, moves toward inside the prosecution compound; main gate boundary in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방검찰청 정문 (차량이 반쯤 통과 중) — The gate sides flank the car while the passage opens inward behind the camera; used as Frames the vehicle at the exact threshold and makes the institutional crossing legible; 차량 앞유리 (전택수의 얼굴과 운전대가 보임) — The windshield faces the camera on a front three-quarter angle, with 전택수 visible beyond it; used as Acts as the transparent layer through which 전택수's expression and grip remain visible; 전택수의 차 (정문으로 진입 중) — Its front and one side face the retreating camera; used as Carries the forward movement through the gate while preserving realistic scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright daytime natural light is held to moderate contrast and a restrained neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 정문 현판: \"광주지방검찰청\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S34sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched panels, seats, controls and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S34sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L33B01.png"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_prosecutors_office_sel.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 STRUCTURE LOOK 레퍼런스의 건물 특징과 LOCATION의 정문 공간을 정확히 조합하였으며, 인물의 의상 및 지정된 텍스트를 충실하게 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 건물 형태를 완전히 무시하였으며, 프롬프트와 무관한 잘못된 텍스트가 노출되어 지침 위반이 명확함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "운전자의 시선과 차량의 진행 방향 모두 전방을 향하고 있음.",
      "built_space": "차량 내부 좌측에 운전석과 1개의 스티어링 휠이 있음. 그러나 배경의 건물이 STRUCTURE LOOK 레퍼런스와 전혀 다르게 렌더링됨.",
      "entities": "인물은 레퍼런스 이미지의 모자를 착용하지 않음. 좌측 기둥에 지정되지 않은 '서울고등검찰청' 등의 잘못된 텍스트가 나타남.",
      "hard_violations": [
       "STRUCTURE LOOK 구조물 레퍼런스 불일치 (완전히 다른 형태의 건물)",
       "잘못된 텍스트 노출 (서울고등검찰청)"
      ],
      "physics": "운전자의 두 손이 스티어링 휠을 잡고 있으며, 차량은 지면에 정상적으로 닿아 있음."
     },
     {
      "label": "B",
      "direction": "차량이 정문을 통과하며 카메라를 향해 나아오고, 운전자는 전방 도로를 응시함.",
      "built_space": "좌측 운전석과 1개의 스티어링 휠이 확인됨. STRUCTURE LOOK의 원통형 지붕 구조와 유리 배열, LOCATION의 경비실과 정문 배치가 올바르게 구현됨.",
      "entities": "인물은 레퍼런스와 동일한 의상과 모자를 착용함. 건물에 지정된 '광주지방검찰청' 텍스트가 정확한 철자로 표기됨.",
      "hard_violations": [],
      "physics": "양손이 운전대를 단단히 쥐고 있으며, 차량의 바퀴가 지면에 정상적으로 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 STRUCTURE LOOK 레퍼런스의 건물 특징과 LOCATION의 정문 공간을 정확히 조합하였으며, 인물의 의상 및 지정된 텍스트를 충실하게 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 건물 형태를 완전히 무시하였으며, 프롬프트와 무관한 잘못된 텍스트가 노출되어 지침 위반이 명확함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "운전자의 시선과 차량의 진행 방향 모두 전방을 향하고 있음.",
      "built_space": "차량 내부 좌측에 운전석과 1개의 스티어링 휠이 있음. 그러나 배경의 건물이 STRUCTURE LOOK 레퍼런스와 전혀 다르게 렌더링됨.",
      "entities": "인물은 레퍼런스 이미지의 모자를 착용하지 않음. 좌측 기둥에 지정되지 않은 '서울고등검찰청' 등의 잘못된 텍스트가 나타남.",
      "hard_violations": [
       "STRUCTURE LOOK 구조물 레퍼런스 불일치 (완전히 다른 형태의 건물)",
       "잘못된 텍스트 노출 (서울고등검찰청)"
      ],
      "physics": "운전자의 두 손이 스티어링 휠을 잡고 있으며, 차량은 지면에 정상적으로 닿아 있음."
     },
     {
      "label": "B",
      "direction": "차량이 정문을 통과하며 카메라를 향해 나아오고, 운전자는 전방 도로를 응시함.",
      "built_space": "좌측 운전석과 1개의 스티어링 휠이 확인됨. STRUCTURE LOOK의 원통형 지붕 구조와 유리 배열, LOCATION의 경비실과 정문 배치가 올바르게 구현됨.",
      "entities": "인물은 레퍼런스와 동일한 의상과 모자를 착용함. 건물에 지정된 '광주지방검찰청' 텍스트가 정확한 철자로 표기됨.",
      "hard_violations": [],
      "physics": "양손이 운전대를 단단히 쥐고 있으며, 차량의 바퀴가 지면에 정상적으로 지지됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 4분의 3 측면 카메라 앵글, 인물의 의상(모자 포함), 그리고 레퍼런스의 정문 구조와 건물 형태를 매우 충실하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지시되지 않은 다른 지역의 텍스트가 기둥에 생성되었고, 카메라 앵글, 정문 구조, 인물의 의상이 모두 프롬프트 및 레퍼런스와 불일치합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 차량의 주행 방향인 앞을 향하고 있으며, 차량은 정문 안쪽으로 진입하는 방향을 향함.",
      "built_space": "차량 내부에 1개의 운전대가 있고 좌측 운전석에 인물이 올바르게 탑승함. 배경의 정문 초소와 철책이 위치 레퍼런스와 일치하며 건물도 구조 레퍼런스와 부합함.",
      "entities": "전택수(레퍼런스와 일치하는 모자, 자켓 착용), 차량, 광주지방검찰청 건물 및 정문. 지시된 '광주지방검찰청' 텍스트가 건물에 표현됨.",
      "hard_violations": [],
      "physics": "인물의 두 손이 운전대를 단단히 쥐고 지탱하며, 차량은 지면에 정상적으로 닿아 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 정면을 향하고 있으며, 차량은 4분의 3 측면이 아닌 정면 카메라를 향해 있음.",
      "built_space": "차량 내부에 1개의 운전대가 있고 좌측 운전석에 탑승함. 배경의 정문 기둥 및 주변 환경이 위치 레퍼런스(경비실 부재 등)와 전혀 다름.",
      "entities": "전택수(모자 누락 및 의상 다름), 차량. 정문 기둥에 프롬프트에 없는 '서울고등검찰청' 등의 텍스트가 나타남.",
      "hard_violations": [
       "지시되지 않은 발명된 텍스트(서울고등검찰청 등) 생성",
       "위치 레퍼런스의 정문 구조 및 환경(LOCKED) 무시 및 불일치"
      ],
      "physics": "인물의 손이 운전대를 쥐고 있으며, 차량은 지면 위에 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 4분의 3 측면 카메라 앵글, 인물의 의상(모자 포함), 그리고 레퍼런스의 정문 구조와 건물 형태를 매우 충실하게 재현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시되지 않은 다른 지역의 텍스트가 기둥에 생성되었고, 카메라 앵글, 정문 구조, 인물의 의상이 모두 프롬프트 및 레퍼런스와 불일치합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수의 시선은 차량의 주행 방향인 앞을 향하고 있으며, 차량은 정문 안쪽으로 진입하는 방향을 향함.",
      "built_space": "차량 내부에 1개의 운전대가 있고 좌측 운전석에 인물이 올바르게 탑승함. 배경의 정문 초소와 철책이 위치 레퍼런스와 일치하며 건물도 구조 레퍼런스와 부합함.",
      "entities": "전택수(레퍼런스와 일치하는 모자, 자켓 착용), 차량, 광주지방검찰청 건물 및 정문. 지시된 '광주지방검찰청' 텍스트가 건물에 표현됨.",
      "hard_violations": [],
      "physics": "인물의 두 손이 운전대를 단단히 쥐고 지탱하며, 차량은 지면에 정상적으로 닿아 있음."
     },
     {
      "label": "A",
      "direction": "전택수의 시선은 정면을 향하고 있으며, 차량은 4분의 3 측면이 아닌 정면 카메라를 향해 있음.",
      "built_space": "차량 내부에 1개의 운전대가 있고 좌측 운전석에 탑승함. 배경의 정문 기둥 및 주변 환경이 위치 레퍼런스(경비실 부재 등)와 전혀 다름.",
      "entities": "전택수(모자 누락 및 의상 다름), 차량. 정문 기둥에 프롬프트에 없는 '서울고등검찰청' 등의 텍스트가 나타남.",
      "hard_violations": [
       "지시되지 않은 발명된 텍스트(서울고등검찰청 등) 생성",
       "위치 레퍼런스의 정문 구조 및 환경(LOCKED) 무시 및 불일치"
      ],
      "physics": "인물의 손이 운전대를 쥐고 있으며, 차량은 지면 위에 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "운전자의 시선과 차량의 진행 방향 모두 전방을 향하고 있음.",
    "built_space": "차량 내부 좌측에 운전석과 1개의 스티어링 휠이 있음. 그러나 배경의 건물이 STRUCTURE LOOK 레퍼런스와 전혀 다르게 렌더링됨.",
    "entities": "인물은 레퍼런스 이미지의 모자를 착용하지 않음. 좌측 기둥에 지정되지 않은 '서울고등검찰청' 등의 잘못된 텍스트가 나타남.",
    "hard_violations": [
     "STRUCTURE LOOK 구조물 레퍼런스 불일치 (완전히 다른 형태의 건물)",
     "잘못된 텍스트 노출 (서울고등검찰청)"
    ],
    "physics": "운전자의 두 손이 스티어링 휠을 잡고 있으며, 차량은 지면에 정상적으로 닿아 있음."
   },
   {
    "label": "B",
    "direction": "차량이 정문을 통과하며 카메라를 향해 나아오고, 운전자는 전방 도로를 응시함.",
    "built_space": "좌측 운전석과 1개의 스티어링 휠이 확인됨. STRUCTURE LOOK의 원통형 지붕 구조와 유리 배열, LOCATION의 경비실과 정문 배치가 올바르게 구현됨.",
    "entities": "인물은 레퍼런스와 동일한 의상과 모자를 착용함. 건물에 지정된 '광주지방검찰청' 텍스트가 정확한 철자로 표기됨.",
    "hard_violations": [],
    "physics": "양손이 운전대를 단단히 쥐고 있으며, 차량의 바퀴가 지면에 정상적으로 지지됨."
   }
  ],
  "totals": {
   "A": 6,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 STRUCTURE LOOK 레퍼런스의 건물 특징과 LOCATION의 정문 공간을 정확히 조합하였으며, 인물의 의상 및 지정된 텍스트를 충실하게 구현함."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "지정된 건물 형태를 완전히 무시하였으며, 프롬프트와 무관한 잘못된 텍스트가 노출되어 지침 위반이 명확함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L33B01.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_prosecutors_office_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "운전대를 잡은 전택수의 왼손(화면 우측) 손가락이 운전대와 융합되어 형태가 무너짐.",
     "fix_en": "Redraw the driver's left hand to clearly separate the fingers from the steering wheel. Preserve the driver's face, clothing, lighting, background, and car interior.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "구조물 레퍼런스와 다르게 배경 건물 1층 출입구 캐노피 정면에 '광주지방검찰청' 대형 텍스트가 임의로 추가됨.",
     "fix_en": "Remove the text from the background building's canopy, filling the space with bare stone texture. Preserve the driver, car, gate, lighting, and framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "프롬프트에 명시된 '정문 현판'이 정문 구조물 어디에도 묘사되지 않음.",
     "fix_en": "Add a physical signboard with the text '광주지방검찰청' to the gate structure. Maintain the car, driver, background building, lighting, and framing.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "차량이 청사를 등지고 바깥 도로·카메라 쪽을 향해 정문 진입이 아니라 진출 중이다.",
     "fix_en": "Replace the car's front with its rear exterior to depict it driving away into the compound, removing the driver's face. Preserve the background building, gate, lighting, and framing.",
     "severity": "critical",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경 고정 건물이 구조 참조와 달리 전면에 다른 석조 출입 포치가 붙어 형태가 다르다.",
     "fix_en": "Replace the protruding stone porch with a recessed glass entryway under a flat canopy. Preserve the car, driver, gate, lighting, and framing.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "차가 게이트를 반쯤 통과한 위치가 아니라 앞부분이 이미 바깥 전경을 크게 차지한다.",
     "fix_en": "Push the car deeper into the scene so it sits halfway through the gate, filling the foreground with asphalt. Maintain the driver, car model, background, and lighting.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "한국 배경임에도 차량이 우핸들로 묘사되어 전택수가 차량 우측 좌석에 앉아있음.",
     "severity": "major"
    },
    {
     "issue_ko": "운전대를 잡은 전택수의 왼손(화면 우측) 손가락이 운전대와 융합되어 형태가 무너짐.",
     "severity": "major"
    },
    {
     "issue_ko": "구조물 레퍼런스와 다르게 배경 건물 1층 출입구 캐노피 정면에 '광주지방검찰청' 대형 텍스트가 임의로 추가됨.",
     "severity": "major"
    },
    {
     "issue_ko": "프롬프트에 명시된 '정문 현판'이 정문 구조물 어디에도 묘사되지 않음.",
     "severity": "major"
    },
    {
     "issue_ko": "차량이 청사를 등지고 바깥 도로·카메라 쪽을 향해 정문 진입이 아니라 진출 중이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경 고정 건물이 구조 참조와 달리 전면에 다른 석조 출입 포치가 붙어 형태가 다르다.",
     "severity": "major"
    },
    {
     "issue_ko": "차가 게이트를 반쯤 통과한 위치가 아니라 앞부분이 이미 바깥 전경을 크게 차지한다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "운전대를 잡은 전택수의 왼손(화면 우측) 손가락이 운전대와 융합되어 형태가 무너짐.",
    "fix_en": "Redraw the driver's left hand to clearly separate the fingers from the steering wheel. Preserve the driver's face, clothing, lighting, background, and car interior.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "구조물 레퍼런스와 다르게 배경 건물 1층 출입구 캐노피 정면에 '광주지방검찰청' 대형 텍스트가 임의로 추가됨.",
    "fix_en": "Remove the text from the background building's canopy, filling the space with bare stone texture. Preserve the driver, car, gate, lighting, and framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "프롬프트에 명시된 '정문 현판'이 정문 구조물 어디에도 묘사되지 않음.",
    "fix_en": "Add a physical signboard with the text '광주지방검찰청' to the gate structure. Maintain the car, driver, background building, lighting, and framing.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "배경 고정 건물이 구조 참조와 달리 전면에 다른 석조 출입 포치가 붙어 형태가 다르다.",
    "fix_en": "Replace the protruding stone porch with a recessed glass entryway under a flat canopy. Preserve the car, driver, gate, lighting, and framing.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "차가 게이트를 반쯤 통과한 위치가 아니라 앞부분이 이미 바깥 전경을 크게 차지한다.",
    "fix_en": "Push the car deeper into the scene so it sits halfway through the gate, filling the foreground with asphalt. Maintain the driver, car model, background, and lighting.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the car's front with its rear exterior to depict it driving away into the compound, removing the driver's face. Preserve the background building, gate, lighting, and framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 전면 4분의 3 측면 카메라 앵글을 정확히 구현하여 앞유리 너머로 전택수의 신원과 운전 자세를 훌륭하게 포착해 프롬프트의 최우선 요구사항을 충족함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "차량의 후면을 비추어 앞유리와 지정된 인물인 전택수가 전혀 프레임에 나타나지 않아 카메라 구도 및 샷 텍스트의 핵심 연출을 완전히 실패함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "운전자는 정면 경로를 응시하고 있으며, 차량은 정문을 반쯤 통과하여 카메라를 향해 앞으로 전진 중임.",
      "built_space": "정문 기둥과 경비 초소 사이에 차량이 위치하며, 차량 내부 좌측 좌석에 1개의 운전대가 있고 운전자가 올바르게 착석해 있음.",
      "entities": "전택수의 얼굴, 연령대, 모자와 복장이 레퍼런스와 일치하며, 지정된 검찰청 건물(구조 레퍼런스 형태 반영)과 정문이 확인됨.",
      "hard_violations": [],
      "physics": "차량은 지면에 안정적으로 닿아 이동 중이며, 운전자의 양손은 운전대를 물리적으로 단단히 쥐고 지탱함."
     },
     {
      "label": "B",
      "direction": "차량은 카메라를 등지고 건물 쪽을 향해 진입하고 있음.",
      "built_space": "차량의 후면만 보여 내부의 운전석, 운전대 등의 구조를 확인할 수 없음.",
      "entities": "지정된 장소의 건물과 초소, 차량은 존재하나, 프롬프트가 요구한 핵심 인물(전택수)이 프레임 내에 전혀 존재하지 않음.",
      "hard_violations": [],
      "physics": "차량은 지면에 닿아 정상적으로 위치해 있으나, 인물의 물리적 접촉이나 지탱은 확인할 수 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 전면 4분의 3 측면 카메라 앵글을 정확히 구현하여 앞유리 너머로 전택수의 신원과 운전 자세를 훌륭하게 포착해 프롬프트의 최우선 요구사항을 충족함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "차량의 후면을 비추어 앞유리와 지정된 인물인 전택수가 전혀 프레임에 나타나지 않아 카메라 구도 및 샷 텍스트의 핵심 연출을 완전히 실패함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "운전자는 정면 경로를 응시하고 있으며, 차량은 정문을 반쯤 통과하여 카메라를 향해 앞으로 전진 중임.",
      "built_space": "정문 기둥과 경비 초소 사이에 차량이 위치하며, 차량 내부 좌측 좌석에 1개의 운전대가 있고 운전자가 올바르게 착석해 있음.",
      "entities": "전택수의 얼굴, 연령대, 모자와 복장이 레퍼런스와 일치하며, 지정된 검찰청 건물(구조 레퍼런스 형태 반영)과 정문이 확인됨.",
      "hard_violations": [],
      "physics": "차량은 지면에 안정적으로 닿아 이동 중이며, 운전자의 양손은 운전대를 물리적으로 단단히 쥐고 지탱함."
     },
     {
      "label": "B",
      "direction": "차량은 카메라를 등지고 건물 쪽을 향해 진입하고 있음.",
      "built_space": "차량의 후면만 보여 내부의 운전석, 운전대 등의 구조를 확인할 수 없음.",
      "entities": "지정된 장소의 건물과 초소, 차량은 존재하나, 프롬프트가 요구한 핵심 인물(전택수)이 프레임 내에 전혀 존재하지 않음.",
      "hard_violations": [],
      "physics": "차량은 지면에 닿아 정상적으로 위치해 있으나, 인물의 물리적 접촉이나 지탱은 확인할 수 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 앞유리 너머의 전측면 앵글, 인물의 묘사, 참조 구조물 건물의 결합을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시된 카메라 앵글과 피사체를 완전히 누락하고 위치 참조 사진의 구도를 단순 복제하여 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "운전자의 시선은 정면을 향함. 차량은 카메라 쪽으로 다가오며 진행 중임.",
      "built_space": "좌측 운전석에 인물이 탑승함. 위치 참조의 정문과 구조물 참조의 건물이 지시대로 배치됨.",
      "entities": "전택수(참조 이미지의 얼굴, 모자, 재킷 일치), 운전대, 차량, 광주지방검찰청 건물과 간판 모두 확인됨.",
      "hard_violations": [],
      "physics": "운전자의 두 손이 운전대를 안정적으로 잡고 있음."
     },
     {
      "label": "A",
      "direction": "차량이 카메라를 등지고 멀어짐. 운전자의 시선 확인 불가.",
      "built_space": "위치 참조 사진의 정문과 배경 건물만 보이며, 구조물 참조 사진의 외관은 무시됨.",
      "entities": "차량 후면만 보이며, 필수 인물(전택수)과 내부 요소가 전혀 없음.",
      "hard_violations": [
       "필수 피사체(전택수) 누락",
       "명시된 전측면 앞유리 카메라 앵글 위반"
      ],
      "physics": "차량이 바닥에 지지되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 앞유리 너머의 전측면 앵글, 인물의 묘사, 참조 구조물 건물의 결합을 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지시된 카메라 앵글과 피사체를 완전히 누락하고 위치 참조 사진의 구도를 단순 복제하여 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "운전자의 시선은 정면을 향함. 차량은 카메라 쪽으로 다가오며 진행 중임.",
      "built_space": "좌측 운전석에 인물이 탑승함. 위치 참조의 정문과 구조물 참조의 건물이 지시대로 배치됨.",
      "entities": "전택수(참조 이미지의 얼굴, 모자, 재킷 일치), 운전대, 차량, 광주지방검찰청 건물과 간판 모두 확인됨.",
      "hard_violations": [],
      "physics": "운전자의 두 손이 운전대를 안정적으로 잡고 있음."
     },
     {
      "label": "B",
      "direction": "차량이 카메라를 등지고 멀어짐. 운전자의 시선 확인 불가.",
      "built_space": "위치 참조 사진의 정문과 배경 건물만 보이며, 구조물 참조 사진의 외관은 무시됨.",
      "entities": "차량 후면만 보이며, 필수 인물(전택수)과 내부 요소가 전혀 없음.",
      "hard_violations": [
       "필수 피사체(전택수) 누락",
       "명시된 전측면 앞유리 카메라 앵글 위반"
      ],
      "physics": "차량이 바닥에 지지되어 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S34sh1__bgfirst_bg.png",
   "bg_asset_id": "fc19036e-3169-44cc-8535-53f033eef6fa",
   "bg_record_key": "S34sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S34sh1::cine": {
  "applied": true,
  "fingerprint": "f7be0ddcc66b63ee1c6a48114ad750aaa8df1b86dc7ac2c1dbe3d21a5a0a84d4",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S34sh1_sel.png",
  "source_sha256": "7e5579a24bd67bad90eb17bde2b741a36722dbe79402dca3e86ee5635f685fd2",
  "file": "S34sh1_cine.png",
  "latency_ms": 11360
 },
 "S35sh1::signage": {
  "fp": "deadc919ceae8d48",
  "inscriptions": [
   {
    "surface_native": "책상 위 삼각 명패",
    "text_native": "검사 장원섭",
    "reason_ko": "검사의 개인 집무실 분위기를 생생하게 재현하기 위해 책상 위에 검사의 직책과 이름이 적힌 명패를 배치합니다."
   }
  ]
 },
 "era_assess::763614a75064d879": {
  "subjects": [
   {
    "subject_native": "대한민국 검사실 내부 (2001년~2017년)",
    "search_terms_native": [
     "검사실 내부",
     "검사 사무실",
     "검찰청 검사실"
    ],
    "language_lock_native": "검색어는 오직 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "한국 검사실 고유의 서류철 종류, 가구 스타일, 검찰 마크 및 내부 레이아웃은 일반 오피스 이미지와 다릅니다."
   }
  ]
 },
 "era_ref::38d93f514d85103b": {
  "subject": "대한민국 검사실 내부 (2001년~2017년)",
  "terms": [
   "검사실 내부",
   "검사 사무실",
   "검찰청 검사실"
  ],
  "queries": [
   [
    "대한민국 검찰청 검사실 내부 2001년 2017년",
    "검사 사무실 검찰청 검사실 내부"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://cdn.welfarehello.com/naver-blog/production/seocho88/2024-07/223505349348/seocho88_223505349348_13.jpg?f=webp&q=80&w=800",
  "picked_reason_ko": "검사실 표지와 유리 칸막이, 업무용 책상·의자·컴퓨터·서류 보관 상태가 한 화면에 명확히 보여 대한민국의 일상적인 검사실 내부를 가장 충실하게 참고할 수 있다.",
  "sha256": "e3a62d21ff7fe0471ee65c0966c1d00dcdc83c9b0968b546e9c553647d6eda24",
  "file": "eraref_38d93f514d85103b.png"
 },
 "S35sh1::bgfirst_bg": {
  "input_fingerprint": "f263ae4d7a43c9b4",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 책상을 사이에 두고 마주 선 채 서로의 오른손을 꽉 맞잡은 장원섭과 전택수의 측면 전경\n\nLOCATION (lock): Beside the desk inside the prosecutor's private office.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the open side of the desk and upper-torso height, a static medium-wide profile composition places 장원섭 on the left and 전택수 on the right, with their clasped right hands centered above the desk edge. Both men incline slightly into the formal greeting while looking at one another rather than the lens; the single emphasized axis is their matched placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground; desk separating both men beneath the handshake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원섭의 책상 (두 사람이 책상을 사이에 두고 마주함) — The open side faces the camera, and the top plane lies beneath their clasped hands; used as Separates the men institutionally while leaving the open side available for the handshake and later seating move; 맞잡은 오른손 (서로 꽉 맞잡고 악수 중); used as Forms the central action and visual bridge between the two profiles.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is treated with subdued neutral-to-cool color, moderate contrast, and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 검사실 내부 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 책상을 사이에 두고 마주 선 채 서로의 오른손을 꽉 맞잡은 장원섭과 전택수의 측면 전경\n\nLOCATION (lock): Beside the desk inside the prosecutor's private office.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the open side of the desk and upper-torso height, a static medium-wide profile composition places 장원섭 on the left and 전택수 on the right, with their clasped right hands centered above the desk edge. Both men incline slightly into the formal greeting while looking at one another rather than the lens; the single emphasized axis is their matched placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground; desk separating both men beneath the handshake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원섭의 책상 (두 사람이 책상을 사이에 두고 마주함) — The open side faces the camera, and the top plane lies beneath their clasped hands; used as Separates the men institutionally while leaving the open side available for the handshake and later seating move; 맞잡은 오른손 (서로 꽉 맞잡고 악수 중); used as Forms the central action and visual bridge between the two profiles.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is treated with subdued neutral-to-cool color, moderate contrast, and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 검사실 내부 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S35sh1__bgfirst_bg.png",
  "asset_id": "fd608d68-5fec-40b7-8aae-4bc018aaaea5",
  "input_asset_ids": [
   "e1a5a04b-d06a-4927-a3e3-45aecca3f70e",
   "f8ad9ceb-c611-4bf9-ad05-eb373d191391"
  ],
  "era_research": {
   "subject": "대한민국 검사실 내부 (2001년~2017년)",
   "queries": [
    [
     "대한민국 검찰청 검사실 내부 2001년 2017년",
     "검사 사무실 검찰청 검사실 내부"
    ]
   ],
   "picked_url": "https://cdn.welfarehello.com/naver-blog/production/seocho88/2024-07/223505349348/seocho88_223505349348_13.jpg?f=webp&q=80&w=800",
   "sha256": "e3a62d21ff7fe0471ee65c0966c1d00dcdc83c9b0968b546e9c553647d6eda24",
   "file": "eraref_38d93f514d85103b.png"
  }
 },
 "S35sh1": {
  "input_fingerprint": "492395d283d4f09a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상을 사이에 두고 마주 선 채 서로의 오른손을 꽉 맞잡은 장원섭과 전택수의 측면 전경\n\nLOCATION (lock): Beside the desk inside the prosecutor's private office. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the open side of the desk and upper-torso height, a static medium-wide profile composition places 장원섭 on the left and 전택수 on the right, with their clasped right hands centered above the desk edge. Both men incline slightly into the formal greeting while looking at one another rather than the lens; the single emphasized axis is their matched placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground; desk separating both men beneath the handshake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원섭의 책상 (두 사람이 책상을 사이에 두고 마주함) — The open side faces the camera, and the top plane lies beneath their clasped hands; used as Separates the men institutionally while leaving the open side available for the handshake and later seating move; 맞잡은 오른손 (서로 꽉 맞잡고 악수 중); used as Forms the central action and visual bridge between the two profiles.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is treated with subdued neutral-to-cool color, moderate contrast, and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 삼각 명패: \"검사 장원섭\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상을 사이에 두고 마주 선 채 서로의 오른손을 꽉 맞잡은 장원섭과 전택수의 측면 전경\n\nLOCATION (lock): Beside the desk inside the prosecutor's private office. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the open side of the desk and upper-torso height, a static medium-wide profile composition places 장원섭 on the left and 전택수 on the right, with their clasped right hands centered above the desk edge. Both men incline slightly into the formal greeting while looking at one another rather than the lens; the single emphasized axis is their matched placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground; desk separating both men beneath the handshake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원섭의 책상 (두 사람이 책상을 사이에 두고 마주함) — The open side faces the camera, and the top plane lies beneath their clasped hands; used as Separates the men institutionally while leaving the open side available for the handshake and later seating move; 맞잡은 오른손 (서로 꽉 맞잡고 악수 중); used as Forms the central action and visual bridge between the two profiles.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is treated with subdued neutral-to-cool color, moderate contrast, and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 삼각 명패: \"검사 장원섭\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 책상을 사이에 두고 마주 선 채 서로의 오른손을 꽉 맞잡은 장원섭과 전택수의 측면 전경\n\nLOCATION (lock): Beside the desk inside the prosecutor's private office. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the open side of the desk and upper-torso height, a static medium-wide profile composition places 장원섭 on the left and 전택수 on the right, with their clasped right hands centered above the desk edge. Both men incline slightly into the formal greeting while looking at one another rather than the lens; the single emphasized axis is their matched placement across the desk.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground; 전택수 in the middle-right of the frame, midground; desk separating both men beneath the handshake in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 원섭의 책상 (두 사람이 책상을 사이에 두고 마주함) — The open side faces the camera, and the top plane lies beneath their clasped hands; used as Separates the men institutionally while leaving the open side available for the handshake and later seating move; 맞잡은 오른손 (서로 꽉 맞잡고 악수 중); used as Forms the central action and visual bridge between the two profiles.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime office ambience is treated with subdued neutral-to-cool color, moderate contrast, and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 삼각 명패: \"검사 장원섭\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S35sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S35sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L34B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 전택수의 머리스타일을 정확히 표현했고, 핵심 액션인 악수하는 두 손의 형태가 자연스러우며 명패 텍스트와 프레이밍을 훌륭히 충족함."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "배경과 명패 텍스트는 훌륭하나, 중심 액션인 맞잡은 두 손이 심하게 뭉개졌고 전택수의 머리스타일이 레퍼런스의 모자에 가려져 보이지 않음."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "장원섭과 전택수가 서로의 얼굴을 향해 시선을 맞추고 있음.",
      "built_space": "배경에 메인 책상과 창문, 우측 캐비닛이 정확히 위치하나, 두 사람은 전경의 테이블을 사이에 두고 서 있음.",
      "entities": "장원섭(좌)은 정장 차림이며, 전택수(우)는 모자 없이 '짧게 정돈된 검은색 머리'를 정확히 보여줌. 명패 텍스트 '검사 장원섭'은 완벽하나 재킷에 'POLICS' 오타가 있음.",
      "hard_violations": [],
      "physics": "두 사람 모두 바닥에 안정적으로 서 있으며, 맞잡은 오른손의 파지법과 손가락 형태가 해부학적으로 매우 자연스러움."
     },
     {
      "label": "A",
      "direction": "장원섭과 전택수가 서로의 얼굴을 향해 시선을 맞추고 있음.",
      "built_space": "사무실의 소파와 배경의 메인 책상이 레퍼런스와 일치하게 배치되었으며, 두 사람은 커피 테이블을 사이에 두고 있음.",
      "entities": "장원섭(좌)은 인물 묘사가 일치하나, 전택수(우)는 레퍼런스의 모자를 그대로 써서 지시문의 '짧은 검은색 머리'가 가려짐. 명패 텍스트 '검사 장원섭'은 정확함.",
      "hard_violations": [],
      "physics": "바닥에 서서 자세를 유지하고 있으나, 두 사람이 맞잡은 오른손이 형체를 알아보기 힘들 정도로 뭉개져 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 전택수의 머리스타일을 정확히 표현했고, 핵심 액션인 악수하는 두 손의 형태가 자연스러우며 명패 텍스트와 프레이밍을 훌륭히 충족함."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "배경과 명패 텍스트는 훌륭하나, 중심 액션인 맞잡은 두 손이 심하게 뭉개졌고 전택수의 머리스타일이 레퍼런스의 모자에 가려져 보이지 않음."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "장원섭과 전택수가 서로의 얼굴을 향해 시선을 맞추고 있음.",
      "built_space": "배경에 메인 책상과 창문, 우측 캐비닛이 정확히 위치하나, 두 사람은 전경의 테이블을 사이에 두고 서 있음.",
      "entities": "장원섭(좌)은 정장 차림이며, 전택수(우)는 모자 없이 '짧게 정돈된 검은색 머리'를 정확히 보여줌. 명패 텍스트 '검사 장원섭'은 완벽하나 재킷에 'POLICS' 오타가 있음.",
      "hard_violations": [],
      "physics": "두 사람 모두 바닥에 안정적으로 서 있으며, 맞잡은 오른손의 파지법과 손가락 형태가 해부학적으로 매우 자연스러움."
     },
     {
      "label": "A",
      "direction": "장원섭과 전택수가 서로의 얼굴을 향해 시선을 맞추고 있음.",
      "built_space": "사무실의 소파와 배경의 메인 책상이 레퍼런스와 일치하게 배치되었으며, 두 사람은 커피 테이블을 사이에 두고 있음.",
      "entities": "장원섭(좌)은 인물 묘사가 일치하나, 전택수(우)는 레퍼런스의 모자를 그대로 써서 지시문의 '짧은 검은색 머리'가 가려짐. 명패 텍스트 '검사 장원섭'은 정확함.",
      "hard_violations": [],
      "physics": "바닥에 서서 자세를 유지하고 있으나, 두 사람이 맞잡은 오른손이 형체를 알아보기 힘들 정도로 뭉개져 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "전택수의 모자가 누락되고 재킷에 오타가 있으며, 배경에 이미 책상과 탁자가 있음에도 전경에 불필요한 책상을 임의로 추가하여 큰 감점 요소가 되었습니다.  ★위반: [gemini-pro] invented objects (원본 공간에 없는 세 번째 책상을 전경에 생성함)"
     },
     {
      "label": "B",
      "score": 1800,
      "verdict_ko": "지시된 두 인물의 의상(모자 포함)과 외모를 충실히 반영했고, 명패의 텍스트 및 악수하는 구도를 자연스럽고 정확하게 구현하여 우수한 결과를 냈습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.8
     },
     "adjusted": {
      "A": 1.25,
      "B": 1.8
     },
     "violations": {
      "A": [
       "[gemini-pro] invented objects (원본 공간에 없는 세 번째 책상을 전경에 생성함)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.2,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1250,
      "verdict_ko": "전택수의 모자가 누락되고 재킷에 오타가 있으며, 배경에 이미 책상과 탁자가 있음에도 전경에 불필요한 책상을 임의로 추가하여 큰 감점 요소가 되었습니다.  ★위반: [gemini-pro] invented objects (원본 공간에 없는 세 번째 책상을 전경에 생성함)"
     },
     {
      "label": "A",
      "score": 1800,
      "verdict_ko": "지시된 두 인물의 의상(모자 포함)과 외모를 충실히 반영했고, 명패의 텍스트 및 악수하는 구도를 자연스럽고 정확하게 구현하여 우수한 결과를 냈습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1806,
     "B": 1258
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "B",
    "direction": "장원섭과 전택수가 서로의 얼굴을 향해 시선을 맞추고 있음.",
    "built_space": "배경에 메인 책상과 창문, 우측 캐비닛이 정확히 위치하나, 두 사람은 전경의 테이블을 사이에 두고 서 있음.",
    "entities": "장원섭(좌)은 정장 차림이며, 전택수(우)는 모자 없이 '짧게 정돈된 검은색 머리'를 정확히 보여줌. 명패 텍스트 '검사 장원섭'은 완벽하나 재킷에 'POLICS' 오타가 있음.",
    "hard_violations": [],
    "physics": "두 사람 모두 바닥에 안정적으로 서 있으며, 맞잡은 오른손의 파지법과 손가락 형태가 해부학적으로 매우 자연스러움."
   },
   {
    "label": "A",
    "direction": "장원섭과 전택수가 서로의 얼굴을 향해 시선을 맞추고 있음.",
    "built_space": "사무실의 소파와 배경의 메인 책상이 레퍼런스와 일치하게 배치되었으며, 두 사람은 커피 테이블을 사이에 두고 있음.",
    "entities": "장원섭(좌)은 인물 묘사가 일치하나, 전택수(우)는 레퍼런스의 모자를 그대로 써서 지시문의 '짧은 검은색 머리'가 가려짐. 명패 텍스트 '검사 장원섭'은 정확함.",
    "hard_violations": [],
    "physics": "바닥에 서서 자세를 유지하고 있으나, 두 사람이 맞잡은 오른손이 형체를 알아보기 힘들 정도로 뭉개져 있음."
   }
  ],
  "totals": {
   "A": 1806,
   "B": 1258
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "지시된 전택수의 머리스타일을 정확히 표현했고, 핵심 액션인 악수하는 두 손의 형태가 자연스러우며 명패 텍스트와 프레이밍을 훌륭히 충족함."
   },
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "배경과 명패 텍스트는 훌륭하나, 중심 액션인 맞잡은 두 손이 심하게 뭉개졌고 전택수의 머리스타일이 레퍼런스의 모자에 가려져 보이지 않음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L34B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "원본 배경(SHOT BACKGROUND)의 건축물과 고정된 요소를 그대로 유지하라는 지시를 어기고, 우측 수납장의 재질(어두운 목재로 변경), 좌측 벽면의 창문 추가, 중앙 책상과 앞쪽 소파의 디자인 등 사무실 배경 전체가 원본과 다르게 새롭게 생성되었습니다.",
     "fix_en": "Replace the current brown cabinets, gray sofas, left-wall window, and room architecture with the exact beige cabinets, green sofas, bare left wall, and specific desk design from the reference background photograph. Maintain the two men, their handshake, their clothing, and the camera framing exactly as they are.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "책상 위 삼각 명패가 없고 앞쪽 낮은 탁자에 직사각형 명패가 놓여 있다",
     "fix_en": "Remove the rectangular nameplate from the foreground coffee table, showing the bare wooden table surface in its place. Add a triangular wooden nameplate on the main desk behind the men's clasped hands. Maintain the two men, their clothing, their pose, and the rest of the room exactly as they are.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "왼쪽 장원섭의 얼굴·나이가 캐릭터 참조의 40대 초반 갸름한 얼굴과 다르다",
     "fix_en": "Refine the left man's face to match the younger, slimmer features of the reference character. Maintain his dark suit, his pose, the right man, and the entire background exactly as they are.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "원본 배경(SHOT BACKGROUND)의 건축물과 고정된 요소를 그대로 유지하라는 지시를 어기고, 우측 수납장의 재질(어두운 목재로 변경), 좌측 벽면의 창문 추가, 중앙 책상과 앞쪽 소파의 디자인 등 사무실 배경 전체가 원본과 다르게 새롭게 생성되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "책상 위 삼각 명패가 없고 앞쪽 낮은 탁자에 직사각형 명패가 놓여 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "참조 빈 배경과 달리 왼쪽 벽·창문·시계·캐비닛·소파·책상 등 건축과 고정 가구가 바뀌었다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 장원섭의 얼굴·나이가 캐릭터 참조의 40대 초반 갸름한 얼굴과 다르다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 장원섭의 얼굴·나이가 캐릭터 참조의 40대 초반 갸름한 얼굴과 다르다",
    "fix_en": "Refine the left man's face to match the younger, slimmer features of the reference character. Maintain his dark suit, his pose, the right man, and the entire background exactly as they are.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the current brown cabinets, gray sofas, left-wall window, and room architecture with the exact beige cabinets, green sofas, bare left wall, and specific desk design from the reference background photograph. Maintain the two men, their handshake, their clothing, and the camera framing exactly as they are.\n- Remove the rectangular nameplate from the foreground coffee table, showing the bare wooden table surface in its place. Add a triangular wooden nameplate on the main desk behind the men's clasped hands. Maintain the two men, their clothing, their pose, and the rest of the room exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "스케치에 따른 인물 배치(왼쪽 장원섭, 오른쪽 전택수), 서로 마주보는 시선, '검사 장원섭' 명패 텍스트까지 프롬프트의 요구사항을 정확하게 구현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 장원섭이 레퍼런스와 다른 인물로 대체됨 / [openrouter:x-ai/grok-4.6] 원본 배경에 없는 좌측 고창 등 고정 구조가 발명·변조됨"
     },
     {
      "label": "B",
      "score": 679,
      "verdict_ko": "인물의 좌우 배치가 지시문과 반대로 뒤바뀌었고, 서로를 마주보지 않고 카메라를 응시하며, 명패 텍스트도 틀려 다수의 핵심 지시사항을 위반했습니다.  ★위반: [gemini-pro] 인물 위치 위반 (왼쪽에 장원섭, 오른쪽에 전택수가 있어야 함) / [gemini-pro] 시선 지시 위반 (서로를 마주보지 않고 카메라를 응시함) / [openrouter:x-ai/grok-4.6] 장원섭·전택수가 스테이징이 지정한 자리의 반대편에 서 있음"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.429
     },
     "adjusted": {
      "A": 1.0,
      "B": 0.679
     },
     "violations": {
      "B": [
       "[gemini-pro] 인물 위치 위반 (왼쪽에 장원섭, 오른쪽에 전택수가 있어야 함)",
       "[gemini-pro] 시선 지시 위반 (서로를 마주보지 않고 카메라를 응시함)",
       "[openrouter:x-ai/grok-4.6] 장원섭·전택수가 스테이징이 지정한 자리의 반대편에 서 있음"
      ],
      "A": [
       "[openrouter:x-ai/grok-4.6] 장원섭이 레퍼런스와 다른 인물로 대체됨",
       "[openrouter:x-ai/grok-4.6] 원본 배경에 없는 좌측 고창 등 고정 구조가 발명·변조됨"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "스케치에 따른 인물 배치(왼쪽 장원섭, 오른쪽 전택수), 서로 마주보는 시선, '검사 장원섭' 명패 텍스트까지 프롬프트의 요구사항을 정확하게 구현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 장원섭이 레퍼런스와 다른 인물로 대체됨 / [openrouter:x-ai/grok-4.6] 원본 배경에 없는 좌측 고창 등 고정 구조가 발명·변조됨"
     },
     {
      "label": "B",
      "score": 679,
      "verdict_ko": "인물의 좌우 배치가 지시문과 반대로 뒤바뀌었고, 서로를 마주보지 않고 카메라를 응시하며, 명패 텍스트도 틀려 다수의 핵심 지시사항을 위반했습니다.  ★위반: [gemini-pro] 인물 위치 위반 (왼쪽에 장원섭, 오른쪽에 전택수가 있어야 함) / [gemini-pro] 시선 지시 위반 (서로를 마주보지 않고 카메라를 응시함) / [openrouter:x-ai/grok-4.6] 장원섭·전택수가 스테이징이 지정한 자리의 반대편에 서 있음"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 인물의 좌우 배치, 마주 보는 시선, 맞잡은 손의 동작 및 명패 텍스트(검사 장원섭)를 완벽히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물의 배치가 뒤바뀌었고, 렌즈를 응시하고 있으며 명패에 잘못된 문자가 출력되어 주요 지시를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 인물이 서로를 보지 않고 카메라 렌즈를 정면으로 응시함.",
      "built_space": "참조된 사무실의 가구 배치가 반영되었으나 전경의 탁자와 명패 구도가 다소 다름.",
      "entities": "왼쪽(전택수)과 오른쪽(장원섭)의 인물이 지시와 반대로 배치됨. 명패 글자가 '수화돝'으로 깨짐.",
      "hard_violations": [
       "지시된 인물의 좌우 위치 뒤바뀜",
       "서로 마주보는 대신 카메라 렌즈 응시",
       "명패 텍스트 지정 문구 위반"
      ],
      "physics": "바닥에 서서 손을 맞잡는 동작 자체는 물리적으로 정상임."
     },
     {
      "label": "B",
      "direction": "두 인물이 측면으로 서서 서로의 얼굴을 정확하게 마주보고 있음.",
      "built_space": "원본 사무실 배경의 소파, 전경의 탁자, 배경의 책상 배치가 정확히 일치함.",
      "entities": "왼쪽에 장원섭, 오른쪽에 전택수가 올바르게 배치됨. 삼각 명패에 '검사 장원섭' 텍스트가 선명함.",
      "hard_violations": [],
      "physics": "안정적으로 서서 올바르게 오른손을 단단히 맞잡고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 인물의 좌우 배치, 마주 보는 시선, 맞잡은 손의 동작 및 명패 텍스트(검사 장원섭)를 완벽히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 배치가 뒤바뀌었고, 렌즈를 응시하고 있으며 명패에 잘못된 문자가 출력되어 주요 지시를 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "두 인물이 서로를 보지 않고 카메라 렌즈를 정면으로 응시함.",
      "built_space": "참조된 사무실의 가구 배치가 반영되었으나 전경의 탁자와 명패 구도가 다소 다름.",
      "entities": "왼쪽(전택수)과 오른쪽(장원섭)의 인물이 지시와 반대로 배치됨. 명패 글자가 '수화돝'으로 깨짐.",
      "hard_violations": [
       "지시된 인물의 좌우 위치 뒤바뀜",
       "서로 마주보는 대신 카메라 렌즈 응시",
       "명패 텍스트 지정 문구 위반"
      ],
      "physics": "바닥에 서서 손을 맞잡는 동작 자체는 물리적으로 정상임."
     },
     {
      "label": "A",
      "direction": "두 인물이 측면으로 서서 서로의 얼굴을 정확하게 마주보고 있음.",
      "built_space": "원본 사무실 배경의 소파, 전경의 탁자, 배경의 책상 배치가 정확히 일치함.",
      "entities": "왼쪽에 장원섭, 오른쪽에 전택수가 올바르게 배치됨. 삼각 명패에 '검사 장원섭' 텍스트가 선명함.",
      "hard_violations": [],
      "physics": "안정적으로 서서 올바르게 오른손을 단단히 맞잡고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1007,
     "B": 682
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S35sh1__bgfirst_bg.png",
   "bg_asset_id": "fd608d68-5fec-40b7-8aae-4bc018aaaea5",
   "bg_record_key": "S35sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S35sh1::cine": {
  "applied": true,
  "fingerprint": "19ac2e265eabc647cf0798e1a20bed301da86d62b3e953ccce2dca7ab9ae690d",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S35sh1_sel.png",
  "source_sha256": "e7282e00b240ceb6a962121177754d19ddf8bf535612eda8ac2a5c6f1618de85",
  "file": "S35sh1_cine.png",
  "latency_ms": 10464
 },
 "S35sh12::signage": {
  "fp": "87e96df92d0b5948",
  "inscriptions": []
 },
 "S35sh12": {
  "input_fingerprint": "f484866c966ceada",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입술을 굳게 다문 채 깊은 생각에 잠긴 표정으로 전택수의 눈을 가만히 들여다보는 장원섭의 클로즈업\n\nLOCATION (lock): In the seated interview area of the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At close-up distance on 전택수's side of the desk, the camera settles slightly below 장원섭's eye line and off his frontal axis after the final dolly-in. 장원섭 occupies the center-right in a three-quarter view, his closed lips and unwavering eyes held in sharp focus as he studies 전택수 just outside the frame; the near desk edge anchors the lower corner without distracting from his absorbed expression.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office desk (positioned between the two men) — Its near corner enters from the lower left and recedes behind 장원섭; used as The near edge provides a restrained lower-frame anchor for the office confrontation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime office ambience keeps facial tones natural and restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecutor's office, desk arrangement, daylight, and the prosecutor's formal clothing. Exclude the handshake and the visiting officer from the close frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입술을 굳게 다문 채 깊은 생각에 잠긴 표정으로 전택수의 눈을 가만히 들여다보는 장원섭의 클로즈업\n\nLOCATION (lock): In the seated interview area of the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At close-up distance on 전택수's side of the desk, the camera settles slightly below 장원섭's eye line and off his frontal axis after the final dolly-in. 장원섭 occupies the center-right in a three-quarter view, his closed lips and unwavering eyes held in sharp focus as he studies 전택수 just outside the frame; the near desk edge anchors the lower corner without distracting from his absorbed expression.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office desk (positioned between the two men) — Its near corner enters from the lower left and recedes behind 장원섭; used as The near edge provides a restrained lower-frame anchor for the office confrontation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime office ambience keeps facial tones natural and restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecutor's office, desk arrangement, daylight, and the prosecutor's formal clothing. Exclude the handshake and the visiting officer from the close frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입술을 굳게 다문 채 깊은 생각에 잠긴 표정으로 전택수의 눈을 가만히 들여다보는 장원섭의 클로즈업\n\nLOCATION (lock): In the seated interview area of the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At close-up distance on 전택수's side of the desk, the camera settles slightly below 장원섭's eye line and off his frontal axis after the final dolly-in. 장원섭 occupies the center-right in a three-quarter view, his closed lips and unwavering eyes held in sharp focus as he studies 전택수 just outside the frame; the near desk edge anchors the lower corner without distracting from his absorbed expression.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office desk (positioned between the two men) — Its near corner enters from the lower left and recedes behind 장원섭; used as The near edge provides a restrained lower-frame anchor for the office confrontation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime office ambience keeps facial tones natural and restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecutor's office, desk arrangement, daylight, and the prosecutor's formal clothing. Exclude the handshake and the visiting officer from the close frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선은 프레임 왼쪽 밖의 대상을 향하고 있음.",
    "built_space": "응접실 테이블에 명패가 있으나 인물이 베이지색 소파가 아닌 검은색 사무용 의자에 앉아 있으며 뒤쪽에 소파가 잘못 배치됨.",
    "entities": "장원섭의 외모 의상 및 명패 텍스트가 레퍼런스와 일치함.",
    "hard_violations": [
     "레퍼런스의 응접실 구조에 없는 검은색 의자가 등장해 인물이 잘못된 가구에 앉아 있음"
    ],
    "physics": "테이블에 기댄 착석 자세와 지지 상태는 안정적임."
   },
   {
    "label": "B",
    "direction": "시선은 프레임 왼쪽 밖의 대상을 향해 고정됨.",
    "built_space": "인물이 베이지색 소파에 올바르게 앉아 있으며 테이블과 배경 가구 배치가 레퍼런스의 응접실 환경과 부합함.",
    "entities": "장원섭의 외모 의상 명패 텍스트 모두 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "소파에 앉아 무게중심을 잡고 있는 자세가 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지시된 클로즈업 앵글과 인물의 진지한 표정을 잘 살렸으며 베이지색 소파와 의상 등 레퍼런스 요소들을 충실히 반영했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "인물의 외모와 의상은 잘 구현되었으나 지정된 응접실 위치에 맞지 않는 사무용 의자가 등장하여 공간 구조를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S35sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 장면 스틸컷과 달리 장원섭의 양복 왼쪽 가슴 주머니에 흰색 행거치프가 추가되어 복장 연속성에 어긋남.",
     "fix_en": "Remove the white pocket square from the left breast pocket of the man's dark suit jacket, leaving the pocket empty and flush with the suit fabric. Preserve the man present and his position, his remaining clothing, the desk, the nameplate, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "장원섭의 입술이 굳게 다물리지 않고 살짝 벌어져 있다.",
     "fix_en": "Close the man's lips firmly together to convey a serious, absorbed expression. Preserve the man present and his position, his eye gaze, his clothing, the desk, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이전 장면 스틸컷과 달리 장원섭의 양복 왼쪽 가슴 주머니에 흰색 행거치프가 추가되어 복장 연속성에 어긋남.",
     "severity": "major"
    },
    {
     "issue_ko": "장원섭의 입술이 굳게 다물리지 않고 살짝 벌어져 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 가슴까지와 사무실 배경이 넓게 들어온 미디엄 클로즈업이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 장면 스틸컷과 달리 장원섭의 양복 왼쪽 가슴 주머니에 흰색 행거치프가 추가되어 복장 연속성에 어긋남.",
    "fix_en": "Remove the white pocket square from the left breast pocket of the man's dark suit jacket, leaving the pocket empty and flush with the suit fabric. Preserve the man present and his position, his remaining clothing, the desk, the nameplate, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "장원섭의 입술이 굳게 다물리지 않고 살짝 벌어져 있다.",
    "fix_en": "Close the man's lips firmly together to convey a serious, absorbed expression. Preserve the man present and his position, his eye gaze, his clothing, the desk, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S35sh1"
  }
 },
 "S35sh12::cine": {
  "applied": true,
  "fingerprint": "e67e998eeeb202fd655de7ce67a4d877db0d165b2593d7a85d04426ebcfa495a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S35sh12_sel.png",
  "source_sha256": "962309677bf6793fe4e79d12104b81d10c0c10d4faa9aab0a0ca1a912b7c2c0b",
  "file": "S35sh12_cine.png",
  "latency_ms": 9614
 },
 "S36sh2::confined_fp_apt": {
  "applies": true,
  "reason_ko": "이 숏은 승합차 뒷좌석이라는 차량 실내 공간을 배경으로 하며 인물이 앉은 위치와 창밖을 바라보는 방향이 중요합니다. 좌석과 창문의 공간 배치가 어긋나면 이야기의 흐름이 깨질 수 있으므로 평면도 레이아웃 보조가 필요합니다.",
  "input_fingerprint": "3458bb91a5dd6356"
 },
 "S36sh2::signage": {
  "fp": "6d7b69e798fcb384",
  "inscriptions": [
   {
    "surface_native": "교도소 정문 현판",
    "text_native": "교도소",
    "reason_ko": "전택수가 승합차를 타고 교도소 경내로 진입하고 있음을 시각적으로 명확히 전달하기 위해 외벽에 교도소 표지판이 필요합니다."
   }
  ]
 },
 "confinedfp::e51f25f0af4f": {
  "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/confinedfp_base_e51f25f0af4f.png",
  "place_text": "Inside the van's rear passenger compartment as it enters the prison grounds, enclosed by close seats, windows, and vehicle fittings.",
  "input_fingerprint": "6aabcd1e239b26c4"
 },
 "S36sh2::confined_fp": {
  "reads": {
   "controls": "Steering wheel at the DRIVER station.",
   "mirrors": "No mirrors are marked on the diagram.",
   "camera": "Positioned outside the van on the left side, pointing rightwards (inwards) through the window next to the REAR LEFT seat.",
   "occupants": "전택수 is in the REAR LEFT seat, facing forward (upward on the diagram)."
  },
  "mismatches": [
   "The text states the location is inside the van's passenger compartment, but the diagram places the camera outside looking in.",
   "The text dictates that the window and prison building are in the background, which is impossible with the diagram's camera positioned outside looking into the van interior."
  ],
  "scene_description_en": "The camera is stationed outside the van on the left side, pointing inwards through the window. It captures a profile view of 전택수, who occupies the rear left seat in the center of the frame, facing towards the left side of the screen. The window frame acts as a foreground element between the camera and the subject. Behind him, the interior of the van fills the background, showing the empty rear center and rear right seats to the right of the frame. The steering wheel at the front driver's station is out of view. No mirrors are visible in this shot.",
  "fixed": true,
  "input_fingerprint": "01ff69f28d3706f1"
 },
 "era_assess::c71f92e2f9913ecb": {
  "subjects": [
   {
    "subject_native": "대한민국 법무부 교도소 호송차량 내부 (2001년~2017년)",
    "search_terms_native": [
     "법무부 호송차 내부",
     "교도소 호송버스 내부",
     "호송차량 내부 철창"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 다른 언어로 번역하거나 혼용하지 마십시오.",
    "reason_ko": "2000년대 한국 교도소 호송차 내부의 시트 구조, 격벽, 창문 철창 및 내부 마감은 서구식 경찰 밴이나 현대식 차량과 확연히 다릅니다."
   }
  ]
 },
 "era_ref::316a4b9a78cd15a9": {
  "subject": "대한민국 법무부 교도소 호송차량 내부 (2001년~2017년)",
  "terms": [
   "법무부 호송차 내부",
   "교도소 호송버스 내부",
   "호송차량 내부 철창"
  ],
  "queries": [
   [
    "대한민국 법무부 호송차 내부 교도소 호송버스 내부 철창 2001년 2017년",
    "법무부 교정본부 수용자 호송차량 내부 철창"
   ],
   [
    "법무부 호송버스 내부 철창 사진",
    "교도소 호송차량 내부 공개 사진",
    "교정본부 호송차 내부 체험",
    "죄수복 입고 호송버스 한 번 타 보자"
   ]
  ],
  "candidates": 3,
  "picked_index": 1,
  "picked_url": "https://www.sotongsinmun.com/data/cheditor/1911/10_copy1.jpg",
  "picked_reason_ko": "1번은 대한민국 교도소 호송차량의 객실 내부를 넓고 선명하게 보여 주어 좌석 배치, 철망 창문, 운전석 분리 격벽과 출입문 등의 구조를 가장 잘 파악할 수 있다.",
  "sha256": "045e4921d472a28702cc23614b130abfe52f7250a602eaa24eafbd0527c4fe7e",
  "file": "eraref_316a4b9a78cd15a9.png"
 },
 "S36sh2": {
  "input_fingerprint": "709b619300733210",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 승합차 뒷좌석에 앉아 반쯤 열린 차창 밖의 높은 교도소 건물을 묵묵히 올려다보는 전택수의 얼굴\n\nLOCATION (lock): Inside the van's rear passenger compartment as it enters the prison grounds, enclosed by close seats, windows, and vehicle fittings. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Tracking alongside the rear passenger window, the camera finishes slightly below 전택수's seated eye line in a close oblique three-quarter view through the half-open opening. His face occupies the lower-left half while his lifted gaze leads diagonally toward the high prison building held beyond the window in the upper-right background.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, foreground, looks toward high prison building; high prison building in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rear passenger window (half-open) — The glazing is lowered halfway, with 전택수 visible at the opening and the prison building beyond; used as The half-open opening frames 전택수 while preserving the visual route of his upward gaze; high prison building (standing beyond the van) — An elevated exterior face is visible beyond the rear passenger window; used as Held beyond 전택수 as the destination of his upward look.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime ambience maintains natural facial detail and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 교도소 정문 현판: \"교도소\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed outside the van on the left side, pointing inwards through the window. It captures a profile view of 전택수, who occupies the rear left seat in the center of the frame, facing towards the left side of the screen. The window frame acts as a foreground element between the camera and the subject. Behind him, the interior of the van fills the background, showing the empty rear center and rear right seats to the right of the frame. The steering wheel at the front driver's station is out of view. No mirrors are visible in this shot.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 승합차 뒷좌석에 앉아 반쯤 열린 차창 밖의 높은 교도소 건물을 묵묵히 올려다보는 전택수의 얼굴\n\nLOCATION (lock): Inside the van's rear passenger compartment as it enters the prison grounds, enclosed by close seats, windows, and vehicle fittings. The shot takes place here.\n",
   "B": "Create ONE 16:9 photorealistic film still for the brief below.\n\nThe FIRST attached image is a top-down FLOOR PLAN of this interior and\nthe SCENE LAYOUT text below is what a careful reader saw in it.\nTogether they are the ONLY authority for physical arrangement: which\nseat/station each person occupies, which station every primary control\nbelongs to, where any mirror/reflective surface sits and what it can\nphysically reflect, where the camera stands and what appears on which\nside of the screen. If any other sentence seems to contradict them, the\nfloor plan wins. The floor plan is a diagram, not scenery — none of its\nlines, arrows or labels may appear in the photograph. WHO the people\nare and what they do comes from the SHOT TEXT and the attached\nCHARACTER/PROP references — never add a person the SHOT TEXT does not\nplace here. No text, no watermarks.\n\nSCENE LAYOUT (what a careful reader saw in the attached floor plan):\nThe camera is stationed outside the van on the left side, pointing inwards through the window. It captures a profile view of 전택수, who occupies the rear left seat in the center of the frame, facing towards the left side of the screen. The window frame acts as a foreground element between the camera and the subject. Behind him, the interior of the van fills the background, showing the empty rear center and rear right seats to the right of the frame. The steering wheel at the front driver's station is out of view. No mirrors are visible in this shot.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 승합차 뒷좌석에 앉아 반쯤 열린 차창 밖의 높은 교도소 건물을 묵묵히 올려다보는 전택수의 얼굴\n\nLOCATION (lock): Inside the van's rear passenger compartment as it enters the prison grounds, enclosed by close seats, windows, and vehicle fittings. The shot takes place here.\n\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S36sh2_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 대한민국 법무부 교도소 호송차량 내부 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_316a4b9a78cd15a9.png"
    },
    {
     "label": "전택수",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S36sh2_confinedfp.png"
    },
    {
     "label": "PERIOD REFERENCE — 대한민국 법무부 교도소 호송차량 내부 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_316a4b9a78cd15a9.png"
    },
    {
     "label": "전택수",
     "path": "<bytes:917702>"
    }
   ]
  },
  "gq": {
   "route": "combined",
   "gap": 0.5,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.667
   },
   "adjusted": {
    "A": 0.5,
    "B": 0.917
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스의 호송차량 좌석 형태(비닐 시트, 철제 프레임) 완전 누락 및 일반 가죽 시트 적용",
     "[gemini-pro] 건물이 우측 상단 배경이 아닌 좌측 배경에 위치",
     "[gemini-pro] 인물의 시선이 배경의 교도소 건물을 향하지 않고 반대 방향을 응시함",
     "[openrouter:x-ai/grok-4.6] 샷·평면도상 뒷좌석(REAR LEFT)이 아닌 뒤에 빈 좌석이 있는 중간열에 앉혀 있다"
    ],
    "B": [
     "[gemini-pro] 레퍼런스의 호송차량 좌석 구조를 무시하고 일반 차량 시트로 묘사함",
     "[gemini-pro] 수직으로 반쯤 내려간 창문 지시를 어기고 슬라이딩 창문 묘사",
     "[gemini-pro] 건물이 우측 상단 배경이 아닌 좌측 전경 유리에 반사되어 프레이밍 지시 위반"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 500,
   "B": 917
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 500,
    "verdict_ko": "레퍼런스의 호송차량 시트 구조를 무시하고 건물의 위치와 시선 방향 지시를 어겼으나, 반쯤 내려간 창문과 실제 배경 묘사는 B보다 프롬프트에 근접함.  ★위반: [gemini-pro] 레퍼런스의 호송차량 좌석 형태(비닐 시트, 철제 프레임) 완전 누락 및 일반 가죽 시트 적용 / [gemini-pro] 건물이 우측 상단 배경이 아닌 좌측 배경에 위치 / [gemini-pro] 인물의 시선이 배경의 교도소 건물을 향하지 않고 반대 방향을 응시함 / [openrouter:x-ai/grok-4.6] 샷·평면도상 뒷좌석(REAR LEFT)이 아닌 뒤에 빈 좌석이 있는 중간열에 앉혀 있다"
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "레퍼런스 시트 구조 무시 및 슬라이딩 창문 묘사로 지시를 어겼으며, 건물이 유리에 반사된 형태로 렌더링되어 프레이밍과 시선 방향을 크게 위반함.  ★위반: [gemini-pro] 레퍼런스의 호송차량 좌석 구조를 무시하고 일반 차량 시트로 묘사함 / [gemini-pro] 수직으로 반쯤 내려간 창문 지시를 어기고 슬라이딩 창문 묘사 / [gemini-pro] 건물이 우측 상단 배경이 아닌 좌측 전경 유리에 반사되어 프레이밍 지시 위반"
   }
  ],
  "refs": [
   {
    "label": "FLOOR PLAN — layout authority, a diagram, never scenery",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S36sh2_confinedfp.png"
   },
   {
    "label": "PERIOD REFERENCE — 대한민국 법무부 교도소 호송차량 내부 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_316a4b9a78cd15a9.png"
   },
   {
    "label": "전택수",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "평면도 지시와 달리 인물이 맨 뒷좌석(Rear Left)이 아닌 중간 열에 앉아 있으며(우측 뒤로 또 다른 측면 창문과 좌석이 보임), 화면 우측에 보여야 할 빈 중앙 및 우측 뒷좌석이 묘사되지 않았습니다.",
     "fix_en": "Convert the space immediately behind the character into the solid rear wall of the van, removing the extra side window and the seat headrest behind him to make his seat the rearmost one. Preserve the character, his clothing, his pose, the foreground window frame, and the exterior reflection of the building.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "차량 내부가 지정된 레퍼런스와 다릅니다. 창문의 굵은 철망 덮개가 누락되었고, 파란색 일체형 플라스틱/비닐 시트 대신 헤드레스트가 분리된 일반 쿠션 시트가 렌더링되었습니다.",
     "fix_en": "Replace the visible vehicle seats with solid blue vinyl seats without separate headrests, and add a heavy black metal wire mesh over the interior of the rightmost window. Preserve the character, his outfit, his pose, the exterior reflections, and the camera angle.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "평면도 지시와 달리 인물이 맨 뒷좌석(Rear Left)이 아닌 중간 열에 앉아 있으며(우측 뒤로 또 다른 측면 창문과 좌석이 보임), 화면 우측에 보여야 할 빈 중앙 및 우측 뒷좌석이 묘사되지 않았습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "차량 내부가 지정된 레퍼런스와 다릅니다. 창문의 굵은 철망 덮개가 누락되었고, 파란색 일체형 플라스틱/비닐 시트 대신 헤드레스트가 분리된 일반 쿠션 시트가 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 배치도의 맨 뒷줄 왼쪽(rear left)이 아니라 그보다 앞 열에 앉아 있고, 오른쪽 뒷창으로 더 뒤쪽 빈 좌석 머리받침이 보인다.",
     "severity": "critical"
    },
    {
     "issue_ko": "차 안이 고증 호송차량의 파란 시트·철망 칸막이가 아니라 일반 승합차의 어두운 가죽 시트로 되어 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "차량 내부가 지정된 레퍼런스와 다릅니다. 창문의 굵은 철망 덮개가 누락되었고, 파란색 일체형 플라스틱/비닐 시트 대신 헤드레스트가 분리된 일반 쿠션 시트가 렌더링되었습니다.",
    "fix_en": "Replace the visible vehicle seats with solid blue vinyl seats without separate headrests, and add a heavy black metal wire mesh over the interior of the rightmost window. Preserve the character, his outfit, his pose, the exterior reflections, and the camera angle.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Convert the space immediately behind the character into the solid rear wall of the van, removing the extra side window and the seat headrest behind him to make his seat the rearmost one. Preserve the character, his clothing, his pose, the foreground window frame, and the exterior reflection of the building.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지시된 차량 외부 카메라 구도를 정확히 따랐으며, 전택수의 외모와 복장을 완벽히 반영하고 창문에 교도소 건물을 표현하여 프롬프트를 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "카메라 위치가 내부로 잘못 설정되었고, 지정된 캐릭터가 아닌 레퍼런스의 인물을 그대로 복사했으며 워터마크까지 유출되어 심각한 오류가 다수 존재합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 왼쪽 상단으로 향하며 차창 밖(반사된 형태)의 건물을 올려다보고 있음.",
      "built_space": "카메라가 차량 외부에 위치하여 창문을 통해 내부를 들여다보는 구도이며, 인물이 지시된 대로 화면 중앙 좌측 좌석에 앉아 왼쪽을 향함.",
      "entities": "얼굴, 경찰 재킷, 모자 등 전택수의 레퍼런스와 일치하는 인물이며, 창문에 교도소 건물로 보이는 높은 구조물이 나타남.",
      "hard_violations": [],
      "physics": "인물이 차량 좌석에 자연스럽게 앉아 자세를 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 아래를 향하고 있으며 외부 건물을 쳐다보지 않음.",
      "built_space": "카메라가 차량 외부가 아닌 내부 통로 쪽에 위치하여 인물을 정면에서 바라보는 구도로 배치됨.",
      "entities": "프롬프트가 요구한 전택수가 아닌, 배경 레퍼런스 이미지에 있던 수의를 입은 다른 인물이 등장함.",
      "hard_violations": [
       "지시된 카메라 위치(차량 외부 좌측)를 무시하고 내부에 카메라를 배치함",
       "지정된 인물(전택수)이 아닌 레퍼런스 이미지의 다른 인물을 생성함",
       "우측 하단에 레퍼런스 이미지의 워터마크가 그대로 유출됨"
      ],
      "physics": "인물이 좌석에 앉아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지시된 차량 외부 카메라 구도를 정확히 따랐으며, 전택수의 외모와 복장을 완벽히 반영하고 창문에 교도소 건물을 표현하여 프롬프트를 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "카메라 위치가 내부로 잘못 설정되었고, 지정된 캐릭터가 아닌 레퍼런스의 인물을 그대로 복사했으며 워터마크까지 유출되어 심각한 오류가 다수 존재합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 왼쪽 상단으로 향하며 차창 밖(반사된 형태)의 건물을 올려다보고 있음.",
      "built_space": "카메라가 차량 외부에 위치하여 창문을 통해 내부를 들여다보는 구도이며, 인물이 지시된 대로 화면 중앙 좌측 좌석에 앉아 왼쪽을 향함.",
      "entities": "얼굴, 경찰 재킷, 모자 등 전택수의 레퍼런스와 일치하는 인물이며, 창문에 교도소 건물로 보이는 높은 구조물이 나타남.",
      "hard_violations": [],
      "physics": "인물이 차량 좌석에 자연스럽게 앉아 자세를 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 아래를 향하고 있으며 외부 건물을 쳐다보지 않음.",
      "built_space": "카메라가 차량 외부가 아닌 내부 통로 쪽에 위치하여 인물을 정면에서 바라보는 구도로 배치됨.",
      "entities": "프롬프트가 요구한 전택수가 아닌, 배경 레퍼런스 이미지에 있던 수의를 입은 다른 인물이 등장함.",
      "hard_violations": [
       "지시된 카메라 위치(차량 외부 좌측)를 무시하고 내부에 카메라를 배치함",
       "지정된 인물(전택수)이 아닌 레퍼런스 이미지의 다른 인물을 생성함",
       "우측 하단에 레퍼런스 이미지의 워터마크가 그대로 유출됨"
      ],
      "physics": "인물이 좌석에 앉아 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지정된 차량 외부 카메라 시점을 정확히 지켰으며, 전택수의 인상착의(의상과 모자)를 레퍼런스에 맞게 완벽하게 구현한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "카메라가 차량 외부가 아닌 내부에 위치하는 심각한 구도 위반을 범했으며, 지정된 인물(전택수) 대신 레퍼런스 이미지의 죄수를 그대로 그려내어 신원을 완전히 상실했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선이 화면 왼쪽을 향하며 반쯤 열린 창밖의 건물을 의식하듯 약간 위를 응시하고 있음.",
      "built_space": "카메라가 승합차 왼쪽 외부에 위치하여 창문을 통해 내부를 촬영함. 전경에 창틀이 있고, 유리에 건물이 반사되어 보이며, 차량 내부의 좌석 배치가 평면도와 호응함.",
      "entities": "전택수의 얼굴, 연령대, 경찰 재킷, 체크 셔츠, '수사' 모자가 캐릭터 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "좌석에 안정적으로 앉아 차량의 지지를 받고 있음."
     },
     {
      "label": "A",
      "direction": "고개를 숙이고 시선이 바닥이나 무릎을 향해 있어 창밖의 건물을 올려다보지 않음.",
      "built_space": "카메라가 지정된 외부가 아닌 승합차 내부 통로에 위치하여 뒷좌석을 정면으로 바라보고 있음.",
      "entities": "캐릭터 레퍼런스(전택수)의 외모와 의상이 전혀 반영되지 않았으며, 참고용 배경 레퍼런스의 죄수 인물이 잘못 등장함.",
      "hard_violations": [
       "지정된 인물(전택수) 신원 상실 및 완전히 다른 인물 등장",
       "명시된 카메라 시점(차량 외부 좌측) 및 샷 구도 완전 위반"
      ],
      "physics": "좌석에 앉아 중력을 정상적으로 받고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지정된 차량 외부 카메라 시점을 정확히 지켰으며, 전택수의 인상착의(의상과 모자)를 레퍼런스에 맞게 완벽하게 구현한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "카메라가 차량 외부가 아닌 내부에 위치하는 심각한 구도 위반을 범했으며, 지정된 인물(전택수) 대신 레퍼런스 이미지의 죄수를 그대로 그려내어 신원을 완전히 상실했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 화면 왼쪽을 향하며 반쯤 열린 창밖의 건물을 의식하듯 약간 위를 응시하고 있음.",
      "built_space": "카메라가 승합차 왼쪽 외부에 위치하여 창문을 통해 내부를 촬영함. 전경에 창틀이 있고, 유리에 건물이 반사되어 보이며, 차량 내부의 좌석 배치가 평면도와 호응함.",
      "entities": "전택수의 얼굴, 연령대, 경찰 재킷, 체크 셔츠, '수사' 모자가 캐릭터 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "좌석에 안정적으로 앉아 차량의 지지를 받고 있음."
     },
     {
      "label": "B",
      "direction": "고개를 숙이고 시선이 바닥이나 무릎을 향해 있어 창밖의 건물을 올려다보지 않음.",
      "built_space": "카메라가 지정된 외부가 아닌 승합차 내부 통로에 위치하여 뒷좌석을 정면으로 바라보고 있음.",
      "entities": "캐릭터 레퍼런스(전택수)의 외모와 의상이 전혀 반영되지 않았으며, 참고용 배경 레퍼런스의 죄수 인물이 잘못 등장함.",
      "hard_violations": [
       "지정된 인물(전택수) 신원 상실 및 완전히 다른 인물 등장",
       "명시된 카메라 시점(차량 외부 좌측) 및 샷 구도 완전 위반"
      ],
      "physics": "좌석에 앉아 중력을 정상적으로 받고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 18,
     "B": 1
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "confined_fp": {
   "base_key": "confinedfp::e51f25f0af4f",
   "apt_reason": "이 숏은 승합차 뒷좌석이라는 차량 실내 공간을 배경으로 하며 인물이 앉은 위치와 창밖을 바라보는 방향이 중요합니다. 좌석과 창문의 공간 배치가 어긋나면 이야기의 흐름이 깨질 수 있으므로 평면도 레이아웃 보조가 필요합니다.",
   "fixed": true,
   "mismatches": [
    "The text states the location is inside the van's passenger compartment, but the diagram places the camera outside looking in.",
    "The text dictates that the window and prison building are in the background, which is impossible with the diagram's camera positioned outside looking into the van interior."
   ],
   "era_research": {
    "subject": "대한민국 법무부 교도소 호송차량 내부 (2001년~2017년)",
    "queries": [
     [
      "대한민국 법무부 호송차 내부 교도소 호송버스 내부 철창 2001년 2017년",
      "법무부 교정본부 수용자 호송차량 내부 철창"
     ],
     [
      "법무부 호송버스 내부 철창 사진",
      "교도소 호송차량 내부 공개 사진",
      "교정본부 호송차 내부 체험",
      "죄수복 입고 호송버스 한 번 타 보자"
     ]
    ],
    "picked_url": "https://www.sotongsinmun.com/data/cheditor/1911/10_copy1.jpg",
    "sha256": "045e4921d472a28702cc23614b130abfe52f7250a602eaa24eafbd0527c4fe7e",
    "file": "eraref_316a4b9a78cd15a9.png"
   }
  },
  "ref_mode": "confined_fp: 도면+장면설명+엔티티",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S36sh2::cine": {
  "applied": true,
  "fingerprint": "b3c0710c36bae762674c09bac200a9b078b86b16c33253570719ccec0bfc18bb",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S36sh2_sel.png",
  "source_sha256": "a0a0f7ad083e50436686bf116e4e77df9ec458d7fe3616182f6407d4e1fdf064",
  "file": "S36sh2_cine.png",
  "latency_ms": 12210
 },
 "S37sh2::signage": {
  "fp": "71187500a1bb4c2d",
  "inscriptions": [
   {
    "surface_native": "접견실 벽면의 안내판",
    "text_native": "접견실",
    "reason_ko": "교도소 접견실 내부라는 공간적 배경을 사실적으로 나타내기 위해 벽면에 접견실 표시판이 필요합니다."
   }
  ]
 },
 "era_assess::2e3cd9ee8c81650e": {
  "subjects": [
   {
    "subject_native": "2000년대 대한민국 교도소 조사실",
    "search_terms_native": [
     "교도소 조사실",
     "교정시설 조사실",
     "교도소 접견실 내부",
     "교도소 심문실"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "일반적인 서구식 취조실(일방향 거울 등)과 달리, 2000년대 한국 교도소 내 조사실의 독특한 책상 배치, 철제 집기류, 녹음 장비 및 내부 창틀 구조를 정확하게 표현하기 위함"
   }
  ]
 },
 "era_fail::e170e9acef772e5b": {
  "stage": "research",
  "subject": "2000년대 대한민국 교도소 조사실",
  "attempts": 1
 },
 "S37sh2::bgfirst_bg": {
  "input_fingerprint": "0f7d38738904caac",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블을 사이에 두고 서의용, 나상혁의 맞은편에 앉은 지국현과, 그 뒤편에 떨어져 앉은 전택수 및 광주교도소 교도관의 모습\n\nLOCATION (lock): Inside the prison interview room, around the interrogation table with observer seats set back behind it.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the room-side corner behind 서의용 and 나상혁, a slightly elevated wide frame begins the dolly toward the table on a diagonal axis. Their asynchronous over-shoulder silhouettes occupy opposite foreground edges, 지국현 sits centered across the table, and 전택수 with the 광주교도소 교도관 remain separated in the rear depth, making the institutional alignment immediately legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground; 전택수 in the upper-left of the frame, background; 광주교도소 교도관 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between the opposing sides) — The long edge recedes diagonally from the lower foreground toward 지국현 in the middle of the frame; used as Its diagonal length separates the interrogators in the foreground from 지국현 across the room; rear seating (occupied by 전택수 and the 교도관) — The chair fronts angle toward the interrogation table behind 지국현; used as The rear chairs preserve the physical separation of 전택수 and the 교도관 from the interrogation table.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves a naturalistic procedural tone with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 교도소 조사실: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블을 사이에 두고 서의용, 나상혁의 맞은편에 앉은 지국현과, 그 뒤편에 떨어져 앉은 전택수 및 광주교도소 교도관의 모습\n\nLOCATION (lock): Inside the prison interview room, around the interrogation table with observer seats set back behind it.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the room-side corner behind 서의용 and 나상혁, a slightly elevated wide frame begins the dolly toward the table on a diagonal axis. Their asynchronous over-shoulder silhouettes occupy opposite foreground edges, 지국현 sits centered across the table, and 전택수 with the 광주교도소 교도관 remain separated in the rear depth, making the institutional alignment immediately legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground; 전택수 in the upper-left of the frame, background; 광주교도소 교도관 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between the opposing sides) — The long edge recedes diagonally from the lower foreground toward 지국현 in the middle of the frame; used as Its diagonal length separates the interrogators in the foreground from 지국현 across the room; rear seating (occupied by 전택수 and the 교도관) — The chair fronts angle toward the interrogation table behind 지국현; used as The rear chairs preserve the physical separation of 전택수 and the 교도관 from the interrogation table.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves a naturalistic procedural tone with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 교도소 조사실: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S37sh2__bgfirst_bg.png",
  "asset_id": "8202069a-1153-4424-8c80-067cd858bc9f",
  "input_asset_ids": [
   "9cea5a7f-6955-4e3d-9d5e-9e5d7b6caf66",
   "d845ff3f-7f28-4a29-8a58-6ec486ab3139"
  ],
  "era_research": {
   "subject": "2000년대 대한민국 교도소 조사실",
   "queries": [
    [
     "2000년대 대한민국 교도소 조사실 교정시설 조사실 내부",
     "2000년대 대한민국 교도소 접견실 심문실 내부"
    ],
    [
     "대한민국 교도소 접견실 내부 사진",
     "대한민국 교정시설 조사실 내부 사진"
    ]
   ],
   "picked_url": "https://img3.daumcdn.net/thumb/R658x0.q70/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fnews%2F201910%2F31%2Fmoneytoday%2F20191031060122221xane.jpg",
   "sha256": "9f57044a0f067bfddd4d4b7ee4217f7f7c74efec6f5dc3fb097a5f01aa23a659",
   "file": "eraref_e170e9acef772e5b.png"
  }
 },
 "S37sh2": {
  "input_fingerprint": "47053793e1cc8e6f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 서의용, 나상혁의 맞은편에 앉은 지국현과, 그 뒤편에 떨어져 앉은 전택수 및 광주교도소 교도관의 모습\n\nLOCATION (lock): Inside the prison interview room, around the interrogation table with observer seats set back behind it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the room-side corner behind 서의용 and 나상혁, a slightly elevated wide frame begins the dolly toward the table on a diagonal axis. Their asynchronous over-shoulder silhouettes occupy opposite foreground edges, 지국현 sits centered across the table, and 전택수 with the 광주교도소 교도관 remain separated in the rear depth, making the institutional alignment immediately legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground; 전택수 in the upper-left of the frame, background; 광주교도소 교도관 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between the opposing sides) — The long edge recedes diagonally from the lower foreground toward 지국현 in the middle of the frame; used as Its diagonal length separates the interrogators in the foreground from 지국현 across the room; rear seating (occupied by 전택수 and the 교도관) — The chair fronts angle toward the interrogation table behind 지국현; used as The rear chairs preserve the physical separation of 전택수 and the 교도관 from the interrogation table.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves a naturalistic procedural tone with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리); 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 접견실 벽면의 안내판: \"접견실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 서의용, 나상혁의 맞은편에 앉은 지국현과, 그 뒤편에 떨어져 앉은 전택수 및 광주교도소 교도관의 모습\n\nLOCATION (lock): Inside the prison interview room, around the interrogation table with observer seats set back behind it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the room-side corner behind 서의용 and 나상혁, a slightly elevated wide frame begins the dolly toward the table on a diagonal axis. Their asynchronous over-shoulder silhouettes occupy opposite foreground edges, 지국현 sits centered across the table, and 전택수 with the 광주교도소 교도관 remain separated in the rear depth, making the institutional alignment immediately legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground; 전택수 in the upper-left of the frame, background; 광주교도소 교도관 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between the opposing sides) — The long edge recedes diagonally from the lower foreground toward 지국현 in the middle of the frame; used as Its diagonal length separates the interrogators in the foreground from 지국현 across the room; rear seating (occupied by 전택수 and the 교도관) — The chair fronts angle toward the interrogation table behind 지국현; used as The rear chairs preserve the physical separation of 전택수 and the 교도관 from the interrogation table.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves a naturalistic procedural tone with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리); 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 접견실 벽면의 안내판: \"접견실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 서의용, 나상혁의 맞은편에 앉은 지국현과, 그 뒤편에 떨어져 앉은 전택수 및 광주교도소 교도관의 모습\n\nLOCATION (lock): Inside the prison interview room, around the interrogation table with observer seats set back behind it. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the room-side corner behind 서의용 and 나상혁, a slightly elevated wide frame begins the dolly toward the table on a diagonal axis. Their asynchronous over-shoulder silhouettes occupy opposite foreground edges, 지국현 sits centered across the table, and 전택수 with the 광주교도소 교도관 remain separated in the rear depth, making the institutional alignment immediately legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground; 전택수 in the upper-left of the frame, background; 광주교도소 교도관 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between the opposing sides) — The long edge recedes diagonally from the lower foreground toward 지국현 in the middle of the frame; used as Its diagonal length separates the interrogators in the foreground from 지국현 across the room; rear seating (occupied by 전택수 and the 교도관) — The chair fronts angle toward the interrogation table behind 지국현; used as The rear chairs preserve the physical separation of 전택수 and the 교도관 from the interrogation table.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves a naturalistic procedural tone with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리); 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 접견실 벽면의 안내판: \"접견실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S37sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S37sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    },
    {
     "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:994284>"
    },
    {
     "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:861801>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L36B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    },
    {
     "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:994284>"
    },
    {
     "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:861801>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "인물들의 지정된 좌석 배치와 안내판 텍스트를 잘 준수했으나, 대각선 구도 누락과 전택수의 의상 차이가 아쉽습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "핵심 인물인 지국현과 교도관의 착석 위치가 프롬프트 지시와 완전히 어긋나 심각한 배치 위반에 해당합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전경 좌측과 중앙 인물이 마주보며, 배경 인물들도 중앙을 응시함.",
      "built_space": "대각선 테이블 주위로 5개의 의자에 인물들이 착석함.",
      "entities": "지국현(수의)이 배경 우측에, 교도관(제복)이 전경 우측에 잘못 배치됨. 중앙에는 정장을 입은 인물이 있음. 안내판 '접견실' 텍스트는 일치함.",
      "hard_violations": [
       "지국현과 교도관이 프롬프트가 지정하지 않은 잘못된 좌석에 배치됨"
      ],
      "physics": "모든 인물이 의자에 의해 체중을 지탱하며, 손과 팔이 테이블 표면에 놓여 있음."
     },
     {
      "label": "B",
      "direction": "지국현은 전경 좌측 인물을 바라보고, 배경 인물들은 정면을 주시함.",
      "built_space": "가로로 놓인 테이블을 기준으로 전경 2명, 중/배경 3명이 의자에 착석함.",
      "entities": "지국현(중앙)과 교도관(배경 우측)의 위치와 의상이 정확히 일치함. 전택수(배경 좌측)는 점퍼 대신 정장을 입어 의상이 불일치함. '접견실' 텍스트는 정확함.",
      "hard_violations": [],
      "physics": "모든 인물이 의자에 앉아 안정적으로 지탱되며, 지국현의 두 손은 테이블 위에 얹혀 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "인물들의 지정된 좌석 배치와 안내판 텍스트를 잘 준수했으나, 대각선 구도 누락과 전택수의 의상 차이가 아쉽습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "핵심 인물인 지국현과 교도관의 착석 위치가 프롬프트 지시와 완전히 어긋나 심각한 배치 위반에 해당합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전경 좌측과 중앙 인물이 마주보며, 배경 인물들도 중앙을 응시함.",
      "built_space": "대각선 테이블 주위로 5개의 의자에 인물들이 착석함.",
      "entities": "지국현(수의)이 배경 우측에, 교도관(제복)이 전경 우측에 잘못 배치됨. 중앙에는 정장을 입은 인물이 있음. 안내판 '접견실' 텍스트는 일치함.",
      "hard_violations": [
       "지국현과 교도관이 프롬프트가 지정하지 않은 잘못된 좌석에 배치됨"
      ],
      "physics": "모든 인물이 의자에 의해 체중을 지탱하며, 손과 팔이 테이블 표면에 놓여 있음."
     },
     {
      "label": "B",
      "direction": "지국현은 전경 좌측 인물을 바라보고, 배경 인물들은 정면을 주시함.",
      "built_space": "가로로 놓인 테이블을 기준으로 전경 2명, 중/배경 3명이 의자에 착석함.",
      "entities": "지국현(중앙)과 교도관(배경 우측)의 위치와 의상이 정확히 일치함. 전택수(배경 좌측)는 점퍼 대신 정장을 입어 의상이 불일치함. '접견실' 텍스트는 정확함.",
      "hard_violations": [],
      "physics": "모든 인물이 의자에 앉아 안정적으로 지탱되며, 지국현의 두 손은 테이블 위에 얹혀 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트가 지시한 프레임 레이아웃(중앙 지국현, 후경 전택수와 교도관, 전경 실루엣 2명)과 장소 디테일, 안내판 텍스트를 정확히 구현함."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "중앙에 나상혁을 배치하고 지국현과 교도관의 위치를 잘못 지정하는 등 명시된 인물 배치 지시를 완전히 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "프레임 전경의 두 인물과 후경의 인물들이 중앙에 앉은 지국현을 향해 시선을 두거나 테이블 쪽을 향함.",
      "built_space": "참고 사진의 접견실 구조(벽면 투톤 컬러, 철창창문)가 잘 재현되었으며, 중앙 테이블 앞뒤로 의자가 적절히 배치되어 5명이 지정된 자리에 앉아 있음.",
      "entities": "중앙에 죄수복('교도스 1234')을 입은 지국현, 좌측 후경에 전택수, 우측 후경에 푸른 제복의 교도관, 전경 양쪽에 두 인물이 정확한 위치에 등장하며 '접견실' 안내판 텍스트도 완벽함.",
      "hard_violations": [],
      "physics": "모든 인물이 의자에 안정적으로 착석해 있으며, 중앙 인물의 두 팔은 테이블 위에 자연스럽게 얹혀 있어 물리적 오류가 없음."
     },
     {
      "label": "B",
      "direction": "전경 및 후경 인물들이 중앙의 양복 입은 인물을 향해 시선을 모으고 있음.",
      "built_space": "테이블과 의자 배치는 참고 사진과 유사하나, 인물들이 차지하고 있는 위치가 프롬프트에서 요구한 구조와 맞지 않음.",
      "entities": "중앙에 나상혁이 위치하고, 죄수복을 입은 지국현이 우측 후경으로 밀려났으며, 교도관이 우측 전경에 앉아 있어 프롬프트가 지시한 인물 배치가 완전히 실패함.",
      "hard_violations": [
       "프롬프트가 지정한 좌석 배치 위반 (지국현이 중앙이 아닌 후경에, 교도관이 후경이 아닌 전경에 앉아 있음)"
      ],
      "physics": "모든 인물은 의자에 정상적으로 앉아 있고 손과 팔이 다리나 테이블에 지지되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "프롬프트가 지시한 프레임 레이아웃(중앙 지국현, 후경 전택수와 교도관, 전경 실루엣 2명)과 장소 디테일, 안내판 텍스트를 정확히 구현함."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "중앙에 나상혁을 배치하고 지국현과 교도관의 위치를 잘못 지정하는 등 명시된 인물 배치 지시를 완전히 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "프레임 전경의 두 인물과 후경의 인물들이 중앙에 앉은 지국현을 향해 시선을 두거나 테이블 쪽을 향함.",
      "built_space": "참고 사진의 접견실 구조(벽면 투톤 컬러, 철창창문)가 잘 재현되었으며, 중앙 테이블 앞뒤로 의자가 적절히 배치되어 5명이 지정된 자리에 앉아 있음.",
      "entities": "중앙에 죄수복('교도스 1234')을 입은 지국현, 좌측 후경에 전택수, 우측 후경에 푸른 제복의 교도관, 전경 양쪽에 두 인물이 정확한 위치에 등장하며 '접견실' 안내판 텍스트도 완벽함.",
      "hard_violations": [],
      "physics": "모든 인물이 의자에 안정적으로 착석해 있으며, 중앙 인물의 두 팔은 테이블 위에 자연스럽게 얹혀 있어 물리적 오류가 없음."
     },
     {
      "label": "A",
      "direction": "전경 및 후경 인물들이 중앙의 양복 입은 인물을 향해 시선을 모으고 있음.",
      "built_space": "테이블과 의자 배치는 참고 사진과 유사하나, 인물들이 차지하고 있는 위치가 프롬프트에서 요구한 구조와 맞지 않음.",
      "entities": "중앙에 나상혁이 위치하고, 죄수복을 입은 지국현이 우측 후경으로 밀려났으며, 교도관이 우측 전경에 앉아 있어 프롬프트가 지시한 인물 배치가 완전히 실패함.",
      "hard_violations": [
       "프롬프트가 지정한 좌석 배치 위반 (지국현이 중앙이 아닌 후경에, 교도관이 후경이 아닌 전경에 앉아 있음)"
      ],
      "physics": "모든 인물은 의자에 정상적으로 앉아 있고 손과 팔이 다리나 테이블에 지지되어 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 5,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "전경 좌측과 중앙 인물이 마주보며, 배경 인물들도 중앙을 응시함.",
    "built_space": "대각선 테이블 주위로 5개의 의자에 인물들이 착석함.",
    "entities": "지국현(수의)이 배경 우측에, 교도관(제복)이 전경 우측에 잘못 배치됨. 중앙에는 정장을 입은 인물이 있음. 안내판 '접견실' 텍스트는 일치함.",
    "hard_violations": [
     "지국현과 교도관이 프롬프트가 지정하지 않은 잘못된 좌석에 배치됨"
    ],
    "physics": "모든 인물이 의자에 의해 체중을 지탱하며, 손과 팔이 테이블 표면에 놓여 있음."
   },
   {
    "label": "B",
    "direction": "지국현은 전경 좌측 인물을 바라보고, 배경 인물들은 정면을 주시함.",
    "built_space": "가로로 놓인 테이블을 기준으로 전경 2명, 중/배경 3명이 의자에 착석함.",
    "entities": "지국현(중앙)과 교도관(배경 우측)의 위치와 의상이 정확히 일치함. 전택수(배경 좌측)는 점퍼 대신 정장을 입어 의상이 불일치함. '접견실' 텍스트는 정확함.",
    "hard_violations": [],
    "physics": "모든 인물이 의자에 앉아 안정적으로 지탱되며, 지국현의 두 손은 테이블 위에 얹혀 있음."
   }
  ],
  "totals": {
   "A": 5,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "인물들의 지정된 좌석 배치와 안내판 텍스트를 잘 준수했으나, 대각선 구도 누락과 전택수의 의상 차이가 아쉽습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "핵심 인물인 지국현과 교도관의 착석 위치가 프롬프트 지시와 완전히 어긋나 심각한 배치 위반에 해당합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L36B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   },
   {
    "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:861801>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "좌측 배경에 배치되어야 할 인물(전택수)이 제공된 캐릭터 레퍼런스(경찰 점퍼, 모자 착용)와 완전히 다른 외모와 정장 차림으로 렌더링되었습니다.",
     "fix_en": "Change the man in the left background to match reference 전택수 exactly: face, navy POLICE jacket, plaid shirt, and navy cap. Preserve the remaining people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "테이블이 전경 하단에서 중심 인물 쪽으로 대각선 방향으로 깊어지게 배치되어야 한다는 카메라 구도 지시를 무시하고 화면에 수평으로 놓여 있습니다.",
     "fix_en": "Rotate the table so its long edge recedes diagonally from the foreground toward the center.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경에 앉은 두 인물(좌측 정장 입은 남성, 우측 교도관)이 자연스러운 상황에 있지 않고 카메라 렌즈 쪽을 정면으로 응시하며 경직된 자세를 취하고 있습니다.",
     "fix_en": "Redirect the eyes of the two background characters to look at the center character or the table.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "전경 오버숄더에 서의용의 비니·가죽자켓이 없고 오른쪽은 레퍼런스에 없는 회색 정장이다.",
     "fix_en": "Change the clothing of the left foreground character to a leather jacket and beanie, and the right to a navy bomber jacket.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "장소 레퍼런스의 뒷벽 중앙 철창 창문이 왼쪽 벽으로 옮겨져 있다.",
     "fix_en": "Move the barred window from the left wall to the center of the rear wall.",
     "severity": "major",
     "observation_index": 7,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "좌측 배경에 배치되어야 할 인물(전택수)이 제공된 캐릭터 레퍼런스(경찰 점퍼, 모자 착용)와 완전히 다른 외모와 정장 차림으로 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "테이블이 전경 하단에서 중심 인물 쪽으로 대각선 방향으로 깊어지게 배치되어야 한다는 카메라 구도 지시를 무시하고 화면에 수평으로 놓여 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경에 앉은 두 인물(좌측 정장 입은 남성, 우측 교도관)이 자연스러운 상황에 있지 않고 카메라 렌즈 쪽을 정면으로 응시하며 경직된 자세를 취하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 전경에 등진 인물의 실루엣이 지정된 캐릭터 레퍼런스(가죽 재킷 또는 항공 점퍼)에 없는 정장 상의를 착용하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 배경 남성이 전택수 레퍼런스와 다른 얼굴이며 수사 점퍼·모자가 아닌 회색 정장이다.",
     "severity": "major"
    },
    {
     "issue_ko": "전경 오버숄더에 서의용의 비니·가죽자켓이 없고 오른쪽은 레퍼런스에 없는 회색 정장이다.",
     "severity": "major"
    },
    {
     "issue_ko": "조사용 테이블 긴 변이 하단에서 지국현 쪽으로 대각선 후퇴하지 않고 화면과 거의 평행하다.",
     "severity": "major"
    },
    {
     "issue_ko": "장소 레퍼런스의 뒷벽 중앙 철창 창문이 왼쪽 벽으로 옮겨져 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "테이블이 전경 하단에서 중심 인물 쪽으로 대각선 방향으로 깊어지게 배치되어야 한다는 카메라 구도 지시를 무시하고 화면에 수평으로 놓여 있습니다.",
    "fix_en": "Rotate the table so its long edge recedes diagonally from the foreground toward the center.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경에 앉은 두 인물(좌측 정장 입은 남성, 우측 교도관)이 자연스러운 상황에 있지 않고 카메라 렌즈 쪽을 정면으로 응시하며 경직된 자세를 취하고 있습니다.",
    "fix_en": "Redirect the eyes of the two background characters to look at the center character or the table.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "전경 오버숄더에 서의용의 비니·가죽자켓이 없고 오른쪽은 레퍼런스에 없는 회색 정장이다.",
    "fix_en": "Change the clothing of the left foreground character to a leather jacket and beanie, and the right to a navy bomber jacket.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "장소 레퍼런스의 뒷벽 중앙 철창 창문이 왼쪽 벽으로 옮겨져 있다.",
    "fix_en": "Move the barred window from the left wall to the center of the rear wall.",
    "severity": "major",
    "observation_index": 7,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 7,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the man in the left background to match reference 전택수 exactly: face, navy POLICE jacket, plaid shirt, and navy cap. Preserve the remaining people present and their positions, their clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 교도소 접견실 배경과 오버숄더 구도, 인물 배치를 훌륭하게 구현했으나, 배경 좌측의 전택수 및 전경 인물들의 의상이 레퍼런스와 다르게 표현되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 요구한 장소, 테이블 구도, 다수의 인물 배치를 모두 무시하고 스튜디오 배경에 단 두 명만 서 있는 형태로 렌더링되어 지시 사항을 전혀 이행하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "수용복을 입은 지국현은 전경 인물들을 향해 시선을 두고 있으며, 배경의 교도관과 전택수도 테이블 쪽을 바라보고 있음.",
      "built_space": "교도소 접견실 내부를 정확히 묘사함. 대각선으로 놓인 테이블과 의자들, 벽면의 '접견실' 안내판이 프롬프트의 요구사항에 맞게 배치됨.",
      "entities": "중앙의 지국현(수의 착용), 우측 배경의 교도관(파란색 제복)은 레퍼런스와 일치함. 그러나 좌측 배경의 전택수가 경찰 자켓 대신 정장을 입고 있으며, 전경 인물들의 의상도 레퍼런스와 차이가 있음.",
      "hard_violations": [],
      "physics": "등장인물들 모두 의자에 자연스럽게 앉아 있으며, 신체가 가구나 바닥에 안정적으로 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "등장한 두 인물 모두 정면 카메라를 응시하고 있음.",
      "built_space": "아무런 구조물이나 가구가 없는 단순한 흰색 스튜디오 배경임.",
      "entities": "경찰 자켓을 입은 인물과 남색 자켓을 입은 인물 두 명만 등장하며, 지국현, 교도관 및 지시된 5명의 전체 인원 구성이 완전히 누락됨.",
      "hard_violations": [
       "지정된 장소(접견실)가 아닌 스튜디오 배경에서 촬영됨",
       "지정된 구도(테이블을 둔 착석 상태 및 오버숄더)를 완전히 무시함",
       "필수 등장 인물 대거 누락됨"
      ],
      "physics": "두 인물 모두 평면 바닥에 서서 체중을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 교도소 접견실 배경과 오버숄더 구도, 인물 배치를 훌륭하게 구현했으나, 배경 좌측의 전택수 및 전경 인물들의 의상이 레퍼런스와 다르게 표현되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 요구한 장소, 테이블 구도, 다수의 인물 배치를 모두 무시하고 스튜디오 배경에 단 두 명만 서 있는 형태로 렌더링되어 지시 사항을 전혀 이행하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "수용복을 입은 지국현은 전경 인물들을 향해 시선을 두고 있으며, 배경의 교도관과 전택수도 테이블 쪽을 바라보고 있음.",
      "built_space": "교도소 접견실 내부를 정확히 묘사함. 대각선으로 놓인 테이블과 의자들, 벽면의 '접견실' 안내판이 프롬프트의 요구사항에 맞게 배치됨.",
      "entities": "중앙의 지국현(수의 착용), 우측 배경의 교도관(파란색 제복)은 레퍼런스와 일치함. 그러나 좌측 배경의 전택수가 경찰 자켓 대신 정장을 입고 있으며, 전경 인물들의 의상도 레퍼런스와 차이가 있음.",
      "hard_violations": [],
      "physics": "등장인물들 모두 의자에 자연스럽게 앉아 있으며, 신체가 가구나 바닥에 안정적으로 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "등장한 두 인물 모두 정면 카메라를 응시하고 있음.",
      "built_space": "아무런 구조물이나 가구가 없는 단순한 흰색 스튜디오 배경임.",
      "entities": "경찰 자켓을 입은 인물과 남색 자켓을 입은 인물 두 명만 등장하며, 지국현, 교도관 및 지시된 5명의 전체 인원 구성이 완전히 누락됨.",
      "hard_violations": [
       "지정된 장소(접견실)가 아닌 스튜디오 배경에서 촬영됨",
       "지정된 구도(테이블을 둔 착석 상태 및 오버숄더)를 완전히 무시함",
       "필수 등장 인물 대거 누락됨"
      ],
      "physics": "두 인물 모두 평면 바닥에 서서 체중을 지탱하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 카메라 구도, 공간적 배치, 조명 분위기를 훌륭하게 구현했으나, 배경 좌측 전택수가 레퍼런스와 완전히 다른 인물(정장 착용)로 변형된 점이 감점 요인입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트가 지시한 배경, 테이블 구도, 필수 인물들을 완전히 누락하고 스튜디오 컷으로 연출하여 샷 텍스트의 의도를 전혀 충족하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "전경 인물들은 중앙의 지국현을, 지국현과 배경 인물들은 전경 쪽을 향해 자연스럽게 시선을 두고 있음.",
      "built_space": "지시된 투톤 벽면, 창살, 중앙 테이블, 그리고 그 주위에 배치된 의자들의 위치와 공간감이 텍스트와 정확히 일치함.",
      "entities": "중앙의 지국현과 우측 교도관은 레퍼런스와 일치하며 벽면에 '접견실' 간판이 있음. 반면 좌측 전택수는 얼굴이 다르고 정장을 입고 있으며, 전경 인물들의 복장도 레퍼런스와 다름.",
      "hard_violations": [],
      "physics": "모든 인물이 의자에 안정적으로 착석하여 팔과 체중을 자연스럽게 지지하고 있음."
     },
     {
      "label": "A",
      "direction": "두 인물 모두 정면의 카메라를 똑바로 응시하고 있음.",
      "built_space": "지정된 교도소 세트가 아닌, 구조물이 전혀 없는 흰색 스튜디오 배경임.",
      "entities": "지국현, 서의용, 교도관이 아예 누락됨. 우측 인물만 나상혁의 복장과 일치하며, 좌측 인물은 레퍼런스와 다른 모자를 착용함.",
      "hard_violations": [
       "지정된 교도소 공간 완전 누락",
       "지정된 카메라 구도 및 5인 배치 스태이징 완전 무시"
      ],
      "physics": "두 인물 모두 바닥에 평범하게 서서 체중을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 카메라 구도, 공간적 배치, 조명 분위기를 훌륭하게 구현했으나, 배경 좌측 전택수가 레퍼런스와 완전히 다른 인물(정장 착용)로 변형된 점이 감점 요인입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 지시한 배경, 테이블 구도, 필수 인물들을 완전히 누락하고 스튜디오 컷으로 연출하여 샷 텍스트의 의도를 전혀 충족하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전경 인물들은 중앙의 지국현을, 지국현과 배경 인물들은 전경 쪽을 향해 자연스럽게 시선을 두고 있음.",
      "built_space": "지시된 투톤 벽면, 창살, 중앙 테이블, 그리고 그 주위에 배치된 의자들의 위치와 공간감이 텍스트와 정확히 일치함.",
      "entities": "중앙의 지국현과 우측 교도관은 레퍼런스와 일치하며 벽면에 '접견실' 간판이 있음. 반면 좌측 전택수는 얼굴이 다르고 정장을 입고 있으며, 전경 인물들의 복장도 레퍼런스와 다름.",
      "hard_violations": [],
      "physics": "모든 인물이 의자에 안정적으로 착석하여 팔과 체중을 자연스럽게 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "두 인물 모두 정면의 카메라를 똑바로 응시하고 있음.",
      "built_space": "지정된 교도소 세트가 아닌, 구조물이 전혀 없는 흰색 스튜디오 배경임.",
      "entities": "지국현, 서의용, 교도관이 아예 누락됨. 우측 인물만 나상혁의 복장과 일치하며, 좌측 인물은 레퍼런스와 다른 모자를 착용함.",
      "hard_violations": [
       "지정된 교도소 공간 완전 누락",
       "지정된 카메라 구도 및 5인 배치 스태이징 완전 무시"
      ],
      "physics": "두 인물 모두 바닥에 평범하게 서서 체중을 지탱하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S37sh2__bgfirst_bg.png",
   "bg_asset_id": "8202069a-1153-4424-8c80-067cd858bc9f",
   "bg_record_key": "S37sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S37sh2::cine": {
  "applied": true,
  "fingerprint": "93c924dbdf4e6a3d8dbcc61b2dc52393dc9fa522f4e63c244754fbf8e868cd87",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S37sh2_sel.png",
  "source_sha256": "86b25e7e8c70b45d616178525f79f2def862f09220e88aa7707c3eaef685ace6",
  "file": "S37sh2_cine.png",
  "latency_ms": 14430
 },
 "S37sh17::signage": {
  "fp": "3a0aeeae915ac8cc",
  "inscriptions": [
   {
    "surface_native": "담뱃갑 경고 표면",
    "text_native": "흡연은 폐암 등 각종 질병의 원인이 됩니다",
    "reason_ko": "지국현이 코끝에 바짝 대고 냄새를 맡는 담뱃갑에 당시 한국의 사실적인 디테일을 더하기 위해 시대에 맞는 경고 문구를 표시합니다."
   }
  ]
 },
 "S37sh17": {
  "input_fingerprint": "b311f9b3ff42d861",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손에 든 담뱃갑을 코끝에 바짝 댄 채 눈을 지그시 감고 냄새를 들이마시는 지국현의 얼굴\n\nLOCATION (lock): At the interrogation table inside the prison interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height beside the table's long edge, the lateral pan settles into a tight oblique close-up of 지국현 (현재) as the cigarette pack reaches his nose. His closed eyes and controlled face occupy the center-right, while the pack remains naturally scaled below his nose and a narrow strip of table at frame left preserves the interrogation geometry.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Marlboro Red cigarette pack (held in 지국현's hand) — One branded face is obliquely visible to camera beneath 지국현's nose; used as Held close to the face as the immediate focus of the smelling gesture without exceeding a natural hand-held scale; investigation table (beside 지국현) — Its long edge runs away from the camera along the left side of the frame; used as A narrow lower-left strip retains continuity with the table-side camera path.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime interior light holds natural skin texture and avoids glamorizing the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same stark interrogation room, table placement, institutional lighting, and prison uniform. Exclude the investigators and guard from the close frame, retaining only the prisoner smelling the cigarette pack.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon is holding Sanghyeok's Marlboro Red pack at his nose; the lighter remains with the offered cigarettes on the table side until the pack is returned.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 지국현 right now — 지국현's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 지국현 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 담뱃갑 경고 표면: \"흡연은 폐암 등 각종 질병의 원인이 됩니다\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손에 든 담뱃갑을 코끝에 바짝 댄 채 눈을 지그시 감고 냄새를 들이마시는 지국현의 얼굴\n\nLOCATION (lock): At the interrogation table inside the prison interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height beside the table's long edge, the lateral pan settles into a tight oblique close-up of 지국현 (현재) as the cigarette pack reaches his nose. His closed eyes and controlled face occupy the center-right, while the pack remains naturally scaled below his nose and a narrow strip of table at frame left preserves the interrogation geometry.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Marlboro Red cigarette pack (held in 지국현's hand) — One branded face is obliquely visible to camera beneath 지국현's nose; used as Held close to the face as the immediate focus of the smelling gesture without exceeding a natural hand-held scale; investigation table (beside 지국현) — Its long edge runs away from the camera along the left side of the frame; used as A narrow lower-left strip retains continuity with the table-side camera path.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime interior light holds natural skin texture and avoids glamorizing the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same stark interrogation room, table placement, institutional lighting, and prison uniform. Exclude the investigators and guard from the close frame, retaining only the prisoner smelling the cigarette pack.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon is holding Sanghyeok's Marlboro Red pack at his nose; the lighter remains with the offered cigarettes on the table side until the pack is returned.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 지국현 right now — 지국현's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 지국현 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 담뱃갑 경고 표면: \"흡연은 폐암 등 각종 질병의 원인이 됩니다\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손에 든 담뱃갑을 코끝에 바짝 댄 채 눈을 지그시 감고 냄새를 들이마시는 지국현의 얼굴\n\nLOCATION (lock): At the interrogation table inside the prison interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height beside the table's long edge, the lateral pan settles into a tight oblique close-up of 지국현 (현재) as the cigarette pack reaches his nose. His closed eyes and controlled face occupy the center-right, while the pack remains naturally scaled below his nose and a narrow strip of table at frame left preserves the interrogation geometry.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Marlboro Red cigarette pack (held in 지국현's hand) — One branded face is obliquely visible to camera beneath 지국현's nose; used as Held close to the face as the immediate focus of the smelling gesture without exceeding a natural hand-held scale; investigation table (beside 지국현) — Its long edge runs away from the camera along the left side of the frame; used as A narrow lower-left strip retains continuity with the table-side camera path.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime interior light holds natural skin texture and avoids glamorizing the gesture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same stark interrogation room, table placement, institutional lighting, and prison uniform. Exclude the investigators and guard from the close frame, retaining only the prisoner smelling the cigarette pack.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon is holding Sanghyeok's Marlboro Red pack at his nose; the lighter remains with the offered cigarettes on the table side until the pack is returned.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 지국현 right now — 지국현's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 지국현 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 담뱃갑 경고 표면: \"흡연은 폐암 등 각종 질병의 원인이 됩니다\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.125,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.875,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.875,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1875,
   "B": 1750
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1875,
    "verdict_ko": "지정된 담뱃갑 경고 문구를 완벽하게 렌더링하였으며, 클로즈업 구도와 캐릭터의 외형 및 액션을 프롬프트에 맞게 훌륭히 구현했습니다."
   },
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "전체적인 구도와 캐릭터 구현은 우수하나, 권위 있는 지시 사항인 담뱃갑 경고 문구에 오타('질방의 민인이 됩나다')가 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S37sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "담뱃갑에서 튀어나온 담배 한 개비가 인물의 콧구멍 속으로 직접 들어가 있습니다.",
     "fix_en": "Remove the cigarette protruding into the nostril, replacing it with a flat pack top. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "담뱃갑을 쥐고 있는 손가락들의 마디가 비정상적으로 길고 관절 형태가 심하게 왜곡되어 있습니다.",
     "fix_en": "Redraw the hand to have normal human fingers cleanly gripping the pack. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "타이트한 사선 얼굴 클로즈업이 아니라 상반신·흉패·접견실 표지판까지 넓게 들어와 얼굴이 화면 중앙-오른쪽을 차지하지 않는다",
     "fix_en": "Adjust framing to a tight oblique close-up with the face center-right. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "코 아래 담뱃갑의 브랜드 면이 비스듬히 보이지 않고 카메라에 거의 정면으로 향해 있다",
     "fix_en": "Rotate the pack to an oblique angle. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "담뱃갑에서 튀어나온 담배 한 개비가 인물의 콧구멍 속으로 직접 들어가 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "담뱃갑을 쥐고 있는 손가락들의 마디가 비정상적으로 길고 관절 형태가 심하게 왜곡되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "타이트한 사선 얼굴 클로즈업이 아니라 상반신·흉패·접견실 표지판까지 넓게 들어와 얼굴이 화면 중앙-오른쪽을 차지하지 않는다",
     "severity": "major"
    },
    {
     "issue_ko": "코 아래 담뱃갑의 브랜드 면이 비스듬히 보이지 않고 카메라에 거의 정면으로 향해 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "타이트한 사선 얼굴 클로즈업이 아니라 상반신·흉패·접견실 표지판까지 넓게 들어와 얼굴이 화면 중앙-오른쪽을 차지하지 않는다",
    "fix_en": "Adjust framing to a tight oblique close-up with the face center-right. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "코 아래 담뱃갑의 브랜드 면이 비스듬히 보이지 않고 카메라에 거의 정면으로 향해 있다",
    "fix_en": "Rotate the pack to an oblique angle. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the cigarette protruding into the nostril, replacing it with a flat pack top. Preserve the people present and their positions, their clothing, the set, the light, and the framing.\n- Redraw the hand to have normal human fingers cleanly gripping the pack. Preserve the people present and their positions, their clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 화각을 정확히 구현하였고, 다른 인물을 배제한 채 지국현이 눈을 감고 담뱃갑 냄새를 맡는 핵심 행동과 표면의 텍스트까지 완벽하게 표현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "명시적으로 배제하라고 지시한 다른 인물들을 모두 포함시켰으며, 클로즈업이 아닌 와이드 샷으로 연출하고 지시된 행동(냄새 맡기, 눈 감기)을 전혀 수행하지 않아 프롬프트를 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "지국현이 눈을 감고 코끝에 댄 담뱃갑을 향해 고개를 살짝 숙인 채 냄새를 맡고 있음.",
      "built_space": "프레임 좌측 하단에 테이블의 좁은 가장자리가 보이며, 배경은 교도소 면회실 벽면과 창문 일부임.",
      "entities": "지국현 단독 등장. 수의(교도스 1234)를 입고 있으며, 정확한 경고 문구(흡연은 폐암 등 각종 질병의 원인이 됩니다)가 적힌 말보로 레드 담뱃갑을 들고 있음.",
      "hard_violations": [],
      "physics": "손이 담뱃갑을 지탱하고 있으며, 인물은 의자에 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "지국현이 눈을 뜨고 정면을 응시하고 있으며, 담뱃갑은 가슴 높이에 들려 있을 뿐 코에 닿지 않음.",
      "built_space": "테이블 전체와 면회실 공간이 넓게 보이며 4명의 인물이 각각 자리에 앉아 있음.",
      "entities": "지국현 외에 프롬프트가 배제할 것을 명시한 수사관 2명과 교도관 1명이 등장함. 담뱃갑 텍스트가 형태가 뭉개져 있음.",
      "hard_violations": [
       "프롬프트에서 명시적으로 제외할 것을 요구한 추가 인물(수사관, 교도관) 포함"
      ],
      "physics": "손이 담뱃갑을 쥐고 있고 인물들은 바닥과 의자에 의해 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 화각을 정확히 구현하였고, 다른 인물을 배제한 채 지국현이 눈을 감고 담뱃갑 냄새를 맡는 핵심 행동과 표면의 텍스트까지 완벽하게 표현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "명시적으로 배제하라고 지시한 다른 인물들을 모두 포함시켰으며, 클로즈업이 아닌 와이드 샷으로 연출하고 지시된 행동(냄새 맡기, 눈 감기)을 전혀 수행하지 않아 프롬프트를 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "지국현이 눈을 감고 코끝에 댄 담뱃갑을 향해 고개를 살짝 숙인 채 냄새를 맡고 있음.",
      "built_space": "프레임 좌측 하단에 테이블의 좁은 가장자리가 보이며, 배경은 교도소 면회실 벽면과 창문 일부임.",
      "entities": "지국현 단독 등장. 수의(교도스 1234)를 입고 있으며, 정확한 경고 문구(흡연은 폐암 등 각종 질병의 원인이 됩니다)가 적힌 말보로 레드 담뱃갑을 들고 있음.",
      "hard_violations": [],
      "physics": "손이 담뱃갑을 지탱하고 있으며, 인물은 의자에 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "지국현이 눈을 뜨고 정면을 응시하고 있으며, 담뱃갑은 가슴 높이에 들려 있을 뿐 코에 닿지 않음.",
      "built_space": "테이블 전체와 면회실 공간이 넓게 보이며 4명의 인물이 각각 자리에 앉아 있음.",
      "entities": "지국현 외에 프롬프트가 배제할 것을 명시한 수사관 2명과 교도관 1명이 등장함. 담뱃갑 텍스트가 형태가 뭉개져 있음.",
      "hard_violations": [
       "프롬프트에서 명시적으로 제외할 것을 요구한 추가 인물(수사관, 교도관) 포함"
      ],
      "physics": "손이 담뱃갑을 쥐고 있고 인물들은 바닥과 의자에 의해 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요청된 클로즈업 프레이밍, 눈을 감고 냄새를 맡는 연기, 담뱃갑의 지정된 텍스트와 주변 인물 제외 지시를 모두 정확하게 구현하여 지시사항에 완벽히 부합함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍 및 주변 인물 제외 지시를 전혀 따르지 않았으며, 눈을 감고 냄새를 맡는 핵심 연기 또한 구현되지 않아 지시사항을 크게 위반함."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "고개를 약간 숙이고 눈을 감은 채 코가 담뱃갑 상단을 향하고 있음.",
      "built_space": "좌측 하단에 테이블 가장자리가 좁게 보이며 배경에 취조실 벽면이 위치함.",
      "entities": "지국현 단독 등장. 말보로 레드 담뱃갑에 지정된 한글 경고문 텍스트가 정확히 일치함.",
      "hard_violations": [],
      "physics": "오른손으로 담뱃갑을 쥐고 코 아래에 안정적으로 받치고 있음(담배 한 개비가 튀어나와 코에 닿아 있음)."
     },
     {
      "label": "A",
      "direction": "눈을 뜨고 맞은편에 앉은 수사관을 향해 시선을 던지고 있음.",
      "built_space": "테이블을 사이에 두고 4개의 의자에 인물들이 각각 착석해 있는 취조실 구조가 넓게 보임.",
      "entities": "지국현 외에 제외가 지시된 수사관 2명과 교도관 1명이 모두 프레임에 포함됨. 담뱃갑의 텍스트가 뭉개짐.",
      "hard_violations": [
       "클로즈업 프레이밍 및 샷 스케일 지시 위반",
       "제외가 명시된 주변 인물(수사관, 교도관)을 모두 포함함"
      ],
      "physics": "오른손으로 담뱃갑을 들어 올린 채 유지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요청된 클로즈업 프레이밍, 눈을 감고 냄새를 맡는 연기, 담뱃갑의 지정된 텍스트와 주변 인물 제외 지시를 모두 정확하게 구현하여 지시사항에 완벽히 부합함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍 및 주변 인물 제외 지시를 전혀 따르지 않았으며, 눈을 감고 냄새를 맡는 핵심 연기 또한 구현되지 않아 지시사항을 크게 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "고개를 약간 숙이고 눈을 감은 채 코가 담뱃갑 상단을 향하고 있음.",
      "built_space": "좌측 하단에 테이블 가장자리가 좁게 보이며 배경에 취조실 벽면이 위치함.",
      "entities": "지국현 단독 등장. 말보로 레드 담뱃갑에 지정된 한글 경고문 텍스트가 정확히 일치함.",
      "hard_violations": [],
      "physics": "오른손으로 담뱃갑을 쥐고 코 아래에 안정적으로 받치고 있음(담배 한 개비가 튀어나와 코에 닿아 있음)."
     },
     {
      "label": "B",
      "direction": "눈을 뜨고 맞은편에 앉은 수사관을 향해 시선을 던지고 있음.",
      "built_space": "테이블을 사이에 두고 4개의 의자에 인물들이 각각 착석해 있는 취조실 구조가 넓게 보임.",
      "entities": "지국현 외에 제외가 지시된 수사관 2명과 교도관 1명이 모두 프레임에 포함됨. 담뱃갑의 텍스트가 뭉개짐.",
      "hard_violations": [
       "클로즈업 프레이밍 및 샷 스케일 지시 위반",
       "제외가 명시된 주변 인물(수사관, 교도관)을 모두 포함함"
      ],
      "physics": "오른손으로 담뱃갑을 들어 올린 채 유지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S37sh2"
  }
 },
 "S37sh17::cine": {
  "applied": true,
  "fingerprint": "c4b56633e8b9cb7d19e0d7c7278c3d164f89934a55690687278360eae3faff91",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S37sh17_sel.png",
  "source_sha256": "458b4b1e72c93dcbf176f819cbbbf4ff18a94e520e3193bac3df0b20fcf68b83",
  "file": "S37sh17_cine.png",
  "latency_ms": 11992
 },
 "S37sh28::signage": {
  "fp": "0ebd5fa636c48a29",
  "inscriptions": [
   {
    "surface_native": "수사기록 서류철",
    "text_native": "수사기록",
    "reason_ko": "형사와 피의자 간의 숨 막히는 대치 상황에서 갈등의 원인이 되는 사건 파일임을 직관적으로 보여주기 위해 테이블 위 서류철에 필요한 표기입니다."
   }
  ]
 },
 "S37sh28": {
  "input_fingerprint": "a265b65d30a40b90",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 살벌한 눈빛으로 서로를 쏘아보는 서의용과 지국현의 팽팽한 대치\n\nLOCATION (lock): Across the table inside the prison interview room, between the suspect's and detectives' seats. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The dolly-in slows at the side of the table, slightly below seated chest height and perpendicular to the opponents' eyeline, preserving 서의용 and 지국현 (현재) in balanced opposing profiles. 서의용 strains forward at frame left while 나상혁's restraining presence clips the left edge, and 지국현 remains composed at frame right with the table forming an unbroken barrier between their locked gazes.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, looks toward 지국현 across the table; 지국현 (현재) in the middle-right of the frame, midground, looks toward 서의용 across the table; table separating the opponents in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between 서의용 and 지국현) — Its long side crosses the lower center perpendicular to the opponents' locked eyeline; used as The uninterrupted edge divides the two profiles and sustains the visual stalemate.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime ambience keeps the confrontation severe but naturalistic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same interrogation-room table, bare walls, harsh light, and the prisoner's uniform. Exclude the senior officer and guard in the rear, focusing on the detective and prisoner confronting each other.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok has taken back and put away the Marlboro pack and lighter before Euiyong and Ji Gukhyeon's final confrontation.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사기록 서류철: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 살벌한 눈빛으로 서로를 쏘아보는 서의용과 지국현의 팽팽한 대치\n\nLOCATION (lock): Across the table inside the prison interview room, between the suspect's and detectives' seats. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The dolly-in slows at the side of the table, slightly below seated chest height and perpendicular to the opponents' eyeline, preserving 서의용 and 지국현 (현재) in balanced opposing profiles. 서의용 strains forward at frame left while 나상혁's restraining presence clips the left edge, and 지국현 remains composed at frame right with the table forming an unbroken barrier between their locked gazes.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, looks toward 지국현 across the table; 지국현 (현재) in the middle-right of the frame, midground, looks toward 서의용 across the table; table separating the opponents in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between 서의용 and 지국현) — Its long side crosses the lower center perpendicular to the opponents' locked eyeline; used as The uninterrupted edge divides the two profiles and sustains the visual stalemate.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime ambience keeps the confrontation severe but naturalistic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same interrogation-room table, bare walls, harsh light, and the prisoner's uniform. Exclude the senior officer and guard in the rear, focusing on the detective and prisoner confronting each other.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok has taken back and put away the Marlboro pack and lighter before Euiyong and Ji Gukhyeon's final confrontation.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사기록 서류철: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블을 사이에 두고 살벌한 눈빛으로 서로를 쏘아보는 서의용과 지국현의 팽팽한 대치\n\nLOCATION (lock): Across the table inside the prison interview room, between the suspect's and detectives' seats. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The dolly-in slows at the side of the table, slightly below seated chest height and perpendicular to the opponents' eyeline, preserving 서의용 and 지국현 (현재) in balanced opposing profiles. 서의용 strains forward at frame left while 나상혁's restraining presence clips the left edge, and 지국현 remains composed at frame right with the table forming an unbroken barrier between their locked gazes.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, midground, looks toward 지국현 across the table; 지국현 (현재) in the middle-right of the frame, midground, looks toward 서의용 across the table; table separating the opponents in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: investigation table (positioned between 서의용 and 지국현) — Its long side crosses the lower center perpendicular to the opponents' locked eyeline; used as The uninterrupted edge divides the two profiles and sustains the visual stalemate.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime ambience keeps the confrontation severe but naturalistic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same interrogation-room table, bare walls, harsh light, and the prisoner's uniform. Exclude the senior officer and guard in the rear, focusing on the detective and prisoner confronting each other.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok has taken back and put away the Marlboro pack and lighter before Euiyong and Ji Gukhyeon's final confrontation.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사기록 서류철: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "서의용과 지국현이 테이블을 사이에 두고 서로의 눈을 똑바로 응시하고 있음.",
    "built_space": "두 인물이 테이블을 사이에 두고 마주 앉아 있으며, 카메라가 둘의 측면을 정확히 포착함. 테이블 위에는 '수사기록' 서류가 바르게 놓여 있음.",
    "entities": "서의용(비니, 가죽재킷)과 지국현(죄수복)의 외모가 레퍼런스와 잘 일치함. 왼쪽 가장자리에 나상혁의 얼굴 일부와 제지하는 손이 나타남.",
    "hard_violations": [],
    "physics": "서의용은 몸을 앞으로 기울이고 있으며, 왼쪽 인물의 손이 그의 어깨를 단단히 잡아 제지하고 있음. 지국현은 의자에 차분하게 앉아 있음."
   },
   {
    "label": "A",
    "direction": "서의용과 지국현 모두 서로를 보지 않고 프레임 왼쪽의 서 있는 인물을 향해 시선을 향하고 있음.",
    "built_space": "두 인물이 테이블의 같은 쪽(혹은 모서리)에 나란히 앉아 있으며, 테이블이 둘 사이의 장벽 역할을 하지 못함.",
    "entities": "서의용과 지국현의 외모는 레퍼런스와 일치하나, 왼쪽 가장자리에 얼굴이 잘린 정장 차림의 인물이 크게 서 있음.",
    "hard_violations": [
     "마주 보는 측면 구도(opposing profiles)와 테이블을 사이에 둔 배치라는 핵심 카메라/프레임 지시를 완전히 위반함."
    ],
    "physics": "두 인물 모두 의자에 앉아 중력을 받으며 안정적인 자세를 취하고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 8,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "프롬프트가 요구한 완벽한 측면 대치 구도와 시선 교환, 인물의 복장과 디테일을 충실하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "마주 보는 측면 구도와 테이블 배치라는 핵심 연출 지시를 완전히 무시하고 두 인물을 같은 방향으로 배치했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S37sh17_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프레임 왼쪽 가장자리에 샷 텍스트·인물 목록에 없는 정장 남성이 서의용 어깨를 잡고 있다",
     "fix_en": "Remove the man in the suit on the far left edge, including his hand on the shoulder, replacing him with the background window and wall. Preserve the two main characters, their poses, clothing, the table, and the folder exactly as they are.",
     "severity": "critical",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "PEOPLE 규정(SHOT TEXT에 명시된 서의용과 지국현 외에는 절대 추가 불가)에 어긋나게 화면 왼쪽에 제3의 인물(얼굴 측면과 상체)이 뚜렷하게 등장함.",
     "severity": "critical"
    },
    {
     "issue_ko": "프레임 왼쪽 가장자리에 샷 텍스트·인물 목록에 없는 정장 남성이 서의용 어깨를 잡고 있다",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the man in the suit on the far left edge, including his hand on the shoulder, replacing him with the background window and wall. Preserve the two main characters, their poses, clothing, the table, and the folder exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "두 인물의 팽팽한 시선 교환, 명시된 카메라 프레이밍, 이전 샷에서 담배를 치운 상태, 서류철의 정확한 텍스트까지 프롬프트의 요구사항을 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "프롬프트가 요구한 대치 상황을 전혀 반영하지 않고, 이전 샷의 단독 흡연 장면을 그대로 반복하여 모든 지시사항을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "화면 왼쪽의 서의용과 오른쪽의 지국현이 테이블을 사이에 두고 서로의 눈을 강렬하게 똑바로 응시하고 있습니다.",
      "built_space": "접견실 내부이며, 두 사람 사이에 긴 테이블이 가로놓여 대치를 위한 물리적 장벽 역할을 하고 있습니다. 배경 벽면에는 이전 샷과 동일한 '접견실' 표지판이 있습니다.",
      "entities": "서의용(비니, 가죽 재킷, 배지), 지국현(수의 '교도스 1234')의 외형이 레퍼런스와 일치합니다. 프레임 왼쪽 가장자리에 서의용을 제지하는 나상혁의 모습이 일부 보이며, 테이블 위에는 '수사기록'이 정확히 적힌 서류철이 놓여 있습니다. 이전 샷의 담배는 지시대로 치워져 있습니다.",
      "hard_violations": [],
      "physics": "나상혁의 손이 서의용의 어깨를 단단히 짚고 있으며, 서류철은 테이블 표면에 안정적으로 놓여 빛과 질감을 자연스럽게 반사하고 있습니다."
     },
     {
      "label": "B",
      "direction": "지국현이 시선을 아래로 향한 채 담배갑에 집중하고 있으며, 대치해야 할 대상(서의용)을 바라보지 않습니다.",
      "built_space": "접견실 테이블 앞에 인물이 단독으로 앉아 있으며, 뒤쪽 벽면에 표지판이 위치해 있습니다.",
      "entities": "지국현 단 한 명만 등장하며 샷 텍스트가 요구한 서의용과 나상혁이 누락되었습니다. 또한 이미 치워졌어야 할 담배갑을 그대로 들고 있습니다.",
      "hard_violations": [
       "지정된 카메라 프레이밍(두 사람의 측면 대치) 및 앵글을 완전히 무시한 채 이전 샷의 구도를 반복함"
      ],
      "physics": "손이 담배갑을 안정적으로 쥐고 입에 대고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "두 인물의 팽팽한 시선 교환, 명시된 카메라 프레이밍, 이전 샷에서 담배를 치운 상태, 서류철의 정확한 텍스트까지 프롬프트의 요구사항을 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "프롬프트가 요구한 대치 상황을 전혀 반영하지 않고, 이전 샷의 단독 흡연 장면을 그대로 반복하여 모든 지시사항을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "화면 왼쪽의 서의용과 오른쪽의 지국현이 테이블을 사이에 두고 서로의 눈을 강렬하게 똑바로 응시하고 있습니다.",
      "built_space": "접견실 내부이며, 두 사람 사이에 긴 테이블이 가로놓여 대치를 위한 물리적 장벽 역할을 하고 있습니다. 배경 벽면에는 이전 샷과 동일한 '접견실' 표지판이 있습니다.",
      "entities": "서의용(비니, 가죽 재킷, 배지), 지국현(수의 '교도스 1234')의 외형이 레퍼런스와 일치합니다. 프레임 왼쪽 가장자리에 서의용을 제지하는 나상혁의 모습이 일부 보이며, 테이블 위에는 '수사기록'이 정확히 적힌 서류철이 놓여 있습니다. 이전 샷의 담배는 지시대로 치워져 있습니다.",
      "hard_violations": [],
      "physics": "나상혁의 손이 서의용의 어깨를 단단히 짚고 있으며, 서류철은 테이블 표면에 안정적으로 놓여 빛과 질감을 자연스럽게 반사하고 있습니다."
     },
     {
      "label": "B",
      "direction": "지국현이 시선을 아래로 향한 채 담배갑에 집중하고 있으며, 대치해야 할 대상(서의용)을 바라보지 않습니다.",
      "built_space": "접견실 테이블 앞에 인물이 단독으로 앉아 있으며, 뒤쪽 벽면에 표지판이 위치해 있습니다.",
      "entities": "지국현 단 한 명만 등장하며 샷 텍스트가 요구한 서의용과 나상혁이 누락되었습니다. 또한 이미 치워졌어야 할 담배갑을 그대로 들고 있습니다.",
      "hard_violations": [
       "지정된 카메라 프레이밍(두 사람의 측면 대치) 및 앵글을 완전히 무시한 채 이전 샷의 구도를 반복함"
      ],
      "physics": "손이 담배갑을 안정적으로 쥐고 입에 대고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "서의용과 지국현의 팽팽한 대치 상황을 정확한 앵글과 배치로 구현했으며, 테이블 위 '수사기록' 텍스트와 인물들의 복장까지 레퍼런스를 완벽하게 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "프롬프트가 요구한 두 인물의 대치 상황과 서의용을 완전히 누락했으며, 치워야 할 담뱃갑을 여전히 들고 있어 지시사항을 이행하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "서의용과 지국현이 테이블을 사이에 두고 서로의 눈을 날카롭게 마주보고 있으며, 좌측 가장자리의 인물이 서의용의 어깨를 잡고 있음.",
      "built_space": "두 인물 사이에 수사 테이블이 올바르게 위치하며, 이전 컷의 좁고 차가운 접견실 배경과 조명이 그대로 유지됨.",
      "entities": "서의용과 지국현의 외모 및 복장이 레퍼런스와 정확히 일치하며, 테이블 위 서류철에 '수사기록'이라는 지정된 텍스트가 명확하게 적혀 있음.",
      "hard_violations": [],
      "physics": "인물들이 의자에 앉아 테이블을 향해 상체를 기울인 자세가 자연스러우며, 어깨에 얹어진 손과 테이블 위 서류철이 물리적으로 안정되게 지지됨."
     },
     {
      "label": "A",
      "direction": "지국현이 시선을 아래로 향한 채 담뱃갑을 코에 대고 있으며, 대치해야 할 상대방은 화면에 없음.",
      "built_space": "테이블과 배경은 이전 컷의 형태를 띠고 있으나, 카메라 앵글과 인물 배치가 샷 텍스트가 요구한 마주보는 구도(medium shot)를 따르지 않음.",
      "entities": "지국현(현재) 1명만 등장하여 필수 인물인 서의용이 누락됨. 프롬프트에서 이미 치웠다고 명시한 말보로 담뱃갑이 여전히 존재함.",
      "hard_violations": [
       "샷 텍스트에 명시된 필수 인물(서의용) 누락 및 대치 상황 미구현",
       "프롬프트의 지시(Carried state)를 어기고 치워져야 할 담뱃갑을 들고 있음"
      ],
      "physics": "손으로 담뱃갑을 들고 얼굴에 밀착한 포즈 자체는 물리적으로 성립함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "서의용과 지국현의 팽팽한 대치 상황을 정확한 앵글과 배치로 구현했으며, 테이블 위 '수사기록' 텍스트와 인물들의 복장까지 레퍼런스를 완벽하게 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "프롬프트가 요구한 두 인물의 대치 상황과 서의용을 완전히 누락했으며, 치워야 할 담뱃갑을 여전히 들고 있어 지시사항을 이행하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서의용과 지국현이 테이블을 사이에 두고 서로의 눈을 날카롭게 마주보고 있으며, 좌측 가장자리의 인물이 서의용의 어깨를 잡고 있음.",
      "built_space": "두 인물 사이에 수사 테이블이 올바르게 위치하며, 이전 컷의 좁고 차가운 접견실 배경과 조명이 그대로 유지됨.",
      "entities": "서의용과 지국현의 외모 및 복장이 레퍼런스와 정확히 일치하며, 테이블 위 서류철에 '수사기록'이라는 지정된 텍스트가 명확하게 적혀 있음.",
      "hard_violations": [],
      "physics": "인물들이 의자에 앉아 테이블을 향해 상체를 기울인 자세가 자연스러우며, 어깨에 얹어진 손과 테이블 위 서류철이 물리적으로 안정되게 지지됨."
     },
     {
      "label": "B",
      "direction": "지국현이 시선을 아래로 향한 채 담뱃갑을 코에 대고 있으며, 대치해야 할 상대방은 화면에 없음.",
      "built_space": "테이블과 배경은 이전 컷의 형태를 띠고 있으나, 카메라 앵글과 인물 배치가 샷 텍스트가 요구한 마주보는 구도(medium shot)를 따르지 않음.",
      "entities": "지국현(현재) 1명만 등장하여 필수 인물인 서의용이 누락됨. 프롬프트에서 이미 치웠다고 명시한 말보로 담뱃갑이 여전히 존재함.",
      "hard_violations": [
       "샷 텍스트에 명시된 필수 인물(서의용) 누락 및 대치 상황 미구현",
       "프롬프트의 지시(Carried state)를 어기고 치워져야 할 담뱃갑을 들고 있음"
      ],
      "physics": "손으로 담뱃갑을 들고 얼굴에 밀착한 포즈 자체는 물리적으로 성립함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 20,
     "B": 1
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S37sh17"
  }
 },
 "S37sh28::cine": {
  "applied": true,
  "fingerprint": "0e88f6051c9d14dc1752e8d9a00459fd29d91a870ed2e06c39c1de3f817c49f6",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S37sh28_sel.png",
  "source_sha256": "5bbf2e6fc8771d63b17960dc6361f6e03a9eb99a083b9e3a80aa52ff35060c9e",
  "file": "S37sh28_cine.png",
  "latency_ms": 10225
 },
 "S38sh3::signage": {
  "fp": "c64129cbc171c40f",
  "inscriptions": [
   {
    "surface_native": "독방 문 옆 명패",
    "text_native": "제1독거실",
    "reason_ko": "2000년대 한국 교도소의 독방 분위기를 사실적으로 묘사하기 위해 문 옆에 부착된 독거실 표시판이 필요합니다."
   }
  ]
 },
 "era_assess::9c4f870eb912d75f": {
  "subjects": [
   {
    "subject_native": "대한민국 교도소 독방 및 복도 (2001년~2017년)",
    "search_terms_native": [
     "대한민국 교도소 독방",
     "교도소 복도",
     "교도소 내부 복도",
     "독방 문 철창"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "일반적인 이미지 모델은 서구식 철창 감옥을 그리기 쉬우나, 동시기 한국 교도소의 독방 문(배식구와 관찰창이 있는 불투명한 철제 문), 복도 벽면 색상, 바닥 재질 등은 뚜렷한 국산 양식의 차이가 있습니다."
   }
  ]
 },
 "era_ref::e2ba3f241aca0381": {
  "subject": "대한민국 교도소 독방 및 복도 (2001년~2017년)",
  "terms": [
   "대한민국 교도소 독방",
   "교도소 복도",
   "교도소 내부 복도",
   "독방 문 철창"
  ],
  "queries": [
   [
    "대한민국 교도소 독방 복도 내부 2001년 2017년",
    "대한민국 교도소 독방 문 철창 교도소 내부 복도"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://img3.daumcdn.net/thumb/R658x0.q70/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fnews%2F201811%2F10%2Fhani%2F20181110094605803zkpf.jpg",
  "picked_reason_ko": "대한민국의 2001~2017년 교도소 독방 구역을 가장 평범하고 선명하게 보여 주며, 복도와 반복되는 철제 방문·벽체·조명·바닥의 재료와 비례를 읽기 쉽다.",
  "sha256": "d44ed3a6abd543f167002e5a4cb18ac2a8c4bddfa3ac15761917d7f735c5677d",
  "file": "eraref_e2ba3f241aca0381.png"
 },
 "S38sh3::bgfirst_bg": {
  "input_fingerprint": "63a8e210f6f28479",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 독방 철문의 문턱을 한쪽 발로 딛고 안으로 몸을 들이는 지국현의 뒷모습\n\nLOCATION (lock): At the threshold between the prison corridor and an open solitary cell.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hip height directly behind 지국현 (현재), the camera begins tracking from just outside the open cell doorway on a diagonal line across the threshold. His back occupies the middle of the frame as one foot plants inside and his weight follows forward, with the open iron door and the cell interior defining the direction of entry rather than presenting a static rear pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground, moves toward cell interior; cell interior beyond the threshold in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: solitary-cell iron door (open) — The open leaf is seen from behind and partly edge-on as the doorway faces into the cell; used as Its open leaf and frame create the threshold through which the camera follows 지국현; cell threshold (being crossed); used as Marks the exact crossing point beneath 지국현's leading foot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime institutional ambience maintains moderate contrast and a restrained documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 교도소 독방 및 복도 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 열린 독방 철문의 문턱을 한쪽 발로 딛고 안으로 몸을 들이는 지국현의 뒷모습\n\nLOCATION (lock): At the threshold between the prison corridor and an open solitary cell.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hip height directly behind 지국현 (현재), the camera begins tracking from just outside the open cell doorway on a diagonal line across the threshold. His back occupies the middle of the frame as one foot plants inside and his weight follows forward, with the open iron door and the cell interior defining the direction of entry rather than presenting a static rear pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground, moves toward cell interior; cell interior beyond the threshold in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: solitary-cell iron door (open) — The open leaf is seen from behind and partly edge-on as the doorway faces into the cell; used as Its open leaf and frame create the threshold through which the camera follows 지국현; cell threshold (being crossed); used as Marks the exact crossing point beneath 지국현's leading foot.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime institutional ambience maintains moderate contrast and a restrained documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 교도소 독방 및 복도 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S38sh3__bgfirst_bg.png",
  "asset_id": "7f99f7ae-aac7-4966-8912-4bd7ea031ec5",
  "input_asset_ids": [
   "e8ffb79f-2598-40ed-bc78-dfffb3e7e73c",
   "3b0271f1-31b9-4248-9018-c29ce53afbe5"
  ],
  "era_research": {
   "subject": "대한민국 교도소 독방 및 복도 (2001년~2017년)",
   "queries": [
    [
     "대한민국 교도소 독방 복도 내부 2001년 2017년",
     "대한민국 교도소 독방 문 철창 교도소 내부 복도"
    ]
   ],
   "picked_url": "https://img3.daumcdn.net/thumb/R658x0.q70/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fnews%2F201811%2F10%2Fhani%2F20181110094605803zkpf.jpg",
   "sha256": "d44ed3a6abd543f167002e5a4cb18ac2a8c4bddfa3ac15761917d7f735c5677d",
   "file": "eraref_e2ba3f241aca0381.png"
  }
 },
 "S38sh3": {
  "input_fingerprint": "6db3c5a5e907e7f1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 독방 철문의 문턱을 한쪽 발로 딛고 안으로 몸을 들이는 지국현의 뒷모습\n\nLOCATION (lock): At the threshold between the prison corridor and an open solitary cell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hip height directly behind 지국현 (현재), the camera begins tracking from just outside the open cell doorway on a diagonal line across the threshold. His back occupies the middle of the frame as one foot plants inside and his weight follows forward, with the open iron door and the cell interior defining the direction of entry rather than presenting a static rear pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground, moves toward cell interior; cell interior beyond the threshold in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: solitary-cell iron door (open) — The open leaf is seen from behind and partly edge-on as the doorway faces into the cell; used as Its open leaf and frame create the threshold through which the camera follows 지국현; cell threshold (being crossed); used as Marks the exact crossing point beneath 지국현's leading foot.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime institutional ambience maintains moderate contrast and a restrained documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The solitary-cell door remains open as Ji Gukhyeon crosses its threshold into the cell.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 독방 문 옆 명패: \"제1독거실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 독방 철문의 문턱을 한쪽 발로 딛고 안으로 몸을 들이는 지국현의 뒷모습\n\nLOCATION (lock): At the threshold between the prison corridor and an open solitary cell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hip height directly behind 지국현 (현재), the camera begins tracking from just outside the open cell doorway on a diagonal line across the threshold. His back occupies the middle of the frame as one foot plants inside and his weight follows forward, with the open iron door and the cell interior defining the direction of entry rather than presenting a static rear pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground, moves toward cell interior; cell interior beyond the threshold in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: solitary-cell iron door (open) — The open leaf is seen from behind and partly edge-on as the doorway faces into the cell; used as Its open leaf and frame create the threshold through which the camera follows 지국현; cell threshold (being crossed); used as Marks the exact crossing point beneath 지국현's leading foot.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime institutional ambience maintains moderate contrast and a restrained documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The solitary-cell door remains open as Ji Gukhyeon crosses its threshold into the cell.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 독방 문 옆 명패: \"제1독거실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 열린 독방 철문의 문턱을 한쪽 발로 딛고 안으로 몸을 들이는 지국현의 뒷모습\n\nLOCATION (lock): At the threshold between the prison corridor and an open solitary cell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hip height directly behind 지국현 (현재), the camera begins tracking from just outside the open cell doorway on a diagonal line across the threshold. His back occupies the middle of the frame as one foot plants inside and his weight follows forward, with the open iron door and the cell interior defining the direction of entry rather than presenting a static rear pose.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 지국현 (현재) in the middle-center of the frame, midground, moves toward cell interior; cell interior beyond the threshold in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: solitary-cell iron door (open) — The open leaf is seen from behind and partly edge-on as the doorway faces into the cell; used as Its open leaf and frame create the threshold through which the camera follows 지국현; cell threshold (being crossed); used as Marks the exact crossing point beneath 지국현's leading foot.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime institutional ambience maintains moderate contrast and a restrained documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The solitary-cell door remains open as Ji Gukhyeon crosses its threshold into the cell.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 독방 문 옆 명패: \"제1독거실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S38sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S38sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:994284>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L41B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:994284>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "참조된 복도의 깊이감과 대각선 카메라 앵글을 훌륭히 재현했으나, 수의 앞면의 이름표가 등에 잘못 그려진 점은 아쉽습니다."
     },
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "벽면을 정면으로 바라보는 평면적인 구도로 인해 지시된 대각선 앵글과 복도의 장소감을 상실했으며, A와 동일하게 이름표 위치 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.571,
      "B": 1.714
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "참조된 복도의 깊이감과 대각선 카메라 앵글을 훌륭히 재현했으나, 수의 앞면의 이름표가 등에 잘못 그려진 점은 아쉽습니다."
     },
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "벽면을 정면으로 바라보는 평면적인 구도로 인해 지시된 대각선 앵글과 복도의 장소감을 상실했으며, A와 동일하게 이름표 위치 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프레임 중앙에 감방 내부를 배치하라는 구도 요구를 정확히 따랐으며 공간 구조가 물리적으로 타당합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "감방 내부를 프레임 중앙에 배치하라는 지시를 어겼으며, 불가능한 건축 구조를 묘사하여 치명적인 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물이 감방 안을 향해 걷고 있으며 방향이 내부로 명확히 향함.",
      "built_space": "카메라가 복도 측면 벽을 정면으로 바라보며, 열린 문과 감방 내부가 프레임 중앙에 정확히 위치함.",
      "entities": "지국현의 뒷모습. 파란색 죄수복을 입었으나 앞가슴에 있어야 할 수인번호 표가 등에 잘못 묘사됨.",
      "hard_violations": [],
      "physics": "왼발이 문턱 안쪽에 닿아 있고 오른발은 바닥을 밀며 걷는 동작을 자연스럽게 지지함."
     },
     {
      "label": "B",
      "direction": "인물이 감방 안을 향해 발을 내딛고 있음.",
      "built_space": "측면 벽과 문이 카메라와 평행하게 정면을 향하는데, 바로 옆 복도는 1점 투시로 깊어지는 불가능한 기하학적 구조를 보여줌.",
      "entities": "지국현의 뒷모습. 파란색 죄수복 착용, 수인번호 표가 등에 잘못 묘사됨.",
      "hard_violations": [
       "측면 문과 정면을 향해 깊어지는 복도가 공존하는 물리적으로 불가능한 공간 구조"
      ],
      "physics": "왼발을 문턱에 딛고 오른발을 든 상태로 동작 자체는 지지되고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프레임 중앙에 감방 내부를 배치하라는 구도 요구를 정확히 따랐으며 공간 구조가 물리적으로 타당합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "감방 내부를 프레임 중앙에 배치하라는 지시를 어겼으며, 불가능한 건축 구조를 묘사하여 치명적인 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물이 감방 안을 향해 걷고 있으며 방향이 내부로 명확히 향함.",
      "built_space": "카메라가 복도 측면 벽을 정면으로 바라보며, 열린 문과 감방 내부가 프레임 중앙에 정확히 위치함.",
      "entities": "지국현의 뒷모습. 파란색 죄수복을 입었으나 앞가슴에 있어야 할 수인번호 표가 등에 잘못 묘사됨.",
      "hard_violations": [],
      "physics": "왼발이 문턱 안쪽에 닿아 있고 오른발은 바닥을 밀며 걷는 동작을 자연스럽게 지지함."
     },
     {
      "label": "A",
      "direction": "인물이 감방 안을 향해 발을 내딛고 있음.",
      "built_space": "측면 벽과 문이 카메라와 평행하게 정면을 향하는데, 바로 옆 복도는 1점 투시로 깊어지는 불가능한 기하학적 구조를 보여줌.",
      "entities": "지국현의 뒷모습. 파란색 죄수복 착용, 수인번호 표가 등에 잘못 묘사됨.",
      "hard_violations": [
       "측면 문과 정면을 향해 깊어지는 복도가 공존하는 물리적으로 불가능한 공간 구조"
      ],
      "physics": "왼발을 문턱에 딛고 오른발을 든 상태로 동작 자체는 지지되고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1574,
     "B": 1721
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 1574,
   "B": 1721
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "참조된 복도의 깊이감과 대각선 카메라 앵글을 훌륭히 재현했으나, 수의 앞면의 이름표가 등에 잘못 그려진 점은 아쉽습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "벽면을 정면으로 바라보는 평면적인 구도로 인해 지시된 대각선 앵글과 복도의 장소감을 상실했으며, A와 동일하게 이름표 위치 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L41B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "캐릭터 레퍼런스상 앞가슴에 위치해야 할 수인번호 표('교도소 1234')가 캐릭터의 등 윗부분에 잘못 렌더링됨.",
     "fix_en": "Erase the white patch and text from the character's upper back, filling the area with plain blue denim fabric to match the rest of the shirt. Preserve the person present and his position, his clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "뒤쪽에 있는 왼쪽 발과 신발이 해부학적으로 불가능하게 180도 뒤틀려 발뒤꿈치가 앞을 향하고 있음.",
     "fix_en": "Redraw the trailing left foot and brown shoe so the heel faces the camera and the toe points forward into the cell, fixing the backward anatomical twist. Preserve the person present and his position, his clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "캐릭터 레퍼런스상 앞가슴에 위치해야 할 수인번호 표('교도소 1234')가 캐릭터의 등 윗부분에 잘못 렌더링됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "뒤쪽에 있는 왼쪽 발과 신발이 해부학적으로 불가능하게 180도 뒤틀려 발뒤꿈치가 앞을 향하고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "등 중앙의 교도소 1234 패치가 참조 의상의 앞가슴이 아닌 등에 붙어 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Erase the white patch and text from the character's upper back, filling the area with plain blue denim fabric to match the rest of the shirt. Preserve the person present and his position, his clothing, the set, the light, and the framing.\n- Redraw the trailing left foot and brown shoe so the heel faces the camera and the toe points forward into the cell, fixing the backward anatomical twist. Preserve the person present and his position, his clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "수인번호 표식이 등에 나타나는 오류를 제거하고, 지국현이 독방으로 들어가는 뒷모습과 '제1독거실' 명패 등 프롬프트의 요구사항을 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물이 독방으로 들어가는 샷의 구성은 좋으나, 가슴에 있어야 할 수인번호 패치가 등에 잘못 렌더링되는 치명적인 복장 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물은 열린 감방 문을 통과해 방 안쪽을 향해 나아가고 있음.",
      "built_space": "교도소 복도와 열린 철문, 감방 내부가 보임. 문의 방향과 손잡이 위치가 올바름.",
      "entities": "지국현의 뒷모습이 보이지만 가슴에 위치해야 할 죄수번호 명찰이 등에 잘못 그려져 있음. 벽에 '제1독거실' 명패가 있음.",
      "hard_violations": [
       "가슴에 있어야 할 수인번호 명찰이 등에 위치함 (의상 구조 오류)"
      ],
      "physics": "오른발이 바닥을 딛고 체중을 지탱하며, 왼발은 앞으로 내딛기 위해 들려 있음."
     },
     {
      "label": "B",
      "direction": "인물은 열린 감방 문을 통과해 방 안쪽을 향해 나아가고 있음.",
      "built_space": "교도소 복도, 열린 철문, 감방 내부가 참조 사진과 일치하게 렌더링됨.",
      "entities": "지국현의 뒷모습과 죄수복이 올바르게 묘사됨(등에 명찰 없음). 벽에 '제1독거실' 명패가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "오른발로 문턱 부근을 딛고 체중을 지지하며, 왼발은 앞으로 내딛기 위해 들려 있는 자연스러운 보행 상태임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "수인번호 표식이 등에 나타나는 오류를 제거하고, 지국현이 독방으로 들어가는 뒷모습과 '제1독거실' 명패 등 프롬프트의 요구사항을 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물이 독방으로 들어가는 샷의 구성은 좋으나, 가슴에 있어야 할 수인번호 패치가 등에 잘못 렌더링되는 치명적인 복장 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물은 열린 감방 문을 통과해 방 안쪽을 향해 나아가고 있음.",
      "built_space": "교도소 복도와 열린 철문, 감방 내부가 보임. 문의 방향과 손잡이 위치가 올바름.",
      "entities": "지국현의 뒷모습이 보이지만 가슴에 위치해야 할 죄수번호 명찰이 등에 잘못 그려져 있음. 벽에 '제1독거실' 명패가 있음.",
      "hard_violations": [
       "가슴에 있어야 할 수인번호 명찰이 등에 위치함 (의상 구조 오류)"
      ],
      "physics": "오른발이 바닥을 딛고 체중을 지탱하며, 왼발은 앞으로 내딛기 위해 들려 있음."
     },
     {
      "label": "B",
      "direction": "인물은 열린 감방 문을 통과해 방 안쪽을 향해 나아가고 있음.",
      "built_space": "교도소 복도, 열린 철문, 감방 내부가 참조 사진과 일치하게 렌더링됨.",
      "entities": "지국현의 뒷모습과 죄수복이 올바르게 묘사됨(등에 명찰 없음). 벽에 '제1독거실' 명패가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "오른발로 문턱 부근을 딛고 체중을 지지하며, 왼발은 앞으로 내딛기 위해 들려 있는 자연스러운 보행 상태임."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 뒷모습 앵글에 맞춰 죄수복을 정확히 묘사했으며, 명패 텍스트('제1독거실')와 문턱을 넘는 동작 등 모든 프롬프트 요구사항을 완벽하게 충족했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "명패와 공간 구성은 훌륭하나, 캐릭터 레퍼런스상 앞가슴에 있는 수인번호 패치를 등에 잘못 부착하는 결정적인 의상 오류를 범했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "남자의 몸과 시선이 열린 감방 안쪽을 향하고 있음.",
      "built_space": "교도소 복도, 열린 철문, 감방 내부가 올바른 비례로 배치되어 있으며 문 우측 벽에 명패가 있음.",
      "entities": "지국현의 뒷모습이 레퍼런스와 일치하는 푸른색 죄수복을 입고 나타남. 벽의 명패에 '제1독거실'이 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "오른발은 복도 바닥을, 왼발은 문턱을 자연스럽게 딛고 있어 걷는 동작의 체중 이동이 잘 표현됨."
     },
     {
      "label": "B",
      "direction": "남자의 몸과 시선이 열린 감방 안쪽을 향하고 있음.",
      "built_space": "교도소 복도, 열린 철문, 감방 내부가 올바른 비례로 배치되어 있으며 문 우측 벽에 명패가 있음.",
      "entities": "지국현의 뒷모습이나, 앞가슴에 있어야 할 '교도소 1234' 명찰이 등에 잘못 붙어 있음. 벽의 명패에 '제1독거실'이 적혀 있음.",
      "hard_violations": [],
      "physics": "오른발은 복도 바닥을, 왼발은 문턱을 자연스럽게 딛고 있어 걷는 동작이 물리적으로 타당함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "지정된 뒷모습 앵글에 맞춰 죄수복을 정확히 묘사했으며, 명패 텍스트('제1독거실')와 문턱을 넘는 동작 등 모든 프롬프트 요구사항을 완벽하게 충족했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "명패와 공간 구성은 훌륭하나, 캐릭터 레퍼런스상 앞가슴에 있는 수인번호 패치를 등에 잘못 부착하는 결정적인 의상 오류를 범했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "남자의 몸과 시선이 열린 감방 안쪽을 향하고 있음.",
      "built_space": "교도소 복도, 열린 철문, 감방 내부가 올바른 비례로 배치되어 있으며 문 우측 벽에 명패가 있음.",
      "entities": "지국현의 뒷모습이 레퍼런스와 일치하는 푸른색 죄수복을 입고 나타남. 벽의 명패에 '제1독거실'이 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "오른발은 복도 바닥을, 왼발은 문턱을 자연스럽게 딛고 있어 걷는 동작의 체중 이동이 잘 표현됨."
     },
     {
      "label": "A",
      "direction": "남자의 몸과 시선이 열린 감방 안쪽을 향하고 있음.",
      "built_space": "교도소 복도, 열린 철문, 감방 내부가 올바른 비례로 배치되어 있으며 문 우측 벽에 명패가 있음.",
      "entities": "지국현의 뒷모습이나, 앞가슴에 있어야 할 '교도소 1234' 명찰이 등에 잘못 붙어 있음. 벽의 명패에 '제1독거실'이 적혀 있음.",
      "hard_violations": [],
      "physics": "오른발은 복도 바닥을, 왼발은 문턱을 자연스럽게 딛고 있어 걷는 동작이 물리적으로 타당함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 19
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S38sh3__bgfirst_bg.png",
   "bg_asset_id": "7f99f7ae-aac7-4966-8912-4bd7ea031ec5",
   "bg_record_key": "S38sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S38sh3::cine": {
  "applied": true,
  "fingerprint": "94272abe35f0f9720cd0be13933d58bb27c76f44107dd5493a881edd0de9d16f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S38sh3_sel.png",
  "source_sha256": "05b8aaeb8041b07bb6dafa268a2d7b23726ebbb406db71a86538fbd8500f176a",
  "file": "S38sh3_cine.png",
  "latency_ms": 10278
 },
 "S38sh4::signage": {
  "fp": "31477ec5ffc96b48",
  "inscriptions": [
   {
    "surface_native": "벽에 붙은 쪽지",
    "text_native": "여호와는 나의 목자시니 내게 부족함이 없으리로다",
    "reason_ko": "지국현의 뒤로 보이는 독방 벽면에 다닥다닥 붙어 있는 실제 성경 구절 쪽지의 사실성을 높이기 위함입니다."
   },
   {
    "surface_native": "벽에 붙은 또 다른 쪽지",
    "text_native": "두려워하지 말라 내가 너와 함께 함이라",
    "reason_ko": "독방 내부의 쓸쓸하고 종교적인 분위기를 강조하기 위해 벽에 붙은 구절을 한국어 성경 텍스트로 묘사합니다."
   }
  ]
 },
 "S38sh4": {
  "input_fingerprint": "7c915f6ff25e4389",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성경 구절 쪽지들이 붙은 독방 벽을 배경으로 무표정하게 선 지국현의 감정 없는 얼굴 클로즈업\n\nLOCATION (lock): Inside the sparse solitary cell, directly before the wall covered with Bible-verse notes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Inside the cell at upper-chest height, the completed arc settles into a close front three-quarter angle slightly off 지국현 (현재)'s direct axis as the dolly reaches its endpoint. His emotionless face occupies the center-left, while the Bible-verse notes remain legible as a clustered background layer over his shoulder rather than becoming decorative foreground graphics.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bible-verse notes (attached to the cell wall) — Their written faces are turned toward the camera on the wall behind 지국현; used as The clustered written faces remain behind 지국현 as narrative context for his blank expression; cell wall (bearing the Bible-verse notes); used as Provides the restrained background plane against which the face and notes separate.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves an emotionless, naturalistic facial rendering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Bible-verse notes remain affixed across the wall of Ji Gukhyeon's solitary cell.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽에 붙은 쪽지: \"여호와는 나의 목자시니 내게 부족함이 없으리로다\"\n- 벽에 붙은 또 다른 쪽지: \"두려워하지 말라 내가 너와 함께 함이라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성경 구절 쪽지들이 붙은 독방 벽을 배경으로 무표정하게 선 지국현의 감정 없는 얼굴 클로즈업\n\nLOCATION (lock): Inside the sparse solitary cell, directly before the wall covered with Bible-verse notes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Inside the cell at upper-chest height, the completed arc settles into a close front three-quarter angle slightly off 지국현 (현재)'s direct axis as the dolly reaches its endpoint. His emotionless face occupies the center-left, while the Bible-verse notes remain legible as a clustered background layer over his shoulder rather than becoming decorative foreground graphics.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bible-verse notes (attached to the cell wall) — Their written faces are turned toward the camera on the wall behind 지국현; used as The clustered written faces remain behind 지국현 as narrative context for his blank expression; cell wall (bearing the Bible-verse notes); used as Provides the restrained background plane against which the face and notes separate.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves an emotionless, naturalistic facial rendering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Bible-verse notes remain affixed across the wall of Ji Gukhyeon's solitary cell.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽에 붙은 쪽지: \"여호와는 나의 목자시니 내게 부족함이 없으리로다\"\n- 벽에 붙은 또 다른 쪽지: \"두려워하지 말라 내가 너와 함께 함이라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성경 구절 쪽지들이 붙은 독방 벽을 배경으로 무표정하게 선 지국현의 감정 없는 얼굴 클로즈업\n\nLOCATION (lock): Inside the sparse solitary cell, directly before the wall covered with Bible-verse notes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Inside the cell at upper-chest height, the completed arc settles into a close front three-quarter angle slightly off 지국현 (현재)'s direct axis as the dolly reaches its endpoint. His emotionless face occupies the center-left, while the Bible-verse notes remain legible as a clustered background layer over his shoulder rather than becoming decorative foreground graphics.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: Bible-verse notes (attached to the cell wall) — Their written faces are turned toward the camera on the wall behind 지국현; used as The clustered written faces remain behind 지국현 as narrative context for his blank expression; cell wall (bearing the Bible-verse notes); used as Provides the restrained background plane against which the face and notes separate.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime interior ambience preserves an emotionless, naturalistic facial rendering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The Bible-verse notes remain affixed across the wall of Ji Gukhyeon's solitary cell.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽에 붙은 쪽지: \"여호와는 나의 목자시니 내게 부족함이 없으리로다\"\n- 벽에 붙은 또 다른 쪽지: \"두려워하지 말라 내가 너와 함께 함이라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "이전 샷의 깨끗한 벽면과 복장(수형번호 패치) 디테일을 완벽히 유지했으며, 요구된 프레이밍과 쪽지의 텍스트를 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "인물 묘사와 텍스트 구현은 우수하나, 이전 샷에 없던 심한 벽면 페인트 벗겨짐이 추가되어 장소 일관성(Location Lock)에 위배됩니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 지국현 (현재) — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S38sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   },
   {
    "label": "PROP REFERENCE — 성경 구절 메모 쪽지: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1448620>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "상단 쪽지의 텍스트가 프롬프트에 명시된 '내게'가 아닌 '내가'로 잘못 출력되었습니다.",
     "fix_en": "Change the text '내가' to '내게' on the top note, preserving Ji Gukhyeon's face, posture, clothing, the background wall, and the lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "지정된 텍스트 외에 추가 텍스트를 금지한 프롬프트 지시와 달리, 쪽지 하단에 '(시편 23:1)' 및 '(이사야 41:10)'이라는 내용이 임의로 추가되었습니다.",
     "fix_en": "Remove the text '(시편 23:1)' and '(이사야 41:10)' from the notes, replacing them with blank paper texture, preserving Ji Gukhyeon's face, posture, clothing, the wall, and the lighting.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "셔츠 앞섶의 단추들이 단춧구멍을 통과하는 구조가 아니라 바깥쪽 원단 위에 덧붙여진 형태로 잘못 렌더링되었습니다.",
     "fix_en": "Redraw the shirt placket so the buttons emerge through buttonholes from the underlying fabric layer, preserving Ji Gukhyeon's face, expression, the rest of the uniform, the notes, and the lighting.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "배경 벽이 이전 스틸의 독방 내부(단색)가 아니라 복도와 같은 상단 크림·하단 민트 투톤이다.",
     "fix_en": "Repaint the lower mint-green section of the background wall to an off-white color matching the upper section, preserving Ji Gukhyeon, his uniform, the notes, and the lighting.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "쪽지가 벽에 군집해 붙어 있지 않고 화면 오른쪽 머리 옆에 두 장만 가지런히 붙어 있다.",
     "fix_en": "Add several more overlapping Bible-verse notes around the existing ones to create a clustered arrangement on the wall, preserving Ji Gukhyeon, his pose, uniform, the wall, and the lighting.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "성경 쪽지가 어깨 너머 배경층이 아니라 얼굴과 비슷한 크기의 장식 그래픽처럼 크게 보인다.",
     "fix_en": "Reduce the scale of the Bible-verse notes so they appear as small paper slips on the distant wall, preserving Ji Gukhyeon, his pose, uniform, the wall color, and the lighting.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "수인번호 패치가 인물의 오른쪽 가슴에 위치해 있으나, 캐릭터 레퍼런스 기준 왼쪽 가슴에 있어야 합니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "상단 쪽지의 텍스트가 프롬프트에 명시된 '내게'가 아닌 '내가'로 잘못 출력되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "지정된 텍스트 외에 추가 텍스트를 금지한 프롬프트 지시와 달리, 쪽지 하단에 '(시편 23:1)' 및 '(이사야 41:10)'이라는 내용이 임의로 추가되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "셔츠 앞섶의 단추들이 단춧구멍을 통과하는 구조가 아니라 바깥쪽 원단 위에 덧붙여진 형태로 잘못 렌더링되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "배경 벽이 이전 스틸의 독방 내부(단색)가 아니라 복도와 같은 상단 크림·하단 민트 투톤이다.",
     "severity": "major"
    },
    {
     "issue_ko": "쪽지가 벽에 군집해 붙어 있지 않고 화면 오른쪽 머리 옆에 두 장만 가지런히 붙어 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "성경 쪽지가 어깨 너머 배경층이 아니라 얼굴과 비슷한 크기의 장식 그래픽처럼 크게 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 6,
  "fix_severity_skipped": [
   {
    "issue_ko": "상단 쪽지의 텍스트가 프롬프트에 명시된 '내게'가 아닌 '내가'로 잘못 출력되었습니다.",
    "fix_en": "Change the text '내가' to '내게' on the top note, preserving Ji Gukhyeon's face, posture, clothing, the background wall, and the lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "지정된 텍스트 외에 추가 텍스트를 금지한 프롬프트 지시와 달리, 쪽지 하단에 '(시편 23:1)' 및 '(이사야 41:10)'이라는 내용이 임의로 추가되었습니다.",
    "fix_en": "Remove the text '(시편 23:1)' and '(이사야 41:10)' from the notes, replacing them with blank paper texture, preserving Ji Gukhyeon's face, posture, clothing, the wall, and the lighting.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "셔츠 앞섶의 단추들이 단춧구멍을 통과하는 구조가 아니라 바깥쪽 원단 위에 덧붙여진 형태로 잘못 렌더링되었습니다.",
    "fix_en": "Redraw the shirt placket so the buttons emerge through buttonholes from the underlying fabric layer, preserving Ji Gukhyeon's face, expression, the rest of the uniform, the notes, and the lighting.",
    "severity": "minor",
    "observation_index": 3
   },
   {
    "issue_ko": "배경 벽이 이전 스틸의 독방 내부(단색)가 아니라 복도와 같은 상단 크림·하단 민트 투톤이다.",
    "fix_en": "Repaint the lower mint-green section of the background wall to an off-white color matching the upper section, preserving Ji Gukhyeon, his uniform, the notes, and the lighting.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "쪽지가 벽에 군집해 붙어 있지 않고 화면 오른쪽 머리 옆에 두 장만 가지런히 붙어 있다.",
    "fix_en": "Add several more overlapping Bible-verse notes around the existing ones to create a clustered arrangement on the wall, preserving Ji Gukhyeon, his pose, uniform, the wall, and the lighting.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "성경 쪽지가 어깨 너머 배경층이 아니라 얼굴과 비슷한 크기의 장식 그래픽처럼 크게 보인다.",
    "fix_en": "Reduce the scale of the Bible-verse notes so they appear as small paper slips on the distant wall, preserving Ji Gukhyeon, his pose, uniform, the wall color, and the lighting.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh3"
  }
 },
 "S38sh4::cine": {
  "applied": true,
  "fingerprint": "0d6da14bbf05e3b53e69024934a20f05308c9d1bf91c25c5c0d51c398bd7a408",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S38sh4_sel.png",
  "source_sha256": "8f7f86734a5cb1363266823699fb6b1039a58b7d341acdd7f90f05b6834b5358",
  "file": "S38sh4_cine.png",
  "latency_ms": 12911
 },
 "S39sh2::signage": {
  "fp": "e8b0caa22a0364a2",
  "inscriptions": [
   {
    "surface_native": "벽면 서예 액자",
    "text_native": "정의구현",
    "reason_ko": "2000년대 한국 수사과장실 벽면에 흔히 걸려 있는 한글 서예 액자로, 엄숙한 수사 기관의 공적 분위기를 현실감 있게 묘사합니다."
   }
  ]
 },
 "S39sh2": {
  "input_fingerprint": "bc57a04e1bbb5057",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 주위의 소파에 둘러앉아 심각한 표정으로 바닥을 응시하는 서의용, 나상혁, 강주철의 모습\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the open side of the sofa grouping, the dolly-out ends in a slightly high, diagonal wide composition with 전택수 held at the head position and the three investigators distributed around him. 서의용, 나상혁, and 강주철 each lower their gaze in a different phase of contemplation—one folded forward, one held back, one perched near the edge—so the group reads as individuals sharing the same impasse rather than a repeated pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: sofa grouping (occupied by the four investigators) — The open side faces the camera, with seating surfaces turning inward toward 전택수; used as The grouping defines 전택수's head position and the investigators' surrounding arrangement; office floor (visible within the sofa grouping); used as Visible below the investigators' lowered sightlines as shared negative space.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime office ambience supports a somber, deliberative mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 서예 액자: \"정의구현\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 주위의 소파에 둘러앉아 심각한 표정으로 바닥을 응시하는 서의용, 나상혁, 강주철의 모습\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the open side of the sofa grouping, the dolly-out ends in a slightly high, diagonal wide composition with 전택수 held at the head position and the three investigators distributed around him. 서의용, 나상혁, and 강주철 each lower their gaze in a different phase of contemplation—one folded forward, one held back, one perched near the edge—so the group reads as individuals sharing the same impasse rather than a repeated pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: sofa grouping (occupied by the four investigators) — The open side faces the camera, with seating surfaces turning inward toward 전택수; used as The grouping defines 전택수's head position and the investigators' surrounding arrangement; office floor (visible within the sofa grouping); used as Visible below the investigators' lowered sightlines as shared negative space.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime office ambience supports a somber, deliberative mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 서예 액자: \"정의구현\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 주위의 소파에 둘러앉아 심각한 표정으로 바닥을 응시하는 서의용, 나상혁, 강주철의 모습\n\nLOCATION (lock): In the sofa meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the open side of the sofa grouping, the dolly-out ends in a slightly high, diagonal wide composition with 전택수 held at the head position and the three investigators distributed around him. 서의용, 나상혁, and 강주철 each lower their gaze in a different phase of contemplation—one folded forward, one held back, one perched near the edge—so the group reads as individuals sharing the same impasse rather than a repeated pose.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: sofa grouping (occupied by the four investigators) — The open side faces the camera, with seating surfaces turning inward toward 전택수; used as The grouping defines 전택수's head position and the investigators' surrounding arrangement; office floor (visible within the sofa grouping); used as Visible below the investigators' lowered sightlines as shared negative space.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Moderate-contrast neutral-to-cool daytime office ambience supports a somber, deliberative mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 서예 액자: \"정의구현\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "네 명의 인물이 모두 소파 그룹 중앙의 바닥을 심각하게 응시함.",
    "built_space": "다크 브라운 가죽 소파가 ㄷ자 형태로 배치됨. 배경 벽면에 '정의구현' 서예 액자가 정상적으로 보임.",
    "entities": "서의용, 강주철, 나상혁은 레퍼런스 의상과 일치. 중앙의 전택수는 레퍼런스(경찰복)와 완전히 다른 정장을 착용함.",
    "hard_violations": [
     "로케이션 잠금 위반: 이전 샷의 베이지색 패브릭 소파가 다크 브라운 가죽 소파로 변경됨",
     "캐릭터 의상 위반: 전택수가 레퍼런스의 경찰 점퍼 대신 정장을 착용함"
    ],
    "physics": "모든 인물이 소파 좌석에 체중을 싣고 안정적으로 착석해 있음."
   },
   {
    "label": "B",
    "direction": "네 명의 인물이 각각 바닥 쪽으로 시선을 내리고 있음.",
    "built_space": "다크 브라운 가죽 소파가 거의 일렬이나 얕은 ㄱ자로 배치됨. 뒤편 벽에 '정의구현' 액자가 있음.",
    "entities": "서의용, 강주철, 나상혁의 의상은 레퍼런스와 일치. 중앙 좌측의 전택수는 정장 조끼 차림으로 레퍼런스와 불일치함.",
    "hard_violations": [
     "로케이션 잠금 위반: 이전 샷과 텍스처, 색상이 완전히 다른 가죽 소파 렌더링",
     "캐릭터 의상 위반: 전택수의 경찰복 미반영",
     "구도 위반: 전택수가 헤드 포지션에 위치하는 ㄷ자형 둘러앉은 배치가 아님"
    ],
    "physics": "인물들이 소파에 앉아 있으나 상호작용 없는 일렬 배치 느낌이 강함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 4,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "요구된 ㄷ자형 소파 배치와 헤드 포지션 구도는 잘 구현했으나, 로케이션(베이지색 패브릭 소파)과 전택수의 복장(경찰복)을 무시한 치명적 오류가 있습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "로케이션 및 전택수 복장 오류뿐만 아니라, 지정된 헤드 포지션과 소파 배치 구도마저 따르지 않아 완성도가 매우 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S27sh10_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "소파 안쪽 머리 자리 남성이 전택수 참조와 얼굴·머리·복장이 전혀 다르고 수사 제복 대신 정장을 입고 있다",
     "fix_en": "Replace the central seated man with the Jeon Taek-soo character reference: square face, navy police jacket, cap, plaid shirt, and khaki pants. Maintain the current camera angle, the exact faces, clothing, and positions of the other three men, the sofa, and the room.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "소파가 이전 스틸의 밝은 패브릭이 아니라 갈색 가죽 코너 소파로 바뀌어 있다",
     "fix_en": "Change the dark leather sofa to a light grey fabric sofa matching the previous shot still. Keep the men, their poses, and the rest of the room unchanged.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "오른쪽 끝 강주철이 바닥이 아니라 옆쪽을 보고 있다",
     "fix_en": "Turn the head and gaze of the man on the right downward to the floor. Maintain his posture, the other characters, the sofa, and the room.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "앞쪽 나상혁 재킷 소매에 참조에 없는 로고 패치가 있다",
     "fix_en": "Erase the logo patch on the foreground man's left sleeve, leaving plain navy fabric. Maintain the jacket, the character's pose, the other men, and the room.",
     "severity": "minor",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "중앙에 앉은 전택수가 지정된 레퍼런스 복장(경찰 점퍼와 모자)이 아닌 정장을 입고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "극적인 순간을 포착하는 장면임에도 중앙의 인물이 카메라 렌즈를 정면으로 의식하여 응시하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷 레퍼런스에 등장한 밝은 색상의 패브릭 소파가 어두운 갈색 가죽 소파로 완전히 변경되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "소파 안쪽 머리 자리 남성이 전택수 참조와 얼굴·머리·복장이 전혀 다르고 수사 제복 대신 정장을 입고 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "소파가 이전 스틸의 밝은 패브릭이 아니라 갈색 가죽 코너 소파로 바뀌어 있다",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 끝 강주철이 바닥이 아니라 옆쪽을 보고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "앞쪽 나상혁 재킷 소매에 참조에 없는 로고 패치가 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "소파가 이전 스틸의 밝은 패브릭이 아니라 갈색 가죽 코너 소파로 바뀌어 있다",
    "fix_en": "Change the dark leather sofa to a light grey fabric sofa matching the previous shot still. Keep the men, their poses, and the rest of the room unchanged.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "오른쪽 끝 강주철이 바닥이 아니라 옆쪽을 보고 있다",
    "fix_en": "Turn the head and gaze of the man on the right downward to the floor. Maintain his posture, the other characters, the sofa, and the room.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "앞쪽 나상혁 재킷 소매에 참조에 없는 로고 패치가 있다",
    "fix_en": "Erase the logo patch on the foreground man's left sleeve, leaving plain navy fabric. Maintain the jacket, the character's pose, the other men, and the room.",
    "severity": "minor",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 6,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the central seated man with the Jeon Taek-soo character reference: square face, navy police jacket, cap, plaid shirt, and khaki pants. Maintain the current camera angle, the exact faces, clothing, and positions of the other three men, the sofa, and the room.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 넓은 앵글과 인물들이 소파에 둘러앉아 바닥을 응시하는 연출, '정의구현' 텍스트를 훌륭히 구현했으나, 이전 샷과 소파 재질이 다르고 전택수의 의상이 레퍼런스와 불일치하는 단점이 있습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지시되지 않은 인물이 추가되는 치명적 위반이 있으며, 인물들이 서 있거나 카메라를 응시하여 샷 텍스트의 핵심 연출을 완전히 무시했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "좌측, 우측, 전경의 세 인물은 모두 바닥을 향해 시선을 내리고 있으며, 중앙의 인물은 약간 앞쪽 아래를 응시함.",
      "built_space": "갈색 가죽 소재의 U자형 소파가 중앙에 배치됨. (이전 샷의 옅은 회색 패브릭 소파와 형태 및 재질이 다름).",
      "entities": "서의용(좌), 강주철(우), 나상혁(전경)은 레퍼런스와 외모 및 의상이 일치함. 벽면 액자에 '정의구현' 글자가 정확함. 단, 중앙의 전택수는 지정된 경찰복 대신 정장을 입고 있음.",
      "hard_violations": [],
      "physics": "4명의 인물 모두 소파에 체중을 싣고 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "프레임에 등장하는 4명의 인물 모두 정면의 카메라를 응시하고 있음.",
      "built_space": "갈색 가죽 소파와 어수선한 서류 더미가 있는 어두운 실내로, 이전 샷의 밝은 사무실 로케이션과 불일치함.",
      "entities": "전택수(중앙 앉음), 강주철(좌측 섰음), 나상혁(우측 섰음)의 의상은 일치하나 서의용이 누락됨. 배경에 프롬프트에 없는 회색 스웨터 차림의 남성이 등장함.",
      "hard_violations": [
       "발명된 인물: 지시되지 않은 회색 스웨터 남성이 배경에 추가됨."
      ],
      "physics": "1명은 소파에 앉아 있고, 나머지 3명은 바닥에 서서 체중을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지정된 넓은 앵글과 인물들이 소파에 둘러앉아 바닥을 응시하는 연출, '정의구현' 텍스트를 훌륭히 구현했으나, 이전 샷과 소파 재질이 다르고 전택수의 의상이 레퍼런스와 불일치하는 단점이 있습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지시되지 않은 인물이 추가되는 치명적 위반이 있으며, 인물들이 서 있거나 카메라를 응시하여 샷 텍스트의 핵심 연출을 완전히 무시했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "좌측, 우측, 전경의 세 인물은 모두 바닥을 향해 시선을 내리고 있으며, 중앙의 인물은 약간 앞쪽 아래를 응시함.",
      "built_space": "갈색 가죽 소재의 U자형 소파가 중앙에 배치됨. (이전 샷의 옅은 회색 패브릭 소파와 형태 및 재질이 다름).",
      "entities": "서의용(좌), 강주철(우), 나상혁(전경)은 레퍼런스와 외모 및 의상이 일치함. 벽면 액자에 '정의구현' 글자가 정확함. 단, 중앙의 전택수는 지정된 경찰복 대신 정장을 입고 있음.",
      "hard_violations": [],
      "physics": "4명의 인물 모두 소파에 체중을 싣고 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "프레임에 등장하는 4명의 인물 모두 정면의 카메라를 응시하고 있음.",
      "built_space": "갈색 가죽 소파와 어수선한 서류 더미가 있는 어두운 실내로, 이전 샷의 밝은 사무실 로케이션과 불일치함.",
      "entities": "전택수(중앙 앉음), 강주철(좌측 섰음), 나상혁(우측 섰음)의 의상은 일치하나 서의용이 누락됨. 배경에 프롬프트에 없는 회색 스웨터 차림의 남성이 등장함.",
      "hard_violations": [
       "발명된 인물: 지시되지 않은 회색 스웨터 남성이 배경에 추가됨."
      ],
      "physics": "1명은 소파에 앉아 있고, 나머지 3명은 바닥에 서서 체중을 지탱하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "바닥을 응시하며 둘러앉은 구도와 서예 액자 지시를 훌륭히 수행했으나, 전택수의 의상이 임의로 변경되고 이전 샷의 소파 재질이 유지되지 않아 일부 감점되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정되지 않은 인물이 추가되고 핵심 연출 지시(둘러앉아 바닥 응시)를 전혀 따르지 않았으며, 장소 잠금과 필수 텍스트도 무시되어 크게 실패한 결과물입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "대부분의 인물이 지시된 바닥이 아닌 카메라 정면을 응시하고 있음.",
      "built_space": "이전 샷과 완전히 무관한 어두운 창고 형태의 공간이며, 지정된 액자가 없음.",
      "entities": "전택수만 앉아 있고 나머지는 서 있음. 서의용이 누락되고 신원 미상의 인물(회색 스웨터)이 추가됨.",
      "hard_violations": [
       "지시되지 않은 인물 추가",
       "지정된 인물 누락"
      ],
      "physics": "인물들이 바닥과 소파 표면에 의해 정상적으로 지지됨."
     },
     {
      "label": "B",
      "direction": "네 명의 인물 모두 시선을 아래로 내려 명확히 바닥을 응시하고 있음.",
      "built_space": "사무실 배경에 지정된 텍스트 액자가 구현되었으나, 이전 샷의 밝은 패브릭 소파가 아닌 어두운 가죽 소파가 배치됨.",
      "entities": "네 명 모두 소파에 앉아 있음. 3명은 지정 복장과 일치하나 중앙의 전택수는 지시된 경찰복 대신 정장을 입고 있음.",
      "hard_violations": [],
      "physics": "모든 인물이 소파에 정상적으로 앉아 무게가 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "바닥을 응시하며 둘러앉은 구도와 서예 액자 지시를 훌륭히 수행했으나, 전택수의 의상이 임의로 변경되고 이전 샷의 소파 재질이 유지되지 않아 일부 감점되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정되지 않은 인물이 추가되고 핵심 연출 지시(둘러앉아 바닥 응시)를 전혀 따르지 않았으며, 장소 잠금과 필수 텍스트도 무시되어 크게 실패한 결과물입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "대부분의 인물이 지시된 바닥이 아닌 카메라 정면을 응시하고 있음.",
      "built_space": "이전 샷과 완전히 무관한 어두운 창고 형태의 공간이며, 지정된 액자가 없음.",
      "entities": "전택수만 앉아 있고 나머지는 서 있음. 서의용이 누락되고 신원 미상의 인물(회색 스웨터)이 추가됨.",
      "hard_violations": [
       "지시되지 않은 인물 추가",
       "지정된 인물 누락"
      ],
      "physics": "인물들이 바닥과 소파 표면에 의해 정상적으로 지지됨."
     },
     {
      "label": "A",
      "direction": "네 명의 인물 모두 시선을 아래로 내려 명확히 바닥을 응시하고 있음.",
      "built_space": "사무실 배경에 지정된 텍스트 액자가 구현되었으나, 이전 샷의 밝은 패브릭 소파가 아닌 어두운 가죽 소파가 배치됨.",
      "entities": "네 명 모두 소파에 앉아 있음. 3명은 지정 복장과 일치하나 중앙의 전택수는 지시된 경찰복 대신 정장을 입고 있음.",
      "hard_violations": [],
      "physics": "모든 인물이 소파에 정상적으로 앉아 무게가 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 13,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S27sh10"
  }
 },
 "S39sh2::cine": {
  "applied": true,
  "fingerprint": "91a80fb226a029d752dde186a9944ed13e1ba307106c90455df521eb74cfa37f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S39sh2_sel.png",
  "source_sha256": "551aab4deb3e9f26890b64d5a42fb77b7db2cf494c9ee2e5af785371215a592c",
  "file": "S39sh2_cine.png",
  "latency_ms": 10880
 },
 "S39sh4::signage": {
  "fp": "c9f953993c20a03c",
  "inscriptions": []
 },
 "S39sh4": {
  "input_fingerprint": "ced1876de0ac4c7e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽으로 고개를 홱 돌린 채 단호한 눈빛으로 입술을 뗀 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At the head of the sofa meeting area inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close to 서의용's side of the conversational axis and slightly below 전택수's eye line, the inward move ends on an oblique close-up as 전택수 snaps his head toward 서의용. 전택수's face holds the center-right with firm eye contact and lips just parting to state his decision, while a narrow, soft trace of 서의용's shoulder at the left edge keeps the response grounded in the group exchange.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: head-position sofa (occupied by 전택수) — Its back runs behind 전택수 and angles away toward the other investigators; used as A limited portion behind 전택수 retains continuity with the surrounding meeting.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime office ambience gives the decisive turn moderate contrast without heightened stylization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa arrangement, subdued daytime light, and the seated group's positions. Exclude the three investigators from the close frame and isolate the senior officer turning decisively.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽으로 고개를 홱 돌린 채 단호한 눈빛으로 입술을 뗀 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At the head of the sofa meeting area inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close to 서의용's side of the conversational axis and slightly below 전택수's eye line, the inward move ends on an oblique close-up as 전택수 snaps his head toward 서의용. 전택수's face holds the center-right with firm eye contact and lips just parting to state his decision, while a narrow, soft trace of 서의용's shoulder at the left edge keeps the response grounded in the group exchange.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: head-position sofa (occupied by 전택수) — Its back runs behind 전택수 and angles away toward the other investigators; used as A limited portion behind 전택수 retains continuity with the surrounding meeting.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime office ambience gives the decisive turn moderate contrast without heightened stylization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa arrangement, subdued daytime light, and the seated group's positions. Exclude the three investigators from the close frame and isolate the senior officer turning decisively.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽으로 고개를 홱 돌린 채 단호한 눈빛으로 입술을 뗀 전택수의 얼굴 클로즈업\n\nLOCATION (lock): At the head of the sofa meeting area inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close to 서의용's side of the conversational axis and slightly below 전택수's eye line, the inward move ends on an oblique close-up as 전택수 snaps his head toward 서의용. 전택수's face holds the center-right with firm eye contact and lips just parting to state his decision, while a narrow, soft trace of 서의용's shoulder at the left edge keeps the response grounded in the group exchange.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: head-position sofa (occupied by 전택수) — Its back runs behind 전택수 and angles away toward the other investigators; used as A limited portion behind 전택수 retains continuity with the surrounding meeting.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime office ambience gives the decisive turn moderate contrast without heightened stylization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa arrangement, subdued daytime light, and the seated group's positions. Exclude the three investigators from the close frame and isolate the senior officer turning decisively.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.375,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.625
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.625
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1625,
   "A": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "요청된 입술이 살짝 벌어지는 표정 디테일과 단호한 시선을 잘 구현하였으며, 숏 텍스트의 분위기를 충실히 따랐습니다."
   },
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "전체적인 구도와 인물 외양은 우수하나, 입술을 굳게 다물고 있어 대사를 시작하려는 표정 지시를 놓쳤습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S39sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트에 '입술을 뗀' 상태로 묘사되어 있으나, 사진 속 인물의 입술이 굳게 닫혀 있습니다.",
     "fix_en": "Part the man's lips slightly so they are open as if he is about to speak. Preserve his face, hairstyle, clothing, gaze direction, the camera angle, lighting, and all background and foreground elements.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 중앙-우측 남성이 전택수(50대 중반·각진 얼굴) 및 캐릭터 레퍼런스와 얼굴·나이가 전혀 다른 인물이다.",
     "fix_en": "Replace the man's face and hair to match the mid-50s man with a square face and short neat black hair from the character reference. Preserve his blue bomber jacket and shirt, the camera angle, lighting, the background office set, and the blurred shoulder on the left.",
     "severity": "critical",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트에 '입술을 뗀' 상태로 묘사되어 있으나, 사진 속 인물의 입술이 굳게 닫혀 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙-우측 남성이 전택수(50대 중반·각진 얼굴) 및 캐릭터 레퍼런스와 얼굴·나이가 전혀 다른 인물이다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트에 '입술을 뗀' 상태로 묘사되어 있으나, 사진 속 인물의 입술이 굳게 닫혀 있습니다.",
    "fix_en": "Part the man's lips slightly so they are open as if he is about to speak. Preserve his face, hairstyle, clothing, gaze direction, the camera angle, lighting, and all background and foreground elements.",
    "severity": "major",
    "observation_index": 0
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the man's face and hair to match the mid-50s man with a square face and short neat black hair from the character reference. Preserve his blue bomber jacket and shirt, the camera angle, lighting, the background office set, and the blurred shoulder on the left.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시된 클로즈업 프레이밍을 정확히 구현하였고, 배제해야 할 인물들을 화면에서 제외한 채 좌측 엣지에 서의용의 어깨를 걸고 전택수의 단호한 표정과 살짝 벌어진 입술을 완벽하게 묘사했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "클로즈업 프레이밍 지시를 완전히 무시하고 와이드 샷을 렌더링하였으며, 화면에서 명시적으로 배제하도록 지시된 3명의 인물을 모두 포함시켰으므로 심각한 위반입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 좌측에 걸친 상대방(서의용)의 어깨 쪽을 향하고 있으며, 입술이 대사를 시작하려는 듯 살짝 벌어져 있다.",
      "built_space": "소파의 등받이가 전택수의 등 뒤로 자연스럽게 이어지며, 배경에 '정의구현' 액자와 수납장 등 이전 샷에 설정된 사무실 환경이 올바른 거리감과 크기로 배치되어 있다.",
      "entities": "전택수의 얼굴이 중심 우측에 클로즈업으로 배치되었으며, 복장(남색 재킷, 하늘색 셔츠)과 외형이 레퍼런스와 일치한다. 화면 좌측 가장자리에 서의용의 어깨가 흐릿하게 걸쳐 있고, 나머지 인물들은 성공적으로 배제되었다.",
      "hard_violations": [],
      "physics": "소파에 앉아 상체를 약간 기울인 자연스러운 자세이며, 물리적 지지나 중력 묘사에 이상이 없다."
     },
     {
      "label": "B",
      "direction": "전택수는 화면 좌측을 향해 시선을 던지고 있으나, 입술은 굳게 닫혀 있다.",
      "built_space": "이전 샷의 사무실 공간 전체가 그대로 렌더링되었으며, 4명의 인물이 소파에 앉아 있는 공간 구조가 넓게 묘사되어 있다.",
      "entities": "전택수의 복장 및 외형은 일치하나, 샷 텍스트에서 명시적으로 제외할 것을 요구한 다른 수사관 3명이 그대로 화면에 등장하고 있다. 좌측 엣지의 서의용 어깨 묘사가 없다.",
      "hard_violations": [
       "명시적으로 배제해야 할 인물들(나머지 수사관 3명)을 프레임 안에 포함시킴",
       "요구된 클로즈업(Close-up) 프레이밍을 어기고 이전 샷과 동일한 와이드 샷을 렌더링함"
      ],
      "physics": "4명의 인물 모두 소파와 바닥에 정상적으로 몸을 지지하고 있으며 물리적인 오류는 보이지 않는다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시된 클로즈업 프레이밍을 정확히 구현하였고, 배제해야 할 인물들을 화면에서 제외한 채 좌측 엣지에 서의용의 어깨를 걸고 전택수의 단호한 표정과 살짝 벌어진 입술을 완벽하게 묘사했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "클로즈업 프레이밍 지시를 완전히 무시하고 와이드 샷을 렌더링하였으며, 화면에서 명시적으로 배제하도록 지시된 3명의 인물을 모두 포함시켰으므로 심각한 위반입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 좌측에 걸친 상대방(서의용)의 어깨 쪽을 향하고 있으며, 입술이 대사를 시작하려는 듯 살짝 벌어져 있다.",
      "built_space": "소파의 등받이가 전택수의 등 뒤로 자연스럽게 이어지며, 배경에 '정의구현' 액자와 수납장 등 이전 샷에 설정된 사무실 환경이 올바른 거리감과 크기로 배치되어 있다.",
      "entities": "전택수의 얼굴이 중심 우측에 클로즈업으로 배치되었으며, 복장(남색 재킷, 하늘색 셔츠)과 외형이 레퍼런스와 일치한다. 화면 좌측 가장자리에 서의용의 어깨가 흐릿하게 걸쳐 있고, 나머지 인물들은 성공적으로 배제되었다.",
      "hard_violations": [],
      "physics": "소파에 앉아 상체를 약간 기울인 자연스러운 자세이며, 물리적 지지나 중력 묘사에 이상이 없다."
     },
     {
      "label": "B",
      "direction": "전택수는 화면 좌측을 향해 시선을 던지고 있으나, 입술은 굳게 닫혀 있다.",
      "built_space": "이전 샷의 사무실 공간 전체가 그대로 렌더링되었으며, 4명의 인물이 소파에 앉아 있는 공간 구조가 넓게 묘사되어 있다.",
      "entities": "전택수의 복장 및 외형은 일치하나, 샷 텍스트에서 명시적으로 제외할 것을 요구한 다른 수사관 3명이 그대로 화면에 등장하고 있다. 좌측 엣지의 서의용 어깨 묘사가 없다.",
      "hard_violations": [
       "명시적으로 배제해야 할 인물들(나머지 수사관 3명)을 프레임 안에 포함시킴",
       "요구된 클로즈업(Close-up) 프레이밍을 어기고 이전 샷과 동일한 와이드 샷을 렌더링함"
      ],
      "physics": "4명의 인물 모두 소파와 바닥에 정상적으로 몸을 지지하고 있으며 물리적인 오류는 보이지 않는다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "요청된 클로즈업 앵글을 완벽히 구현했으며, 좌측 가장자리에 흐릿한 어깨를 배치하여 대화의 맥락을 살리면서도 전택수의 단호한 표정과 행동을 지시사항에 맞게 정확하게 포착했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "프레임(클로즈업) 및 다른 세 명의 수사관을 배제하라는 명시적 지시를 무시하고 이전 샷과 거의 동일한 와이드 샷을 생성하여 치명적인 오류를 범했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "전택수가 화면 좌측 전경에 흐릿하게 걸쳐진 인물의 어깨 쪽으로 고개를 돌려 시선을 고정하고 있음.",
      "built_space": "사무실 내부, 전택수의 뒤로 가죽 소파의 등받이가 각도를 틀며 이어져 있고 배경에 액자가 걸려 있음.",
      "entities": "전택수(네이비색 재킷, 짧은 검은 머리)가 중심인물로 클로즈업되어 있으며, 좌측 가장자리에 서의용으로 추정되는 인물의 흐릿한 어깨가 보임.",
      "hard_violations": [],
      "physics": "전택수가 보이지 않는 소파 좌석에 안정적으로 앉아 몸을 돌린 자세를 취하고 있음."
     },
     {
      "label": "A",
      "direction": "전택수가 화면 좌측의 다른 인물을 향해 고개를 돌려 바라보고 있음.",
      "built_space": "사무실 내부, U자형 가죽 소파에 4명의 인물이 앉아 있음.",
      "entities": "전택수를 포함하여 프레임에서 배제되었어야 할 3명의 다른 수사관들이 모두 선명하게 화면에 나타나 있음.",
      "hard_violations": [
       "지시된 카메라 프레임(클로즈업)을 완전히 무시함",
       "화면에서 배제해야 할 3명의 인물을 프레임에 그대로 포함시킴"
      ],
      "physics": "네 명의 인물이 모두 소파에 물리적으로 정상적으로 앉아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "요청된 클로즈업 앵글을 완벽히 구현했으며, 좌측 가장자리에 흐릿한 어깨를 배치하여 대화의 맥락을 살리면서도 전택수의 단호한 표정과 행동을 지시사항에 맞게 정확하게 포착했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "프레임(클로즈업) 및 다른 세 명의 수사관을 배제하라는 명시적 지시를 무시하고 이전 샷과 거의 동일한 와이드 샷을 생성하여 치명적인 오류를 범했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수가 화면 좌측 전경에 흐릿하게 걸쳐진 인물의 어깨 쪽으로 고개를 돌려 시선을 고정하고 있음.",
      "built_space": "사무실 내부, 전택수의 뒤로 가죽 소파의 등받이가 각도를 틀며 이어져 있고 배경에 액자가 걸려 있음.",
      "entities": "전택수(네이비색 재킷, 짧은 검은 머리)가 중심인물로 클로즈업되어 있으며, 좌측 가장자리에 서의용으로 추정되는 인물의 흐릿한 어깨가 보임.",
      "hard_violations": [],
      "physics": "전택수가 보이지 않는 소파 좌석에 안정적으로 앉아 몸을 돌린 자세를 취하고 있음."
     },
     {
      "label": "B",
      "direction": "전택수가 화면 좌측의 다른 인물을 향해 고개를 돌려 바라보고 있음.",
      "built_space": "사무실 내부, U자형 가죽 소파에 4명의 인물이 앉아 있음.",
      "entities": "전택수를 포함하여 프레임에서 배제되었어야 할 3명의 다른 수사관들이 모두 선명하게 화면에 나타나 있음.",
      "hard_violations": [
       "지시된 카메라 프레임(클로즈업)을 완전히 무시함",
       "화면에서 배제해야 할 3명의 인물을 프레임에 그대로 포함시킴"
      ],
      "physics": "네 명의 인물이 모두 소파에 물리적으로 정상적으로 앉아 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 20,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S39sh2"
  }
 },
 "S39sh4::cine": {
  "applied": true,
  "fingerprint": "174735d0d86ff5f88f7ed1419b778d0d1aa1bade0cf5f25b6e719ca57b4b373a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S39sh4_sel.png",
  "source_sha256": "239e0a5af4256d11ff85185bed668300079f7b700e6ea3d7a02c949faf0d59c4",
  "file": "S39sh4_cine.png",
  "latency_ms": 11695
 },
 "S40sh5::signage": {
  "fp": "c8973dd12355aeac",
  "inscriptions": [
   {
    "surface_native": "사무실 출입문 유리",
    "text_native": "사무실",
    "reason_ko": "자동차 정비소 내 사무실 공간이라는 배경을 사실적으로 나타내기 위해 문 유리에 식별용 문구가 필요합니다."
   }
  ]
 },
 "era_assess::6ccdc2ba70240719": {
  "subjects": [
   {
    "subject_native": "2000년대 한국의 카센터 사무실",
    "search_terms_native": [
     "2000년대 카센터 내부",
     "동네 정비소 사무실",
     "카센터 사무실 옛날"
    ],
    "language_lock_native": "이 검색은 오직 한국어로만 수행되어야 하며, 영어나 다른 언어로 번역되거나 다른 언어의 키워드가 추가되어서는 안 됩니다.",
    "reason_ko": "2000년대 한국 카센터 특유의 기름때 묻은 장부, 달력, 집기류의 배치와 정비창이 보이는 유리 파티션 구조는 일반적인 서양식 차고나 현대적인 차량 정비소와 시각적으로 크게 다릅니다."
   }
  ]
 },
 "era_ref::4072d0b89b9187e0": {
  "subject": "2000년대 한국의 카센터 사무실",
  "terms": [
   "2000년대 카센터 내부",
   "동네 정비소 사무실",
   "카센터 사무실 옛날"
  ],
  "queries": [
   [
    "2000년대 한국 카센터 내부 동네 정비소 사무실 옛날",
    "옛날 카센터 사무실 내부 2000년대 한국"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://d9hmpd388i7ko.cloudfront.net/repairShop_web_medias/2023_4_19igc6dj9235jh8.webp",
  "picked_reason_ko": "사진 1은 낡은 장판 바닥, 형광등, 사무용 책상, 고객 대기용 소파, 타이어 안내물 등 2000년대 한국 카센터 사무실의 평범한 구조와 설비를 가장 선명하고 충실하게 보여준다.",
  "sha256": "c9eff6b8d77e785b99ff6ea9f56e0d8af62ab1503c6047e952b0590bbb33b8b0",
  "file": "eraref_4072d0b89b9187e0.png"
 },
 "groupbg::정비소 사무실": {
  "input_fingerprint": "a0ee296b2e18dde9",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "정비소 사무실",
    "tags": [
     "S40sh5",
     "S40sh8"
    ]
   },
   "context_sig": "487553026d2dac95",
   "era_research_sha": "c9eff6b8d77e785b99ff6ea9f56e0d8af62ab1503c6047e952b0590bbb33b8b0"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: Inside the auto-shop office, with the repair bays visible through the interior window.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n카센터 사무실: 정비소 안쪽에 유리 파티션으로 분리된 어수선하고 기름때 묻은 사무 공간이다. (특징: 투명한 유리 창틀 너머로 보이는 정비 리프트와 자동차 부품들; 부품 상자와 달력, 장부들이 널브러진 어수선한 사무용 책상)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 카센터 사무실 - 낮\n- 유리 창틀 너머로 정비시설이 보이는 사무실 안.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 카센터 사무실: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: Inside the auto-shop office, with the repair bays visible through the interior window.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n카센터 사무실: 정비소 안쪽에 유리 파티션으로 분리된 어수선하고 기름때 묻은 사무 공간이다. (특징: 투명한 유리 창틀 너머로 보이는 정비 리프트와 자동차 부품들; 부품 상자와 달력, 장부들이 널브러진 어수선한 사무용 책상)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 카센터 사무실 - 낮\n- 유리 창틀 너머로 정비시설이 보이는 사무실 안.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 카센터 사무실: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_정비소_사무실_98ba69.png",
  "asset_id": "5eee3cc6-afd0-4f3d-944b-c83aa86f5a4f",
  "input_asset_ids": [
   "18b59834-7ccb-4ce1-a9e8-aa1e853f9eea"
  ],
  "origin_tag": "S40sh5",
  "place_text": "Inside the auto-shop office, with the repair bays visible through the interior window.",
  "origin_inputs": {
   "place_text": "Inside the auto-shop office, with the repair bays visible through the interior window.",
   "time_of_day_en": "day",
   "conti_asset_id": "18b59834-7ccb-4ce1-a9e8-aa1e853f9eea"
  },
  "era_research": {
   "subject": "2000년대 한국의 카센터 사무실",
   "terms": [
    "2000년대 카센터 내부",
    "동네 정비소 사무실",
    "카센터 사무실 옛날"
   ],
   "queries": [
    [
     "2000년대 한국 카센터 내부 동네 정비소 사무실 옛날",
     "옛날 카센터 사무실 내부 2000년대 한국"
    ]
   ],
   "candidates": 4,
   "picked_index": 1,
   "picked_url": "https://d9hmpd388i7ko.cloudfront.net/repairShop_web_medias/2023_4_19igc6dj9235jh8.webp",
   "picked_reason_ko": "사진 1은 낡은 장판 바닥, 형광등, 사무용 책상, 고객 대기용 소파, 타이어 안내물 등 2000년대 한국 카센터 사무실의 평범한 구조와 설비를 가장 선명하고 충실하게 보여준다.",
   "sha256": "c9eff6b8d77e785b99ff6ea9f56e0d8af62ab1503c6047e952b0590bbb33b8b0",
   "file": "eraref_4072d0b89b9187e0.png",
   "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_4072d0b89b9187e0.png"
  }
 },
 "era_assess::6c75b357c166359d": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 카센터 사무실 및 정비고",
    "search_terms_native": [
     "한국 카센터 사무실 내부",
     "국산 자동차 정비소 실내",
     "카센터 정비고 유리창"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 결합하지 마십시오.",
    "reason_ko": "2000년대 한국 카센터 특유의 녹색 레자 소파, 맥심 커피믹스 디스펜서, 한글 정비 요금표, 유리창 너머로 보이는 정비소 내의 국산 리프트와 차량 번호판 등의 한국적 디테일을 정확히 표현하기 위해서입니다."
   }
  ]
 },
 "era_ref::9d55cba103025c33": {
  "subject": "2000년대~2010년대 한국의 카센터 사무실 및 정비고",
  "terms": [
   "한국 카센터 사무실 내부",
   "국산 자동차 정비소 실내",
   "카센터 정비고 유리창"
  ],
  "queries": [
   [
    "2000년대 2010년대 한국 카센터 사무실 내부 정비고 유리창",
    "국산 자동차 정비소 실내 카센터 정비고 사무실"
   ],
   [
    "옛날 카센터 내부 사무실 정비소 유리창",
    "2010년 한국 자동차 정비소 정비고 내부"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://www.gongim.com/uploads/store/1723528334_c852jt.jpg",
  "picked_reason_ko": "3번은 타이어·자동차용품 진열과 업무용 컴퓨터, 상담 좌석이 함께 보여 2000~2010년대 한국의 일상적인 카센터 사무·고객대기 공간을 가장 명확하게 읽을 수 있다.",
  "sha256": "6d89ec5332ca943187defa7f158eba6cf32acd8c24c6ed6da3db1ef76030dc62",
  "file": "eraref_9d55cba103025c33.png"
 },
 "S40sh5::bgfirst_bg": {
  "input_fingerprint": "e4a36d016d5ee940",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용을 향해 불쾌한 기색을 드러내며 눈을 부릅뜬 30대 초반 한국인 남성의 굳은 표정\n\nLOCATION (lock): Inside the auto-shop office, with the repair bays visible through the interior window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Slightly below the seated eye line and close behind 서의용's shoulder, the tracked move settles into a tight over-the-shoulder close-up that stays offset from the direct conversational axis. The Korean man in his early thirties occupies the center-right with rigid shoulders, widened eyes fixed on 서의용, and controlled look space to the left; only a narrow blurred trace of 서의용 remains at the frame edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office glass window (in place between the office and maintenance facility) — The window faces the camera obliquely, with the maintenance facility visible beyond it; used as Kept softly behind the man to retain the office's visual connection to the repair facility; maintenance facility (visible beyond the window); used as Appears only as subdued context through the office window.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic neutral-to-cool daytime ambience keeps the hardened expression clear with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 카센터 사무실 및 정비고: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용을 향해 불쾌한 기색을 드러내며 눈을 부릅뜬 30대 초반 한국인 남성의 굳은 표정\n\nLOCATION (lock): Inside the auto-shop office, with the repair bays visible through the interior window.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Slightly below the seated eye line and close behind 서의용's shoulder, the tracked move settles into a tight over-the-shoulder close-up that stays offset from the direct conversational axis. The Korean man in his early thirties occupies the center-right with rigid shoulders, widened eyes fixed on 서의용, and controlled look space to the left; only a narrow blurred trace of 서의용 remains at the frame edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office glass window (in place between the office and maintenance facility) — The window faces the camera obliquely, with the maintenance facility visible beyond it; used as Kept softly behind the man to retain the office's visual connection to the repair facility; maintenance facility (visible beyond the window); used as Appears only as subdued context through the office window.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic neutral-to-cool daytime ambience keeps the hardened expression clear with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 카센터 사무실 및 정비고: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S40sh5__bgfirst_bg.png",
  "asset_id": "d46c4a56-1aba-49bc-a088-547a8f5b4993",
  "input_asset_ids": [
   "18b59834-7ccb-4ce1-a9e8-aa1e853f9eea",
   "5eee3cc6-afd0-4f3d-944b-c83aa86f5a4f"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국의 카센터 사무실 및 정비고",
   "queries": [
    [
     "2000년대 2010년대 한국 카센터 사무실 내부 정비고 유리창",
     "국산 자동차 정비소 실내 카센터 정비고 사무실"
    ],
    [
     "옛날 카센터 내부 사무실 정비소 유리창",
     "2010년 한국 자동차 정비소 정비고 내부"
    ]
   ],
   "picked_url": "https://www.gongim.com/uploads/store/1723528334_c852jt.jpg",
   "sha256": "6d89ec5332ca943187defa7f158eba6cf32acd8c24c6ed6da3db1ef76030dc62",
   "file": "eraref_9d55cba103025c33.png"
  }
 },
 "S40sh5": {
  "input_fingerprint": "afdb1f7a7c82a6b9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용을 향해 불쾌한 기색을 드러내며 눈을 부릅뜬 30대 초반 한국인 남성의 굳은 표정\n\nLOCATION (lock): Inside the auto-shop office, with the repair bays visible through the interior window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Slightly below the seated eye line and close behind 서의용's shoulder, the tracked move settles into a tight over-the-shoulder close-up that stays offset from the direct conversational axis. The Korean man in his early thirties occupies the center-right with rigid shoulders, widened eyes fixed on 서의용, and controlled look space to the left; only a narrow blurred trace of 서의용 remains at the frame edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office glass window (in place between the office and maintenance facility) — The window faces the camera obliquely, with the maintenance facility visible beyond it; used as Kept softly behind the man to retain the office's visual connection to the repair facility; maintenance facility (visible beyond the window); used as Appears only as subdued context through the office window.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic neutral-to-cool daytime ambience keeps the hardened expression clear with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 출입문 유리: \"사무실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용을 향해 불쾌한 기색을 드러내며 눈을 부릅뜬 30대 초반 한국인 남성의 굳은 표정\n\nLOCATION (lock): Inside the auto-shop office, with the repair bays visible through the interior window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Slightly below the seated eye line and close behind 서의용's shoulder, the tracked move settles into a tight over-the-shoulder close-up that stays offset from the direct conversational axis. The Korean man in his early thirties occupies the center-right with rigid shoulders, widened eyes fixed on 서의용, and controlled look space to the left; only a narrow blurred trace of 서의용 remains at the frame edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office glass window (in place between the office and maintenance facility) — The window faces the camera obliquely, with the maintenance facility visible beyond it; used as Kept softly behind the man to retain the office's visual connection to the repair facility; maintenance facility (visible beyond the window); used as Appears only as subdued context through the office window.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic neutral-to-cool daytime ambience keeps the hardened expression clear with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 출입문 유리: \"사무실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용을 향해 불쾌한 기색을 드러내며 눈을 부릅뜬 30대 초반 한국인 남성의 굳은 표정\n\nLOCATION (lock): Inside the auto-shop office, with the repair bays visible through the interior window. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Slightly below the seated eye line and close behind 서의용's shoulder, the tracked move settles into a tight over-the-shoulder close-up that stays offset from the direct conversational axis. The Korean man in his early thirties occupies the center-right with rigid shoulders, widened eyes fixed on 서의용, and controlled look space to the left; only a narrow blurred trace of 서의용 remains at the frame edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office glass window (in place between the office and maintenance facility) — The window faces the camera obliquely, with the maintenance facility visible beyond it; used as Kept softly behind the man to retain the office's visual connection to the repair facility; maintenance facility (visible beyond the window); used as Appears only as subdued context through the office window.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic neutral-to-cool daytime ambience keeps the hardened expression clear with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 출입문 유리: \"사무실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S40sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S40sh5.png"
    },
    {
     "label": "CHARACTER REFERENCE — 30대 초반 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:762177>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_정비소_사무실_98ba69.png"
    },
    {
     "label": "CHARACTER REFERENCE — 30대 초반 남성: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:762177>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 샷 크기를 잘 구현했으며 눈을 부릅뜬 표정과 사무실 텍스트 디테일이 프롬프트와 정확히 일치함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 클로즈업이 아닌 미디엄 샷으로 프레이밍되었으며 인물의 표정에서 요구된 불쾌감과 부릅뜬 눈의 묘사가 부족함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "오른쪽 인물의 시선이 왼쪽 전경에 흐릿하게 걸친 인물을 향해 정확히 고정됨.",
      "built_space": "패턴 벽지, 가죽 소파, 유리창 너머 정비소 등 로케이션 사진의 요소가 올바르게 배치됨.",
      "entities": "30대 남성의 얼굴과 복장이 레퍼런스와 일치하며 문 유리에 사무실 문구가 정확히 적힘.",
      "hard_violations": [],
      "physics": "어깨와 목의 긴장된 자세가 자연스럽게 지탱되며 물리적 오류 없음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 왼쪽 전경의 흐릿한 인물을 향함.",
      "built_space": "사무실 내부 배경 요소가 존재하나 문의 각도와 텍스트의 수평이 다소 어긋남.",
      "entities": "인물의 외모는 일치하나 표정이 밋밋하며 텍스트가 덧붙여진 느낌임.",
      "hard_violations": [],
      "physics": "서 있는 자세에 물리적으로 불가능한 부분은 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 샷 크기를 잘 구현했으며 눈을 부릅뜬 표정과 사무실 텍스트 디테일이 프롬프트와 정확히 일치함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 클로즈업이 아닌 미디엄 샷으로 프레이밍되었으며 인물의 표정에서 요구된 불쾌감과 부릅뜬 눈의 묘사가 부족함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "오른쪽 인물의 시선이 왼쪽 전경에 흐릿하게 걸친 인물을 향해 정확히 고정됨.",
      "built_space": "패턴 벽지, 가죽 소파, 유리창 너머 정비소 등 로케이션 사진의 요소가 올바르게 배치됨.",
      "entities": "30대 남성의 얼굴과 복장이 레퍼런스와 일치하며 문 유리에 사무실 문구가 정확히 적힘.",
      "hard_violations": [],
      "physics": "어깨와 목의 긴장된 자세가 자연스럽게 지탱되며 물리적 오류 없음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 왼쪽 전경의 흐릿한 인물을 향함.",
      "built_space": "사무실 내부 배경 요소가 존재하나 문의 각도와 텍스트의 수평이 다소 어긋남.",
      "entities": "인물의 외모는 일치하나 표정이 밋밋하며 텍스트가 덧붙여진 느낌임.",
      "hard_violations": [],
      "physics": "서 있는 자세에 물리적으로 불가능한 부분은 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "화면에 평면적으로 합성된 듯한 텍스트 위반이 있으며, 요구된 클로즈업이 아닌 미디엄 샷으로 렌더링됨."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 타이트한 오버더숄더 클로즈업과 앉은 눈높이 아래의 카메라 앵글을 완벽하게 구현함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "중앙의 남성이 왼쪽 전경의 흐릿한 인물을 노려봄.",
      "built_space": "정비소 사무실 내부. 창문, 소파, 출입문이 배치됨.",
      "entities": "레퍼런스와 일치하는 30대 남성(얼굴, 작업복, 모자).",
      "hard_violations": [
       "출입문의 '사무실' 텍스트가 원근감과 조명을 무시하고 화면에 붙인 평면적인 스티커처럼 보임 (MATERIAL REALISM 위반)"
      ],
      "physics": "서 있는 자세이며 지면의 지지를 받음."
     },
     {
      "label": "B",
      "direction": "중앙의 남성이 왼쪽 전경의 흐릿한 인물을 쏘아봄.",
      "built_space": "정비소 사무실 내부. 창문 너머의 정비 공간과 출입문, 소파가 올바르게 위치함.",
      "entities": "레퍼런스와 일치하는 30대 남성(얼굴, 작업복, 모자).",
      "hard_violations": [],
      "physics": "의자에 앉아 있는 자세이며 물리적 오류 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "화면에 평면적으로 합성된 듯한 텍스트 위반이 있으며, 요구된 클로즈업이 아닌 미디엄 샷으로 렌더링됨."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 타이트한 오버더숄더 클로즈업과 앉은 눈높이 아래의 카메라 앵글을 완벽하게 구현함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "중앙의 남성이 왼쪽 전경의 흐릿한 인물을 노려봄.",
      "built_space": "정비소 사무실 내부. 창문, 소파, 출입문이 배치됨.",
      "entities": "레퍼런스와 일치하는 30대 남성(얼굴, 작업복, 모자).",
      "hard_violations": [
       "출입문의 '사무실' 텍스트가 원근감과 조명을 무시하고 화면에 붙인 평면적인 스티커처럼 보임 (MATERIAL REALISM 위반)"
      ],
      "physics": "서 있는 자세이며 지면의 지지를 받음."
     },
     {
      "label": "A",
      "direction": "중앙의 남성이 왼쪽 전경의 흐릿한 인물을 쏘아봄.",
      "built_space": "정비소 사무실 내부. 창문 너머의 정비 공간과 출입문, 소파가 올바르게 위치함.",
      "entities": "레퍼런스와 일치하는 30대 남성(얼굴, 작업복, 모자).",
      "hard_violations": [],
      "physics": "의자에 앉아 있는 자세이며 물리적 오류 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 7
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "오른쪽 인물의 시선이 왼쪽 전경에 흐릿하게 걸친 인물을 향해 정확히 고정됨.",
    "built_space": "패턴 벽지, 가죽 소파, 유리창 너머 정비소 등 로케이션 사진의 요소가 올바르게 배치됨.",
    "entities": "30대 남성의 얼굴과 복장이 레퍼런스와 일치하며 문 유리에 사무실 문구가 정확히 적힘.",
    "hard_violations": [],
    "physics": "어깨와 목의 긴장된 자세가 자연스럽게 지탱되며 물리적 오류 없음."
   },
   {
    "label": "B",
    "direction": "인물의 시선이 왼쪽 전경의 흐릿한 인물을 향함.",
    "built_space": "사무실 내부 배경 요소가 존재하나 문의 각도와 텍스트의 수평이 다소 어긋남.",
    "entities": "인물의 외모는 일치하나 표정이 밋밋하며 텍스트가 덧붙여진 느낌임.",
    "hard_violations": [],
    "physics": "서 있는 자세에 물리적으로 불가능한 부분은 없음."
   }
  ],
  "totals": {
   "A": 14,
   "B": 7
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 클로즈업 샷 크기를 잘 구현했으며 눈을 부릅뜬 표정과 사무실 텍스트 디테일이 프롬프트와 정확히 일치함."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "요구된 클로즈업이 아닌 미디엄 샷으로 프레이밍되었으며 인물의 표정에서 요구된 불쾌감과 부릅뜬 눈의 묘사가 부족함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_정비소_사무실_98ba69.png"
   },
   {
    "label": "CHARACTER REFERENCE — 30대 초반 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:762177>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "중앙 남성의 얼굴 생김새가 제공된 캐릭터 레퍼런스 이미지의 특정 인물과 일치하지 않습니다.",
     "fix_en": "Redraw the central man's face to exactly match the specific facial features and identity of the provided character reference. Keep his angry expression, gaze direction, blue cap, stained coveralls, the foreground man's shoulder, and all background elements including the door, cabinet, sofa, and window perfectly intact.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "'사무실' 텍스트가 지시문에서 요구한 유리 재질이 아닌, 불투명한 흰색 문 표면에 렌더링되었습니다.",
     "fix_en": "Remove the text '사무실' from the solid white door and render it on the office glass window instead. Keep the characters, their poses, clothing, expressions, and the rest of the room's geometry exactly as they are.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "배경 레퍼런스 원본에 있던 좌측 파란색 수납장 위의 전자레인지 및 소품들과, 우측 정수기의 상단 부속품이 누락 및 변형되었습니다.",
     "fix_en": "Restore the microwave and small items on the blue cabinet, and the missing top bottle on the water dispenser to exactly match the original background reference. Maintain the characters, their poses, clothing, and all other room fixtures unchanged.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "중앙 남성의 얼굴 생김새가 제공된 캐릭터 레퍼런스 이미지의 특정 인물과 일치하지 않습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "'사무실' 텍스트가 지시문에서 요구한 유리 재질이 아닌, 불투명한 흰색 문 표면에 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 레퍼런스 원본에 있던 좌측 파란색 수납장 위의 전자레인지 및 소품들과, 우측 정수기의 상단 부속품이 누락 및 변형되었습니다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "'사무실' 텍스트가 지시문에서 요구한 유리 재질이 아닌, 불투명한 흰색 문 표면에 렌더링되었습니다.",
    "fix_en": "Remove the text '사무실' from the solid white door and render it on the office glass window instead. Keep the characters, their poses, clothing, expressions, and the rest of the room's geometry exactly as they are.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "배경 레퍼런스 원본에 있던 좌측 파란색 수납장 위의 전자레인지 및 소품들과, 우측 정수기의 상단 부속품이 누락 및 변형되었습니다.",
    "fix_en": "Restore the microwave and small items on the blue cabinet, and the missing top bottle on the water dispenser to exactly match the original background reference. Maintain the characters, their poses, clothing, and all other room fixtures unchanged.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the central man's face to exactly match the specific facial features and identity of the provided character reference. Keep his angry expression, gaze direction, blue cap, stained coveralls, the foreground man's shoulder, and all background elements including the door, cabinet, sofa, and window perfectly intact.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트와 스케치가 요구한 타이트한 오버더숄더 클로즈업 구도를 정확히 구현했으며, 눈을 부릅뜬 굳은 표정과 배경의 텍스트('사무실') 지시를 훌륭하게 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업으로 제한된 프레이밍 지시를 어기고 피사체의 전신과 공간 전체를 보여주는 와이드 샷으로 렌더링하여 높은 우선순위의 샷 규격을 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "중앙 우측의 남성이 왼쪽 프레임 가장자리에 걸친 흐릿한 인물(서의용)을 강렬하게 노려보고 있음.",
      "built_space": "정비소 사무실 내부. 좁은 화각에 맞춰 정수기, 문, 유리창 등 배경 요소의 크기와 위치가 적절히 배치되었으며 문에 '사무실' 텍스트가 있음.",
      "entities": "30대 남성은 레퍼런스의 얼굴, 모자, 작업복과 완벽히 일치함. 왼쪽 프레임에 흐릿한 대화 상대방의 어깨와 머리 일부가 존재함.",
      "hard_violations": [],
      "physics": "어깨와 목에 힘이 들어간 상태로 하단에 앉아 자세를 유지하고 있음."
     },
     {
      "label": "B",
      "direction": "남성이 왼쪽 프레임의 흐릿한 인물을 향해 시선을 고정하고 있음.",
      "built_space": "정비소 사무실 내부. 요구된 샷보다 훨씬 넓게 렌더링되어 책상, 가죽 소파, 유리창 전체가 모두 노출됨.",
      "entities": "주요 인물의 얼굴과 복장은 레퍼런스와 일치하나, 요구된 '눈을 부릅뜬' 표정보다는 다소 무덤덤함. 좌측에 대화 상대방이 걸쳐 있음.",
      "hard_violations": [],
      "physics": "가죽 소파에 엉덩이를 대고 앉아 있으며, 두 발은 바닥을, 두 손은 무릎 위를 안정적으로 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트와 스케치가 요구한 타이트한 오버더숄더 클로즈업 구도를 정확히 구현했으며, 눈을 부릅뜬 굳은 표정과 배경의 텍스트('사무실') 지시를 훌륭하게 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업으로 제한된 프레이밍 지시를 어기고 피사체의 전신과 공간 전체를 보여주는 와이드 샷으로 렌더링하여 높은 우선순위의 샷 규격을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "중앙 우측의 남성이 왼쪽 프레임 가장자리에 걸친 흐릿한 인물(서의용)을 강렬하게 노려보고 있음.",
      "built_space": "정비소 사무실 내부. 좁은 화각에 맞춰 정수기, 문, 유리창 등 배경 요소의 크기와 위치가 적절히 배치되었으며 문에 '사무실' 텍스트가 있음.",
      "entities": "30대 남성은 레퍼런스의 얼굴, 모자, 작업복과 완벽히 일치함. 왼쪽 프레임에 흐릿한 대화 상대방의 어깨와 머리 일부가 존재함.",
      "hard_violations": [],
      "physics": "어깨와 목에 힘이 들어간 상태로 하단에 앉아 자세를 유지하고 있음."
     },
     {
      "label": "B",
      "direction": "남성이 왼쪽 프레임의 흐릿한 인물을 향해 시선을 고정하고 있음.",
      "built_space": "정비소 사무실 내부. 요구된 샷보다 훨씬 넓게 렌더링되어 책상, 가죽 소파, 유리창 전체가 모두 노출됨.",
      "entities": "주요 인물의 얼굴과 복장은 레퍼런스와 일치하나, 요구된 '눈을 부릅뜬' 표정보다는 다소 무덤덤함. 좌측에 대화 상대방이 걸쳐 있음.",
      "hard_violations": [],
      "physics": "가죽 소파에 엉덩이를 대고 앉아 있으며, 두 발은 바닥을, 두 손은 무릎 위를 안정적으로 지지하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 숏 크기(클로즈업)와 스케치를 무시하고 카메라를 뒤로 빼어 미디엄 숏으로 렌더링했으며, 요구된 출입문 텍스트가 누락되었습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "스케치에 제시된 타이트한 오버더숄더 클로즈업 구도를 정확히 구현하였고, 출입문 텍스트와 인물의 표정 및 정체성을 훌륭하게 반영했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "중앙 우측의 남성이 화면 좌측 전경에 흐릿하게 걸쳐진 남성을 정확히 응시하고 있습니다.",
      "built_space": "자동차 정비소 사무실 내부로, 책상, 소파, 창문 너머 정비 공간 등 배경 레퍼런스의 요소들이 제대로 배치되어 있습니다.",
      "entities": "30대 초반 남성의 외모, 작업복, 모자가 캐릭터 레퍼런스와 일치하나, 요구된 출입문 텍스트('사무실')가 없습니다.",
      "hard_violations": [
       "지정된 클로즈업 프레이밍 및 레이아웃 스케치를 무시하고 미디엄 숏으로 임의 확장함"
      ],
      "physics": "남성은 소파에 안정적으로 앉아 있으며 손은 무릎 위에 놓여 있어 물리적 지지에 이상이 없습니다."
     },
     {
      "label": "B",
      "direction": "남성이 눈을 부릅뜨고 화면 좌측 전경의 흐릿한 인물을 향해 시선을 고정하고 있습니다.",
      "built_space": "배경 레퍼런스의 사무실 공간을 클로즈업 구도에 맞게 잘랐으며, 뒤쪽 열린 문과 창문 너머 정비소가 알맞게 보입니다.",
      "entities": "캐릭터 레퍼런스의 외모와 복장을 정확히 재현했으며, 지정된 '사무실' 텍스트가 문에 올바르게 적용되었습니다.",
      "hard_violations": [],
      "physics": "프레임 하단에 인물의 몸통이 자연스럽게 자리 잡고 있어 자세와 지지에 어색함이 없습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 숏 크기(클로즈업)와 스케치를 무시하고 카메라를 뒤로 빼어 미디엄 숏으로 렌더링했으며, 요구된 출입문 텍스트가 누락되었습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "스케치에 제시된 타이트한 오버더숄더 클로즈업 구도를 정확히 구현하였고, 출입문 텍스트와 인물의 표정 및 정체성을 훌륭하게 반영했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "중앙 우측의 남성이 화면 좌측 전경에 흐릿하게 걸쳐진 남성을 정확히 응시하고 있습니다.",
      "built_space": "자동차 정비소 사무실 내부로, 책상, 소파, 창문 너머 정비 공간 등 배경 레퍼런스의 요소들이 제대로 배치되어 있습니다.",
      "entities": "30대 초반 남성의 외모, 작업복, 모자가 캐릭터 레퍼런스와 일치하나, 요구된 출입문 텍스트('사무실')가 없습니다.",
      "hard_violations": [
       "지정된 클로즈업 프레이밍 및 레이아웃 스케치를 무시하고 미디엄 숏으로 임의 확장함"
      ],
      "physics": "남성은 소파에 안정적으로 앉아 있으며 손은 무릎 위에 놓여 있어 물리적 지지에 이상이 없습니다."
     },
     {
      "label": "A",
      "direction": "남성이 눈을 부릅뜨고 화면 좌측 전경의 흐릿한 인물을 향해 시선을 고정하고 있습니다.",
      "built_space": "배경 레퍼런스의 사무실 공간을 클로즈업 구도에 맞게 잘랐으며, 뒤쪽 열린 문과 창문 너머 정비소가 알맞게 보입니다.",
      "entities": "캐릭터 레퍼런스의 외모와 복장을 정확히 재현했으며, 지정된 '사무실' 텍스트가 문에 올바르게 적용되었습니다.",
      "hard_violations": [],
      "physics": "프레임 하단에 인물의 몸통이 자연스럽게 자리 잡고 있어 자세와 지지에 어색함이 없습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S40sh5__bgfirst_bg.png",
   "bg_asset_id": "d46c4a56-1aba-49bc-a088-547a8f5b4993",
   "bg_record_key": "S40sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "정비소 사무실",
   "groupbg_asset_id": "5eee3cc6-afd0-4f3d-944b-c83aa86f5a4f"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S40sh5::cine": {
  "applied": true,
  "fingerprint": "aca4337cb9d33f2836b3e22ee00b4547020c4d8034f39e052d33a550c39b0ce1",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S40sh5_sel.png",
  "source_sha256": "97d9fe2ff98b6fcc90365ceefc423ae012fe32af5915b238b1477127dced7d6d",
  "file": "S40sh5_cine.png",
  "latency_ms": 10166
 },
 "S40sh8::signage": {
  "fp": "0f22d2fd295f0d7e",
  "inscriptions": [
   {
    "surface_native": "사무실 유리문",
    "text_native": "사무실",
    "reason_ko": "자동차 정비소의 사무실 구역임을 나타내는 표식으로, 공간의 현실감과 시대적 배경을 사실적으로 묘사하기 위해 필요합니다."
   }
  ]
 },
 "S40sh8": {
  "input_fingerprint": "aec23a853b17577c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밖으로 나간 30대 초반 한국인 남성의 뒷모습을 착잡한 눈빛으로 응시하는 서의용과 나상혁의 모습\n\nLOCATION (lock): Near the doorway inside the auto-shop office, looking out toward the departing man and repair area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just outside and to one side of the office doorway at chest height, hold a slightly widened oblique view: the departing Korean man crosses the left foreground with his back turned, while Seo Eui-yong and Na Sang-hyeok remain clearly readable inside in the right midground. Keep the camera static and let both investigators' troubled attention travel past the doorway toward his receding figure, with the office window and maintenance area providing depth behind them.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: departing man in the middle-left of the frame, foreground, moves toward space beyond the office doorway; Seo Eui-yong in the middle-center of the frame, midground, looks toward departing man; Na Sang-hyeok in the middle-right of the frame, midground, looks toward departing man.\n- KEY BACKGROUND ELEMENTS: office doorway (open) — Seen obliquely from its outer side, with the office visible through the opening; used as The doorway divides the departing foreground figure from the two investigators inside without fully separating the spaces; glass window frame (visible behind the investigators) — The maintenance facilities are visible beyond the glass; used as Adds rear depth behind the investigators and retains the car-center context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is rendered in a restrained neutral-to-cool palette with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 30대 초반 남성 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 유리문: \"사무실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밖으로 나간 30대 초반 한국인 남성의 뒷모습을 착잡한 눈빛으로 응시하는 서의용과 나상혁의 모습\n\nLOCATION (lock): Near the doorway inside the auto-shop office, looking out toward the departing man and repair area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just outside and to one side of the office doorway at chest height, hold a slightly widened oblique view: the departing Korean man crosses the left foreground with his back turned, while Seo Eui-yong and Na Sang-hyeok remain clearly readable inside in the right midground. Keep the camera static and let both investigators' troubled attention travel past the doorway toward his receding figure, with the office window and maintenance area providing depth behind them.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: departing man in the middle-left of the frame, foreground, moves toward space beyond the office doorway; Seo Eui-yong in the middle-center of the frame, midground, looks toward departing man; Na Sang-hyeok in the middle-right of the frame, midground, looks toward departing man.\n- KEY BACKGROUND ELEMENTS: office doorway (open) — Seen obliquely from its outer side, with the office visible through the opening; used as The doorway divides the departing foreground figure from the two investigators inside without fully separating the spaces; glass window frame (visible behind the investigators) — The maintenance facilities are visible beyond the glass; used as Adds rear depth behind the investigators and retains the car-center context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is rendered in a restrained neutral-to-cool palette with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 30대 초반 남성 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 유리문: \"사무실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 밖으로 나간 30대 초반 한국인 남성의 뒷모습을 착잡한 눈빛으로 응시하는 서의용과 나상혁의 모습\n\nLOCATION (lock): Near the doorway inside the auto-shop office, looking out toward the departing man and repair area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just outside and to one side of the office doorway at chest height, hold a slightly widened oblique view: the departing Korean man crosses the left foreground with his back turned, while Seo Eui-yong and Na Sang-hyeok remain clearly readable inside in the right midground. Keep the camera static and let both investigators' troubled attention travel past the doorway toward his receding figure, with the office window and maintenance area providing depth behind them.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: departing man in the middle-left of the frame, foreground, moves toward space beyond the office doorway; Seo Eui-yong in the middle-center of the frame, midground, looks toward departing man; Na Sang-hyeok in the middle-right of the frame, midground, looks toward departing man.\n- KEY BACKGROUND ELEMENTS: office doorway (open) — Seen obliquely from its outer side, with the office visible through the opening; used as The doorway divides the departing foreground figure from the two investigators inside without fully separating the spaces; glass window frame (visible behind the investigators) — The maintenance facilities are visible beyond the glass; used as Adds rear depth behind the investigators and retains the car-center context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is rendered in a restrained neutral-to-cool palette with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 30대 초반 남성 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 유리문: \"사무실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "gq": {
   "route": "combined",
   "gap": 0.667,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.083,
    "B": 1.0
   },
   "violations": {
    "A": [
     "[gemini-pro] 프레임 레이아웃 및 인물 식별 위반: 전경에 배치되어야 할 떠나는 30대 초반 남성과 중경에 머물러야 할 서의용의 위치와 역할이 서로 뒤바뀜."
    ],
    "B": [
     "[gemini-pro] 프레임 레이아웃 및 인물 식별 위반: 전경에 배치되어야 할 떠나는 30대 초반 남성과 중경에 머물러야 할 서의용의 위치와 역할이 서로 뒤바뀜.",
     "[gemini-pro] 금지된 텍스트 위반: 우측 유리창 파티션에 '행복 사무실' 등 프롬프트가 지시하지 않은 임의의 텍스트가 생성됨."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1083,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1083,
    "verdict_ko": "전경에서 떠나는 30대 남성과 중경에 있어야 할 서의용의 역할 및 위치가 완전히 뒤바뀌어 프레임 레이아웃 지시를 심각하게 위반했습니다.  ★위반: [gemini-pro] 프레임 레이아웃 및 인물 식별 위반: 전경에 배치되어야 할 떠나는 30대 초반 남성과 중경에 머물러야 할 서의용의 위치와 역할이 서로 뒤바뀜."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "인물들의 위치와 역할이 뒤바뀐 치명적인 오류가 동일하게 존재하며, 우측 배경에 허용되지 않은 임의의 텍스트까지 생성되었습니다.  ★위반: [gemini-pro] 프레임 레이아웃 및 인물 식별 위반: 전경에 배치되어야 할 떠나는 30대 초반 남성과 중경에 머물러야 할 서의용의 위치와 역할이 서로 뒤바뀜. / [gemini-pro] 금지된 텍스트 위반: 우측 유리창 파티션에 '행복 사무실' 등 프롬프트가 지시하지 않은 임의의 텍스트가 생성됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 30대 초반 남성 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S40sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   },
   {
    "label": "CHARACTER REFERENCE — 30대 초반 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:762177>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 왼쪽 전경에서 뒷모습으로 떠나는 인물이 30대 초반 남성(정비복)이 아닌 서의용(가죽 재킷, 비니)이며, 중앙에서 지켜보는 인물이 서의용이 아닌 30대 초반 남성으로 렌더링되어 프롬프트가 지시한 두 인물의 위치와 역할이 완전히 뒤바뀌었습니다.",
     "fix_en": "Change the clothing of the foreground left character from a leather jacket and beanie to blue mechanic coveralls and a blue cap, and replace the center character with Seo Eui-yong wearing a black leather jacket and blue beanie, preserving the right character, the camera framing, the doorway, and the background repair area.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 왼쪽 전경에서 뒷모습으로 떠나는 인물이 30대 초반 남성(정비복)이 아닌 서의용(가죽 재킷, 비니)이며, 중앙에서 지켜보는 인물이 서의용이 아닌 30대 초반 남성으로 렌더링되어 프롬프트가 지시한 두 인물의 위치와 역할이 완전히 뒤바뀌었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "떠나는 사람이 잠긴 작업복 남성이 아니라 가죽재킷과 비니를 입은 서의용 뒷모습이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "프레임 한가운데에 잠긴 작업복의 30대 초반 남성이 정면을 보고 서 있어 떠나는 뒷모습이 아니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "카메라가 문 바깥이 아니라 사무실 안쪽에서 바깥 정비장을 향해 찍혀 구도가 반대다.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용과 나상혁이 사무실 안 중경이 아니라 문 밖 정비장에 나란히 서 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "가운데 작업복 남성과 오른쪽 나상혁이 떠나는 뒷모습이 아니라 렌즈 쪽을 보고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷에 없던 회색 후드가 가죽재킷 안에서 밖으로 나와 있다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 6
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the clothing of the foreground left character from a leather jacket and beanie to blue mechanic coveralls and a blue cap, and replace the center character with Seo Eui-yong wearing a black leather jacket and blue beanie, preserving the right character, the camera framing, the doorway, and the background repair area.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시된 오버더숄더 구도를 시도했으나, 떠나는 인물과 지켜보는 인물의 역할이 완전히 뒤바뀌어 프롬프트의 핵심 연출을 위반했습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "지시된 연출(뒷모습, 사무실 문간)을 무시하고 정면으로 나열했으며, 우측 인물의 의상도 레퍼런스와 다릅니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "배경의 두 남성이 전경에 등을 돌린 남성을 바라보고 있음.",
      "built_space": "화면 좌측에 '사무실'이 적힌 열린 문이 있고, 뒤로 정비소 내부와 차량이 보임. 레퍼런스의 불투명한 철문이 유리문으로 변형됨.",
      "entities": "전경에 등 돌린 인물은 30대 남성(정비사)이어야 하나 서의용의 의상(가죽 재킷, 비니)을 입고 있음. 배경의 두 인물은 각각 정비사와 나상혁의 복장과 일치함. '사무실' 텍스트는 유리창에 바르게 렌더링됨.",
      "hard_violations": [
       "전경에서 떠나는 인물이 30대 남성(정비사)이 아닌 서의용으로 잘못 배치됨 (역할 뒤바뀜)",
       "레퍼런스의 고정된 공간 요소(불투명한 흰색 문, 벽지 등)가 유리문과 콘크리트 벽으로 변경됨"
      ],
      "physics": "세 사람 모두 바닥에 지지되어 서 있음."
     },
     {
      "label": "B",
      "direction": "세 명의 남성 모두 카메라 쪽을 정면으로 응시하고 있음.",
      "built_space": "사무실 문간이 생략되고 공구대와 셔터가 있는 정비소 작업장 내부로 완전히 대체됨.",
      "entities": "좌측 인물(정비사)과 중앙 인물(서의용)의 복장은 일치하나, 우측 인물은 나상혁의 복장(남색 재킷)이 아닌 체크 셔츠와 작업용 공구 벨트를 착용하고 있음.",
      "hard_violations": [
       "떠나는 뒷모습을 바라보는 연출이 전혀 반영되지 않음 (전원 정면 응시)",
       "지시된 공간(사무실 문간)을 렌더링하지 않음",
       "우측 인물의 의상이 나상혁의 레퍼런스와 다름 (체크 셔츠, 공구 벨트)"
      ],
      "physics": "세 사람 모두 바닥에 정상적으로 서 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시된 오버더숄더 구도를 시도했으나, 떠나는 인물과 지켜보는 인물의 역할이 완전히 뒤바뀌어 프롬프트의 핵심 연출을 위반했습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "지시된 연출(뒷모습, 사무실 문간)을 무시하고 정면으로 나열했으며, 우측 인물의 의상도 레퍼런스와 다릅니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "배경의 두 남성이 전경에 등을 돌린 남성을 바라보고 있음.",
      "built_space": "화면 좌측에 '사무실'이 적힌 열린 문이 있고, 뒤로 정비소 내부와 차량이 보임. 레퍼런스의 불투명한 철문이 유리문으로 변형됨.",
      "entities": "전경에 등 돌린 인물은 30대 남성(정비사)이어야 하나 서의용의 의상(가죽 재킷, 비니)을 입고 있음. 배경의 두 인물은 각각 정비사와 나상혁의 복장과 일치함. '사무실' 텍스트는 유리창에 바르게 렌더링됨.",
      "hard_violations": [
       "전경에서 떠나는 인물이 30대 남성(정비사)이 아닌 서의용으로 잘못 배치됨 (역할 뒤바뀜)",
       "레퍼런스의 고정된 공간 요소(불투명한 흰색 문, 벽지 등)가 유리문과 콘크리트 벽으로 변경됨"
      ],
      "physics": "세 사람 모두 바닥에 지지되어 서 있음."
     },
     {
      "label": "B",
      "direction": "세 명의 남성 모두 카메라 쪽을 정면으로 응시하고 있음.",
      "built_space": "사무실 문간이 생략되고 공구대와 셔터가 있는 정비소 작업장 내부로 완전히 대체됨.",
      "entities": "좌측 인물(정비사)과 중앙 인물(서의용)의 복장은 일치하나, 우측 인물은 나상혁의 복장(남색 재킷)이 아닌 체크 셔츠와 작업용 공구 벨트를 착용하고 있음.",
      "hard_violations": [
       "떠나는 뒷모습을 바라보는 연출이 전혀 반영되지 않음 (전원 정면 응시)",
       "지시된 공간(사무실 문간)을 렌더링하지 않음",
       "우측 인물의 의상이 나상혁의 레퍼런스와 다름 (체크 셔츠, 공구 벨트)"
      ],
      "physics": "세 사람 모두 바닥에 정상적으로 서 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 숄더뷰 구도를 시도했으나, 뒷모습으로 퇴장하는 인물과 지켜보는 인물의 역할(의상 및 신원)이 뒤바뀌었으며 인물들의 위치도 사무실 내부가 아닌 정비 구역으로 잘못 묘사되었습니다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "요구된 구도와 연출을 완전히 무시한 채 세 인물이 정면을 보고 서 있으며, 오른쪽 인물은 레퍼런스와 전혀 다른 새로운 인물로 창조되었습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "중경에 있는 두 남성이 전경 좌측에 등 돌린 남성을 바라보고 있음.",
      "built_space": "화면 좌측에 '사무실'이라고 적힌 열린 문이 있으며, 중경의 두 인물은 사무실 내부가 아닌 차량이 리프트에 올려진 정비 구역에 서 있음.",
      "entities": "전경 좌측(뒷모습): 서의용의 의상(가죽 재킷, 비니). 중경 중앙: 30대 초반 남성의 의상(파란색 모자, 정비복). 중경 우측: 나상혁(네이비 재킷, 셔츠).",
      "hard_violations": [
       "등을 돌리고 퇴장하는 인물과 지켜보는 인물의 역할 및 의상이 완전히 뒤바뀜 (프롬프트는 30대 초반 남성이 퇴장하고 서의용이 지켜봐야 함).",
       "서의용과 나상혁이 사무실 내부가 아닌 정비 구역에 서 있어 공간 연출 오류."
      ],
      "physics": "세 인물 모두 바닥에 안정적으로 서 있음."
     },
     {
      "label": "A",
      "direction": "세 명의 남성 모두 퇴장하지 않고 카메라 정면을 응시하고 있음.",
      "built_space": "차량 수리 도구와 리프트가 있는 정비소 내부이며, 프롬프트에서 요구한 사무실 문이나 창문은 프레임 내에 존재하지 않음.",
      "entities": "좌측: 30대 초반 남성(파란 모자, 정비복). 중앙: 서의용(비니, 가죽 재킷). 우측: 지시문과 무관한 새로운 인물(체크 셔츠).",
      "hard_violations": [
       "지시문에 없는 새로운 인물(오른쪽 체크 셔츠 남성)을 창조함.",
       "뒷모습으로 퇴장하는 인물이 없고 모두 정면을 바라보는 등 프롬프트의 구도와 연출을 완전히 무시함."
      ],
      "physics": "세 인물 모두 바닥에 안정적으로 서 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "요구된 숄더뷰 구도를 시도했으나, 뒷모습으로 퇴장하는 인물과 지켜보는 인물의 역할(의상 및 신원)이 뒤바뀌었으며 인물들의 위치도 사무실 내부가 아닌 정비 구역으로 잘못 묘사되었습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "요구된 구도와 연출을 완전히 무시한 채 세 인물이 정면을 보고 서 있으며, 오른쪽 인물은 레퍼런스와 전혀 다른 새로운 인물로 창조되었습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "중경에 있는 두 남성이 전경 좌측에 등 돌린 남성을 바라보고 있음.",
      "built_space": "화면 좌측에 '사무실'이라고 적힌 열린 문이 있으며, 중경의 두 인물은 사무실 내부가 아닌 차량이 리프트에 올려진 정비 구역에 서 있음.",
      "entities": "전경 좌측(뒷모습): 서의용의 의상(가죽 재킷, 비니). 중경 중앙: 30대 초반 남성의 의상(파란색 모자, 정비복). 중경 우측: 나상혁(네이비 재킷, 셔츠).",
      "hard_violations": [
       "등을 돌리고 퇴장하는 인물과 지켜보는 인물의 역할 및 의상이 완전히 뒤바뀜 (프롬프트는 30대 초반 남성이 퇴장하고 서의용이 지켜봐야 함).",
       "서의용과 나상혁이 사무실 내부가 아닌 정비 구역에 서 있어 공간 연출 오류."
      ],
      "physics": "세 인물 모두 바닥에 안정적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "세 명의 남성 모두 퇴장하지 않고 카메라 정면을 응시하고 있음.",
      "built_space": "차량 수리 도구와 리프트가 있는 정비소 내부이며, 프롬프트에서 요구한 사무실 문이나 창문은 프레임 내에 존재하지 않음.",
      "entities": "좌측: 30대 초반 남성(파란 모자, 정비복). 중앙: 서의용(비니, 가죽 재킷). 우측: 지시문과 무관한 새로운 인물(체크 셔츠).",
      "hard_violations": [
       "지시문에 없는 새로운 인물(오른쪽 체크 셔츠 남성)을 창조함.",
       "뒷모습으로 퇴장하는 인물이 없고 모두 정면을 바라보는 등 프롬프트의 구도와 연출을 완전히 무시함."
      ],
      "physics": "세 인물 모두 바닥에 안정적으로 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S40sh5"
  }
 },
 "S40sh8::cine": {
  "applied": true,
  "fingerprint": "7aca1f9e0b347a64435c583586686c2a85eccb18dc2ad0618bf1c16d0450c1ed",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S40sh8_sel.png",
  "source_sha256": "3b54e42dd89365cc2196a20e7d56e344ed68c48df3045e590f1d2725855f51c6",
  "file": "S40sh8_cine.png",
  "latency_ms": 11720
 },
 "S41sh1::signage": {
  "fp": "5a3aa5a7fea77c21",
  "inscriptions": [
   {
    "surface_native": "놀이터 안내판",
    "text_native": "어린이놀이터",
    "reason_ko": "2000년대 한국 동네 놀이터 배경의 사실성을 높이기 위해 전형적인 놀이터 표지판을 추가합니다."
   }
  ]
 },
 "era_assess::807bd1e6a368ede4": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 아파트 단지 놀이터",
    "search_terms_native": [
     "2000년대 아파트 놀이터",
     "조합놀이대",
     "동네 놀이터"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대와 2010년대 한국 아파트 단지 놀이터 특유의 알록달록한 조합놀이대 구조물과 바닥재 재질은 일반적인 이미지 모델이 정확히 묘사하기 어렵습니다."
   }
  ]
 },
 "era_ref::2a15119d1fd101b6": {
  "subject": "2000년대~2010년대 한국의 아파트 단지 놀이터",
  "terms": [
   "2000년대 아파트 놀이터",
   "조합놀이대",
   "동네 놀이터"
  ],
  "queries": [
   [
    "2000년대 한국 아파트 놀이터 조합놀이대",
    "2000년대 동네 놀이터 아파트 단지"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://file.kbland.kr/image/kbstar/land/img/alian/kms/complex/photo/objctidnfr/15140/MTUxNDAxNzE4OTY3Mg%3D%3D.jpg",
  "picked_reason_ko": "아파트 동 사이의 일상적인 단지 놀이터를 주제로 삼아 조합 놀이대의 구조, 비례, 재료, 색상과 바닥 포장을 가장 선명하고 온전하게 보여 준다.",
  "sha256": "6d5c94b229f17d3ab874da3226bf0d76821491fa884aa6ef893c5f0a7869d295",
  "file": "eraref_2a15119d1fd101b6.png"
 },
 "S41sh1::bgfirst_bg": {
  "input_fingerprint": "1162d4da274ea794",
  "prompt": "Turn the attached storyboard sketch into a photorealistic film still of\nits LOCATION ONLY, keeping the figures exactly as they are.\n\nKEEP EXACTLY: the camera framing, the horizon, and where every element\nsits in the frame. Each mannequin figure stays a plain grey featureless\nmannequin standing in the very same spot, at the same size, in the same\npose, turned the same way — do not turn them into people, do not move,\nrotate, mirror or re-pose them, do not add or remove figures.\n\nBUILD PHOTOREALISTICALLY: everything that is not a figure — ground,\nsurfacing, structures, vegetation, sky, water, distance. The bare lines\nof the sketch are a layout guide; replace them with the real materials,\ndepth and lighting of the place described below.\n\nSHOT TEXT this background must serve (Korean): 한 발을 들고 달리는 자세의 한국인 어린이들이 있는 놀이터를 배경으로 벤치에 홀로 앉아있는 이수정의 전경\n\nLOCATION (lock): At a bench overlooking the open neighborhood playground and its play area.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin above head height several meters from the bench, looking diagonally across Lee Soo-jung's entire seated figure toward the full playground in a high-angle wide composition. Place her alone in the lower foreground while three or four Korean children remain dispersed in the background, each caught at a different phase of running—different lifted feet, shoulder angles, and gaze directions—so their activity contrasts with her stillness without becoming a synchronized formation.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lee Soo-jung seated alone in the lower-left of the frame, foreground; children playing across the playground in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: bench (occupied by Lee Soo-jung) — Viewed diagonally, with Lee Soo-jung seated along its near-visible side; used as Provides the isolated foreground seat and the compositional base for Lee Soo-jung's full figure; playground (three or four children are playing); used as Extends across the background as the active setting around the isolated interview subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light maintains a subdued neutral-to-cool investigative-drama palette and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHINGS AT THIS PLACE:\nThe neighborhood playground is organized around a main outdoor play area. A fixed bench occupies the playground perimeter and faces inward, providing an overview across the play area. A pedestrian approach directly meets the perimeter and opens into the playground so the site remains connected to the surrounding neighborhood. The bench zone and the play area provide the two camera-capable portions of the site.\n- playground: neighborhood outdoor play area\n- approach: neighborhood pedestrian approach\n- bench: playground-view bench\n\nWORLD FACTS (creator-confirmed — always true):\n- Region (real-world reference): 대한민국 전라남도 및 광주광역시권\n- Era: 2001년의 과거 시점과 2015~2017년의 현대 시점이 병존하는 21세기 초·중반\n- 2001년 장면은 당시 한국의 의복, 차량, 전자기기, 웹 인터페이스와 영상 화질을 적용하고, 2015~2017년 장면은 해당 시기의 현대적 환경과 디지털 기기를 적용한다. 서로 다른 시기의 요소를 임의로 혼용하지 않는다.\n- 2001년에는 유선전화, 초기 휴대전화, 데스크톱 기반 채팅 사이트, 브라운관 또는 당대형 모니터, 당시 보급형 소형차를 사용한다. 2015~2017년에는 스마트폰, 노트북, 디지털 프레젠테이션, 도어락, 현대식 수사·법의학 장비를 사용하되 이후 세대의 기기나 인터페이스는 배제한다.\n\nTHINGS THAT LIVE AT THIS PLACE (what they are, not where they go): the\nlist above names what permanently belongs to this site. Where each one\nsits in this frame is decided by the attached sketch alone — do not add\nor reposition anything on the strength of the list, build only what the\nsketch already has lines for, and do not invent facilities that are not\nlisted.\n\nSignage and other writing that permanently belongs to this place may\nappear, in the place's native language and period style. No captions,\nwatermarks or overlay text, and none of the sketch's diagram symbols,\nlines or markers anywhere in the image.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 아파트 단지 놀이터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Turn the attached storyboard sketch into a photorealistic film still of\nits LOCATION ONLY, keeping the figures exactly as they are.\n\nKEEP EXACTLY: the camera framing, the horizon, and where every element\nsits in the frame. Each mannequin figure stays a plain grey featureless\nmannequin standing in the very same spot, at the same size, in the same\npose, turned the same way — do not turn them into people, do not move,\nrotate, mirror or re-pose them, do not add or remove figures.\n\nBUILD PHOTOREALISTICALLY: everything that is not a figure — ground,\nsurfacing, structures, vegetation, sky, water, distance. The bare lines\nof the sketch are a layout guide; replace them with the real materials,\ndepth and lighting of the place described below.\n\nSHOT TEXT this background must serve (Korean): 한 발을 들고 달리는 자세의 한국인 어린이들이 있는 놀이터를 배경으로 벤치에 홀로 앉아있는 이수정의 전경\n\nLOCATION (lock): At a bench overlooking the open neighborhood playground and its play area.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Begin above head height several meters from the bench, looking diagonally across Lee Soo-jung's entire seated figure toward the full playground in a high-angle wide composition. Place her alone in the lower foreground while three or four Korean children remain dispersed in the background, each caught at a different phase of running—different lifted feet, shoulder angles, and gaze directions—so their activity contrasts with her stillness without becoming a synchronized formation.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lee Soo-jung seated alone in the lower-left of the frame, foreground; children playing across the playground in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: bench (occupied by Lee Soo-jung) — Viewed diagonally, with Lee Soo-jung seated along its near-visible side; used as Provides the isolated foreground seat and the compositional base for Lee Soo-jung's full figure; playground (three or four children are playing); used as Extends across the background as the active setting around the isolated interview subject.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light maintains a subdued neutral-to-cool investigative-drama palette and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHINGS AT THIS PLACE:\nThe neighborhood playground is organized around a main outdoor play area. A fixed bench occupies the playground perimeter and faces inward, providing an overview across the play area. A pedestrian approach directly meets the perimeter and opens into the playground so the site remains connected to the surrounding neighborhood. The bench zone and the play area provide the two camera-capable portions of the site.\n- playground: neighborhood outdoor play area\n- approach: neighborhood pedestrian approach\n- bench: playground-view bench\n\nWORLD FACTS (creator-confirmed — always true):\n- Region (real-world reference): 대한민국 전라남도 및 광주광역시권\n- Era: 2001년의 과거 시점과 2015~2017년의 현대 시점이 병존하는 21세기 초·중반\n- 2001년 장면은 당시 한국의 의복, 차량, 전자기기, 웹 인터페이스와 영상 화질을 적용하고, 2015~2017년 장면은 해당 시기의 현대적 환경과 디지털 기기를 적용한다. 서로 다른 시기의 요소를 임의로 혼용하지 않는다.\n- 2001년에는 유선전화, 초기 휴대전화, 데스크톱 기반 채팅 사이트, 브라운관 또는 당대형 모니터, 당시 보급형 소형차를 사용한다. 2015~2017년에는 스마트폰, 노트북, 디지털 프레젠테이션, 도어락, 현대식 수사·법의학 장비를 사용하되 이후 세대의 기기나 인터페이스는 배제한다.\n\nTHINGS THAT LIVE AT THIS PLACE (what they are, not where they go): the\nlist above names what permanently belongs to this site. Where each one\nsits in this frame is decided by the attached sketch alone — do not add\nor reposition anything on the strength of the list, build only what the\nsketch already has lines for, and do not invent facilities that are not\nlisted.\n\nSignage and other writing that permanently belongs to this place may\nappear, in the place's native language and period style. No captions,\nwatermarks or overlay text, and none of the sketch's diagram symbols,\nlines or markers anywhere in the image.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 아파트 단지 놀이터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S41sh1__bgfirst_bg.png",
  "asset_id": "2713f373-4e1a-4cea-9b44-16ba3cd9ad3e",
  "input_asset_ids": [
   "e616ac9d-7433-4f9d-a42c-8ea477a3cf02"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국의 아파트 단지 놀이터",
   "queries": [
    [
     "2000년대 한국 아파트 놀이터 조합놀이대",
     "2000년대 동네 놀이터 아파트 단지"
    ]
   ],
   "picked_url": "https://file.kbland.kr/image/kbstar/land/img/alian/kms/complex/photo/objctidnfr/15140/MTUxNDAxNzE4OTY3Mg%3D%3D.jpg",
   "sha256": "6d5c94b229f17d3ab874da3226bf0d76821491fa884aa6ef893c5f0a7869d295",
   "file": "eraref_2a15119d1fd101b6.png"
  }
 },
 "S41sh1": {
  "input_fingerprint": "3f8703e10c21a49a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한 발을 들고 달리는 자세의 한국인 어린이들이 있는 놀이터를 배경으로 벤치에 홀로 앉아있는 이수정의 전경\n\nLOCATION (lock): At a bench overlooking the open neighborhood playground and its play area. The shot takes place here — the attached STORYBOARD SKETCH fixes the staging, camera and figure placement of this exact place. No location photograph is attached — build the location itself strictly from the location text above and the shot text, inventing nothing beyond them.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light maintains a subdued neutral-to-cool investigative-drama palette and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이수정 (현재) (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 놀이터 안내판: \"어린이놀이터\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Replace every grey or outlined mannequin figure in the FIRST attached\nimage with the real people the CHARACTER REFERENCE images show, and\noutput a photorealistic film still. Each mannequin becomes the character\nwhose POSE CANON and the shot text match that figure's pose and\nposition; never repeat one person across two mannequins. If no CHARACTER\nREFERENCE image is attached for a figure, still replace it with a\nplausible real person consistent with the text. Every mannequin becomes\na person — erasing a figure, or leaving one out, is not a replacement,\nand the number of figures never changes.\n\nKEEP EXACTLY: the background of the first image, the camera framing,\nand each mannequin's position, size, pose and the direction it is\nturned. A person must be exactly where their mannequin was, at the same\nscale, in the same pose, facing the same way — never mirrored, never\nre-staged, never straightened up, stood upright or otherwise re-posed.\nNo mannequin, grey figure or sketch line may remain anywhere in the\noutput.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한 발을 들고 달리는 자세의 한국인 어린이들이 있는 놀이터를 배경으로 벤치에 홀로 앉아있는 이수정의 전경\n\nLOCATION (lock): At a bench overlooking the open neighborhood playground and its play area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin above head height several meters from the bench, looking diagonally across Lee Soo-jung's entire seated figure toward the full playground in a high-angle wide composition. Place her alone in the lower foreground while three or four Korean children remain dispersed in the background, each caught at a different phase of running—different lifted feet, shoulder angles, and gaze directions—so their activity contrasts with her stillness without becoming a synchronized formation.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lee Soo-jung seated alone in the lower-left of the frame, foreground; children playing across the playground in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: bench (occupied by Lee Soo-jung) — Viewed diagonally, with Lee Soo-jung seated along its near-visible side; used as Provides the isolated foreground seat and the compositional base for Lee Soo-jung's full figure; playground (three or four children are playing); used as Extends across the background as the active setting around the isolated interview subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light maintains a subdued neutral-to-cool investigative-drama palette and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이수정 (현재) (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Replace every grey or outlined mannequin figure in the FIRST attached\nimage with the real people the CHARACTER REFERENCE images show, and\noutput a photorealistic film still. Each mannequin becomes the character\nwhose POSE CANON and the shot text match that figure's pose and\nposition; never repeat one person across two mannequins. If no CHARACTER\nREFERENCE image is attached for a figure, still replace it with a\nplausible real person consistent with the text. Every mannequin becomes\na person — erasing a figure, or leaving one out, is not a replacement,\nand the number of figures never changes.\n\nKEEP EXACTLY: the background of the first image, the camera framing,\nand each mannequin's position, size, pose and the direction it is\nturned. A person must be exactly where their mannequin was, at the same\nscale, in the same pose, facing the same way — never mirrored, never\nre-staged, never straightened up, stood upright or otherwise re-posed.\nNo mannequin, grey figure or sketch line may remain anywhere in the\noutput.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 한 발을 들고 달리는 자세의 한국인 어린이들이 있는 놀이터를 배경으로 벤치에 홀로 앉아있는 이수정의 전경\n\nLOCATION (lock): At a bench overlooking the open neighborhood playground and its play area. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin above head height several meters from the bench, looking diagonally across Lee Soo-jung's entire seated figure toward the full playground in a high-angle wide composition. Place her alone in the lower foreground while three or four Korean children remain dispersed in the background, each caught at a different phase of running—different lifted feet, shoulder angles, and gaze directions—so their activity contrasts with her stillness without becoming a synchronized formation.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Lee Soo-jung seated alone in the lower-left of the frame, foreground; children playing across the playground in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: bench (occupied by Lee Soo-jung) — Viewed diagonally, with Lee Soo-jung seated along its near-visible side; used as Provides the isolated foreground seat and the compositional base for Lee Soo-jung's full figure; playground (three or four children are playing); used as Extends across the background as the active setting around the isolated interview subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light maintains a subdued neutral-to-cool investigative-drama palette and naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이수정 (현재) (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S41sh1__bgfirst_bg.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이수정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:863505>"
    }
   ],
   "B": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S41sh1__bgfirst_bg.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이수정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:863505>"
    }
   ]
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.5,
    "B": 1.5
   },
   "violations": {
    "B": [
     "[gemini-pro] 어린이 3명이 지면과 닿지 않고 물리적 지지 없이 허공에 떠 있음"
    ],
    "A": [
     "[gemini-pro] 어린이 4명 모두 지면과 닿지 않고 물리적 지지 없이 허공에 떠 있음"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1500,
   "A": 1500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "벤치의 형태는 레퍼런스와 일치하나, 필수 안내판이 누락되었고 다수의 어린이가 허공에 떠 있는 치명적인 물리적 오류가 있습니다.  ★위반: [gemini-pro] 어린이 3명이 지면과 닿지 않고 물리적 지지 없이 허공에 떠 있음"
   },
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "필수 안내판 누락과 더불어 벤치의 형태가 다르고, 모든 어린이가 물리적 지지 없이 공중에 떠 있는 오류가 있습니다.  ★위반: [gemini-pro] 어린이 4명 모두 지면과 닿지 않고 물리적 지지 없이 허공에 떠 있음"
   }
  ],
  "refs": [
   {
    "label": "SHOT BACKGROUND",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S41sh1__bgfirst_bg.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이수정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:863505>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "전경 우측 벤치에 앉은 여성이 좌석에 정상적으로 앉아 있습니다. 이는 벤치 등받이 상단에 걸터앉아 좌석에 발을 디디고 있던 원본 마네킹의 정확한 위치와 자세를 완전히 변형한 것으로, 자세 유지 지시 위반입니다.",
     "fix_en": "Straighten the seated woman's back to sit upright on the seat as an in-place mitigation. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경 좌측의 빨간 옷을 입은 아이와 노란 옷을 입은 아이의 앞으로 뻗은 팔다리 방향이 원본 마네킹들의 팔다리 위치와 반대로(거울상처럼) 뒤바뀌어 있습니다.",
     "fix_en": "Redraw the red and yellow-jacketed children to swap their lifted legs to match the reference. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "모래밭 앞쪽 중앙의 노란 옷 아이가 마네킹과 반대로 왼쪽을 향해 달리고 있다",
     "fix_en": "Redraw the yellow-jacketed child to run towards the right instead of the left. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "왼쪽 빨간 옷 아이가 카메라를 바라보며 웃고 있어 무방비한 순간이 아니다",
     "fix_en": "Redraw the red-jacketed child's face to look forward into the run, not at the camera. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "벤치에 앉은 이수정의 손이 마네킹처럼 벤치 좌판이 아니라 무릎 쪽에 모여 있다",
     "fix_en": "Redraw the seated woman's hands to rest flat on the bench seat rather than clasped in her lap. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "minor",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "전경 우측 벤치에 앉은 여성이 좌석에 정상적으로 앉아 있습니다. 이는 벤치 등받이 상단에 걸터앉아 좌석에 발을 디디고 있던 원본 마네킹의 정확한 위치와 자세를 완전히 변형한 것으로, 자세 유지 지시 위반입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경 중앙에서 파란색 옷을 입은 아이가 왼쪽을 향해 달리고 있습니다. 이는 오른쪽을 향해 달리던 해당 위치의 마네킹 방향과 정반대입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경 좌측의 빨간 옷을 입은 아이와 노란 옷을 입은 아이의 앞으로 뻗은 팔다리 방향이 원본 마네킹들의 팔다리 위치와 반대로(거울상처럼) 뒤바뀌어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "모래밭 앞쪽 중앙의 노란 옷 아이가 마네킹과 반대로 왼쪽을 향해 달리고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 빨간 옷 아이가 카메라를 바라보며 웃고 있어 무방비한 순간이 아니다",
     "severity": "major"
    },
    {
     "issue_ko": "벤치에 앉은 이수정의 손이 마네킹처럼 벤치 좌판이 아니라 무릎 쪽에 모여 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "배경 좌측의 빨간 옷을 입은 아이와 노란 옷을 입은 아이의 앞으로 뻗은 팔다리 방향이 원본 마네킹들의 팔다리 위치와 반대로(거울상처럼) 뒤바뀌어 있습니다.",
    "fix_en": "Redraw the red and yellow-jacketed children to swap their lifted legs to match the reference. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "모래밭 앞쪽 중앙의 노란 옷 아이가 마네킹과 반대로 왼쪽을 향해 달리고 있다",
    "fix_en": "Redraw the yellow-jacketed child to run towards the right instead of the left. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "왼쪽 빨간 옷 아이가 카메라를 바라보며 웃고 있어 무방비한 순간이 아니다",
    "fix_en": "Redraw the red-jacketed child's face to look forward into the run, not at the camera. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "벤치에 앉은 이수정의 손이 마네킹처럼 벤치 좌판이 아니라 무릎 쪽에 모여 있다",
    "fix_en": "Redraw the seated woman's hands to rest flat on the bench seat rather than clasped in her lap. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "minor",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Straighten the seated woman's back to sit upright on the seat as an in-place mitigation. Preserve exactly the people present and their positions, their clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1625,
      "verdict_ko": "지시된 카메라 구도와 이수정의 의상/외모, 배경의 배치를 완벽하게 따랐으나, 4명의 어린이가 모두 지지대 없이 공중에 떠 있는 물리적 오류가 발생하여 치명적인 감점 요인이 되었습니다.  ★위반: [gemini-pro] 4명의 어린이 모두 발이 지면에 닿지 않은 채 지지하는 것 없이 공중에 떠 있음"
     },
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "후보 A와 동일하게 지시사항과 캐릭터 레퍼런스를 잘 반영하였으나, 동일하게 4명의 어린이가 지면에서 떠 있어 물리적 한계를 극복하지 못한 치명적 오류가 존재합니다.  ★위반: [gemini-pro] 4명의 어린이 모두 발이 지면에 닿지 않은 채 지지하는 것 없이 공중에 떠 있음"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.875,
      "B": 2.0
     },
     "adjusted": {
      "A": 1.625,
      "B": 1.75
     },
     "violations": {
      "A": [
       "[gemini-pro] 4명의 어린이 모두 발이 지면에 닿지 않은 채 지지하는 것 없이 공중에 떠 있음"
      ],
      "B": [
       "[gemini-pro] 4명의 어린이 모두 발이 지면에 닿지 않은 채 지지하는 것 없이 공중에 떠 있음"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1625,
      "verdict_ko": "지시된 카메라 구도와 이수정의 의상/외모, 배경의 배치를 완벽하게 따랐으나, 4명의 어린이가 모두 지지대 없이 공중에 떠 있는 물리적 오류가 발생하여 치명적인 감점 요인이 되었습니다.  ★위반: [gemini-pro] 4명의 어린이 모두 발이 지면에 닿지 않은 채 지지하는 것 없이 공중에 떠 있음"
     },
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "후보 A와 동일하게 지시사항과 캐릭터 레퍼런스를 잘 반영하였으나, 동일하게 4명의 어린이가 지면에서 떠 있어 물리적 한계를 극복하지 못한 치명적 오류가 존재합니다.  ★위반: [gemini-pro] 4명의 어린이 모두 발이 지면에 닿지 않은 채 지지하는 것 없이 공중에 떠 있음"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "어린이들이 허공에 뜨는 문제를 해결하는 과정에서 두 어린이(노란 옷, 초록 옷)의 딛고 있는 다리를 마네킹과 다르게 변경하는 포즈 오류가 있었으나, 지지 기반 없는 물리적 위반을 피했으므로 전체적으로 더 나은 결과물입니다."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "어린이들의 다리 자세를 마네킹과 동일하게 유지하려 했으나, 4명 모두 발이 바닥에 닿지 않은 채 허공에 떠 있는 심각한 물리 법칙 위반(Hard Violation)이 발생했습니다.  ★위반: [gemini-pro] 배경의 어린이 4명이 모두 바닥에 발을 딛지 않고 허공에 뜬 채 그림자와 분리되어 있음 (물리적 지지 없음)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.429,
      "B": 1.857
     },
     "adjusted": {
      "A": 1.179,
      "B": 1.857
     },
     "violations": {
      "A": [
       "[gemini-pro] 배경의 어린이 4명이 모두 바닥에 발을 딛지 않고 허공에 뜬 채 그림자와 분리되어 있음 (물리적 지지 없음)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.143,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "어린이들이 허공에 뜨는 문제를 해결하는 과정에서 두 어린이(노란 옷, 초록 옷)의 딛고 있는 다리를 마네킹과 다르게 변경하는 포즈 오류가 있었으나, 지지 기반 없는 물리적 위반을 피했으므로 전체적으로 더 나은 결과물입니다."
     },
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "어린이들의 다리 자세를 마네킹과 동일하게 유지하려 했으나, 4명 모두 발이 바닥에 닿지 않은 채 허공에 떠 있는 심각한 물리 법칙 위반(Hard Violation)이 발생했습니다.  ★위반: [gemini-pro] 배경의 어린이 4명이 모두 바닥에 발을 딛지 않고 허공에 뜬 채 그림자와 분리되어 있음 (물리적 지지 없음)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 3482,
     "B": 2929
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S41sh1__bgfirst_bg.png",
   "bg_asset_id": "2713f373-4e1a-4cea-9b44-16ba3cd9ad3e",
   "bg_record_key": "S41sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "lane_conti_only"
  },
  "ref_mode": "lane(map_marker): 스케치+엔티티",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S41sh1::cine": {
  "applied": true,
  "fingerprint": "1b4b7e5943524a9d8df4de875315b55e9f1241c0c19bd47fa4600bddfa51b1b0",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S41sh1_sel.png",
  "source_sha256": "388b2c2598eddf3cb765406a0e4771cc97946532f1b5bda878399bb644523035",
  "file": "S41sh1_cine.png",
  "latency_ms": 10899
 },
 "S41sh2::signage": {
  "fp": "245c05a3e1e9ef95",
  "inscriptions": [
   {
    "surface_native": "놀이터 안내판",
    "text_native": "이용수칙",
    "reason_ko": "동네 놀이터라는 공간적 배경을 사실적으로 묘사하기 위해 배경의 표지판에 흔히 쓰이는 문구를 표시합니다."
   }
  ]
 },
 "S41sh2": {
  "input_fingerprint": "0bca478aa60d0578",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이수정의 맞은편에 앉아 손에 수첩을 쥔 서의용과 나상혁의 상체\n\nLOCATION (lock): At the facing seats beside a bench in the neighborhood playground. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The descending advance settles at seated shoulder height just behind and slightly beside Lee Soo-jung, using her near shoulder as a restrained foreground edge while Seo Eui-yong and Na Sang-hyeok sit together across the bench axis in a medium composition. Eui-yong leans toward her with his notebook held ready, while Sang-hyeok dips over his own notes; both upper bodies and notebooks remain visible without breaking the conversational axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bench seating area (occupied by the interview group) — Seen from behind Lee Soo-jung toward the investigators seated opposite her; used as Defines the seated interview axis and keeps the three participants physically connected in one space; playground (children continue playing in the background); used as Remains softly readable beyond the investigators to preserve the public setting.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light stays subdued and neutral-to-cool, with faces held in a naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same neighborhood playground, bench placement, playing children, and daylight. Exclude the woman-alone composition and place the two investigators opposite her with notebooks.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong and Sanghyeok remain seated opposite Sujeong with their interview notebooks in hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용과 나상혁 right now — 서의용과 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용과 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 놀이터 안내판: \"이용수칙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이수정의 맞은편에 앉아 손에 수첩을 쥔 서의용과 나상혁의 상체\n\nLOCATION (lock): At the facing seats beside a bench in the neighborhood playground. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The descending advance settles at seated shoulder height just behind and slightly beside Lee Soo-jung, using her near shoulder as a restrained foreground edge while Seo Eui-yong and Na Sang-hyeok sit together across the bench axis in a medium composition. Eui-yong leans toward her with his notebook held ready, while Sang-hyeok dips over his own notes; both upper bodies and notebooks remain visible without breaking the conversational axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bench seating area (occupied by the interview group) — Seen from behind Lee Soo-jung toward the investigators seated opposite her; used as Defines the seated interview axis and keeps the three participants physically connected in one space; playground (children continue playing in the background); used as Remains softly readable beyond the investigators to preserve the public setting.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light stays subdued and neutral-to-cool, with faces held in a naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same neighborhood playground, bench placement, playing children, and daylight. Exclude the woman-alone composition and place the two investigators opposite her with notebooks.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong and Sanghyeok remain seated opposite Sujeong with their interview notebooks in hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용과 나상혁 right now — 서의용과 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용과 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 놀이터 안내판: \"이용수칙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이수정의 맞은편에 앉아 손에 수첩을 쥔 서의용과 나상혁의 상체\n\nLOCATION (lock): At the facing seats beside a bench in the neighborhood playground. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The descending advance settles at seated shoulder height just behind and slightly beside Lee Soo-jung, using her near shoulder as a restrained foreground edge while Seo Eui-yong and Na Sang-hyeok sit together across the bench axis in a medium composition. Eui-yong leans toward her with his notebook held ready, while Sang-hyeok dips over his own notes; both upper bodies and notebooks remain visible without breaking the conversational axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bench seating area (occupied by the interview group) — Seen from behind Lee Soo-jung toward the investigators seated opposite her; used as Defines the seated interview axis and keeps the three participants physically connected in one space; playground (children continue playing in the background); used as Remains softly readable beyond the investigators to preserve the public setting.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light stays subdued and neutral-to-cool, with faces held in a naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same neighborhood playground, bench placement, playing children, and daylight. Exclude the woman-alone composition and place the two investigators opposite her with notebooks.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong and Sanghyeok remain seated opposite Sujeong with their interview notebooks in hand.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용과 나상혁 right now — 서의용과 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용과 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 놀이터 안내판: \"이용수칙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "서의용은 전경의 이수정을 향해 시선을 두고 있으며, 나상혁은 자신의 수첩을 내려다보고 있음.",
    "built_space": "벤치 뒤로 모래사장과 그네, 미끄럼틀이 레퍼런스와 일치하게 배치되어 있으며, 안내판은 얕은 나무 기둥에 세워져 있음.",
    "entities": "서의용과 나상혁의 인물 특징과 의상이 레퍼런스와 일치함. 두 사람 모두 수첩과 펜을 들고 있음. 배경에 달리는 아이들이 존재하며, '이용수칙' 텍스트가 정확함.",
    "hard_violations": [],
    "physics": "두 인물은 벤치에 안정적으로 앉아 수첩을 손으로 받치고 있으며, 배경의 아이들은 달리는 자세의 추진력에 의해 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "서의용은 전경의 이수정 쪽을 향해 시선을 두고, 나상혁은 수첩을 내려다보고 있음.",
    "built_space": "벤치 뒤로 놀이터 구조가 보이나, 레퍼런스에 없는 거대한 벽돌 기둥이 큰 비중으로 자리잡고 있음.",
    "entities": "서의용과 나상혁의 특징과 의상이 일치하며 수첩을 들고 있음. 그러나 프롬프트에서 요구한 배경의 아이들이 누락됨. '이용수칙' 텍스트는 정확함.",
    "hard_violations": [
     "발명된 오브젝트 (레퍼런스에 없는 거대한 벽돌 기둥)"
    ],
    "physics": "인물들은 벤치에 앉아 수첩을 손으로 쥐고 있으며 물리적 지지가 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "프롬프트가 요구한 배경의 뛰어노는 아이들을 포함하였으며, 레퍼런스의 놀이터 구조와 환경을 충실히 유지했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "요청된 배경의 아이들이 누락되었고, 레퍼런스에 존재하지 않는 거대한 벽돌 기둥이 임의로 추가되어 공간의 일관성을 훼손했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S41sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "인물들이 앉아 있는 벤치의 방향이 레퍼런스(모래사장을 향함)와 달리 모래사장을 등지도록 180도 뒤집혀 배치됨.",
     "fix_en": "Redraw the visible bench details to loosely match the reference style. Do not change Seo Eui-yong, Na Sang-hyeok, their positions, their clothing, the set, the light, or the framing.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "인물들이 앉은 벤치의 디자인(곡선형 철제 팔걸이가 없음)이 레퍼런스 이미지의 벤치와 다름.",
     "fix_en": "Add curved iron armrests to the bench to match the reference. Do not change Seo Eui-yong, Na Sang-hyeok, their positions, their clothing, the set, the light, or the framing.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "서의용이 이수정 쪽으로 상체를 기울이며 수첩을 준비해 쥐지 않고 허리를 세운 채 필기하고 있다",
     "fix_en": "Redraw Seo Eui-yong to lean forward with his notebook held ready instead of actively writing. Do not change Na Sang-hyeok, his position, their clothing, the set, the light, or the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "전경이 절제된 어깨 가장자리가 아니라 이수정 모자와 뒷머리가 화면 왼쪽을 크게 가리고 있다",
     "fix_en": "Redraw the left foreground to replace the hat and head with background elements, leaving only a restrained shoulder edge. Do not change Seo Eui-yong, Na Sang-hyeok, their positions, their clothing, the set, the light, or the framing.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "인물들이 앉아 있는 벤치의 방향이 레퍼런스(모래사장을 향함)와 달리 모래사장을 등지도록 180도 뒤집혀 배치됨.",
     "severity": "major"
    },
    {
     "issue_ko": "인물들이 앉은 벤치의 디자인(곡선형 철제 팔걸이가 없음)이 레퍼런스 이미지의 벤치와 다름.",
     "severity": "minor"
    },
    {
     "issue_ko": "서의용이 이수정 쪽으로 상체를 기울이며 수첩을 준비해 쥐지 않고 허리를 세운 채 필기하고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "전경이 절제된 어깨 가장자리가 아니라 이수정 모자와 뒷머리가 화면 왼쪽을 크게 가리고 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "인물들이 앉아 있는 벤치의 방향이 레퍼런스(모래사장을 향함)와 달리 모래사장을 등지도록 180도 뒤집혀 배치됨.",
    "fix_en": "Redraw the visible bench details to loosely match the reference style. Do not change Seo Eui-yong, Na Sang-hyeok, their positions, their clothing, the set, the light, or the framing.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "인물들이 앉은 벤치의 디자인(곡선형 철제 팔걸이가 없음)이 레퍼런스 이미지의 벤치와 다름.",
    "fix_en": "Add curved iron armrests to the bench to match the reference. Do not change Seo Eui-yong, Na Sang-hyeok, their positions, their clothing, the set, the light, or the framing.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "서의용이 이수정 쪽으로 상체를 기울이며 수첩을 준비해 쥐지 않고 허리를 세운 채 필기하고 있다",
    "fix_en": "Redraw Seo Eui-yong to lean forward with his notebook held ready instead of actively writing. Do not change Na Sang-hyeok, his position, their clothing, the set, the light, or the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "전경이 절제된 어깨 가장자리가 아니라 이수정 모자와 뒷머리가 화면 왼쪽을 크게 가리고 있다",
    "fix_en": "Redraw the left foreground to replace the hat and head with background elements, leaving only a restrained shoulder edge. Do not change Seo Eui-yong, Na Sang-hyeok, their positions, their clothing, the set, the light, or the framing.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh1"
  }
 },
 "S41sh2::cine": {
  "applied": true,
  "fingerprint": "36e34f2a5190a41db68f0a3d371e93d4dfe3c733b24457f02c90550bf5ebc75f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S41sh2_sel.png",
  "source_sha256": "63f90d4da1771454ade43180075836275ed154ea48c3e7ba8a7761de5c847c55",
  "file": "S41sh2_cine.png",
  "latency_ms": 10822
 },
 "S41sh7::signage": {
  "fp": "1049c70331612f14",
  "inscriptions": []
 },
 "S41sh7": {
  "input_fingerprint": "ff4850e19f224fca",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무언가 떠오른 듯 눈을 크게 뜬 이수정의 굳은 표정\n\nLOCATION (lock): On the bench overlooking the neighborhood playground. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the push slightly above Lee Soo-jung's seated eye line on her three-quarter side, tightening to her rigid face while leaving a narrow strip of look space beyond her widened eyes. Her shoulders stop mid-gesture and her gaze fixes past the investigators into empty space, allowing the remembered detail to register through stillness rather than visual distortion.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Lee Soo-jung isolated in close framing in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: playground background (out of focus); used as Reduced to a soft contextual trace behind her so the realization remains grounded in the same location.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light keeps the close-up restrained, neutral-to-cool, and moderately contrasted without signaling a literal flashback.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same playground, bench, daylight, and continuing children in the background. Exclude the two investigators from the close frame and isolate the seated woman's sudden realization.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이수정 (현재) (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무언가 떠오른 듯 눈을 크게 뜬 이수정의 굳은 표정\n\nLOCATION (lock): On the bench overlooking the neighborhood playground. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the push slightly above Lee Soo-jung's seated eye line on her three-quarter side, tightening to her rigid face while leaving a narrow strip of look space beyond her widened eyes. Her shoulders stop mid-gesture and her gaze fixes past the investigators into empty space, allowing the remembered detail to register through stillness rather than visual distortion.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Lee Soo-jung isolated in close framing in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: playground background (out of focus); used as Reduced to a soft contextual trace behind her so the realization remains grounded in the same location.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light keeps the close-up restrained, neutral-to-cool, and moderately contrasted without signaling a literal flashback.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same playground, bench, daylight, and continuing children in the background. Exclude the two investigators from the close frame and isolate the seated woman's sudden realization.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이수정 (현재) (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 무언가 떠오른 듯 눈을 크게 뜬 이수정의 굳은 표정\n\nLOCATION (lock): On the bench overlooking the neighborhood playground. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the push slightly above Lee Soo-jung's seated eye line on her three-quarter side, tightening to her rigid face while leaving a narrow strip of look space beyond her widened eyes. Her shoulders stop mid-gesture and her gaze fixes past the investigators into empty space, allowing the remembered detail to register through stillness rather than visual distortion.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Lee Soo-jung isolated in close framing in the middle-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: playground background (out of focus); used as Reduced to a soft contextual trace behind her so the realization remains grounded in the same location.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light keeps the close-up restrained, neutral-to-cool, and moderately contrasted without signaling a literal flashback.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same playground, bench, daylight, and continuing children in the background. Exclude the two investigators from the close frame and isolate the seated woman's sudden realization.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이수정 (현재) (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "여성은 화면 오른쪽을 향해 시선을 고정하고 있음.",
    "built_space": "야외 놀이터 벤치에 앉아 있으며 뒤편에 놀이기구와 안내판이 배치됨.",
    "entities": "이수정(레퍼런스와 외모 및 복장 일치), 배경에서 뛰고 있는 4명의 아이들.",
    "hard_violations": [],
    "physics": "여성은 벤치에 안정적으로 앉아 있고 배경의 아이들은 지면을 딛거나 달리는 도약 동작을 보여줌."
   },
   {
    "label": "B",
    "direction": "여성은 화면 오른쪽의 빈 공간을 응시하며 굳은 표정을 짓고 있음.",
    "built_space": "놀이터 벤치에 앉아 있고 뒤편으로 미끄럼틀과 그네, 안내판이 일관되게 위치함.",
    "entities": "이수정(레퍼런스와 일치하는 모자와 코트 착용), 배경에서 달리는 4명의 아이들.",
    "hard_violations": [],
    "physics": "여성의 몸이 벤치에 닿아 지지되고 있으며 배경의 아이들은 지면에 발을 딛고 뛰는 자세를 유지함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 6,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "클로즈업 샷 요구에 완벽히 부합하지는 않으나 A보다 피사체에 가깝게 프레이밍되어 신체 노출을 줄였다는 점에서 우선순위 규칙을 더 잘 따랐습니다."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "클로즈업을 요구한 프롬프트와 달리 허벅지까지 보이는 미디엄 샷으로 렌더링되어 프레이밍 스케일 지시를 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S41sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이수정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:863505>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 얼굴에 집중된 '클로즈업(close-up)' 샷을 지시했으나, 인물의 허리와 손까지 모두 보이는 미디엄 샷으로 연출되었습니다.",
     "fix_en": "Crop the frame tightly to the woman's head and shoulders. Maintain the woman's identity, attire, lighting, and the background elements.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "인물을 프레임의 '중앙 좌측(middle-left)'에 배치하고 시선 방향에 좁은 여백을 두라는 지시와 달리, 인물이 화면 중앙에서 우측에 치우쳐 배치되었습니다.",
     "fix_en": "Shift the woman to the middle-left of the frame. Maintain her identity, pose, attire, lighting, and the background elements.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "3/4 측면이 아니라 상체가 렌즈에 거의 정면으로 단정히 향해 포즈처럼 앉아 있다",
     "fix_en": "Rotate the woman's torso to a 3/4 side angle. Maintain her face, attire, lighting, and the background playground.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경 놀이터와 아이들이 부드러운 아웃포커스 흔적이 아니라 좌우에서 뚜렷하게 보인다",
     "fix_en": "Blur the background playground and children significantly. Maintain the woman's sharpness, attire, face, bench, and lighting.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": false
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 얼굴에 집중된 '클로즈업(close-up)' 샷을 지시했으나, 인물의 허리와 손까지 모두 보이는 미디엄 샷으로 연출되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "인물을 프레임의 '중앙 좌측(middle-left)'에 배치하고 시선 방향에 좁은 여백을 두라는 지시와 달리, 인물이 화면 중앙에서 우측에 치우쳐 배치되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴에 조여진 클로즈업이 아니라 상반신과 벤치·놀이터가 넓게 보이는 미디엄 샷이다",
     "severity": "major"
    },
    {
     "issue_ko": "3/4 측면이 아니라 상체가 렌즈에 거의 정면으로 단정히 향해 포즈처럼 앉아 있다",
     "severity": "major"
    },
    {
     "issue_ko": "배경 놀이터와 아이들이 부드러운 아웃포커스 흔적이 아니라 좌우에서 뚜렷하게 보인다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 얼굴에 집중된 '클로즈업(close-up)' 샷을 지시했으나, 인물의 허리와 손까지 모두 보이는 미디엄 샷으로 연출되었습니다.",
    "fix_en": "Crop the frame tightly to the woman's head and shoulders. Maintain the woman's identity, attire, lighting, and the background elements.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "인물을 프레임의 '중앙 좌측(middle-left)'에 배치하고 시선 방향에 좁은 여백을 두라는 지시와 달리, 인물이 화면 중앙에서 우측에 치우쳐 배치되었습니다.",
    "fix_en": "Shift the woman to the middle-left of the frame. Maintain her identity, pose, attire, lighting, and the background elements.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "3/4 측면이 아니라 상체가 렌즈에 거의 정면으로 단정히 향해 포즈처럼 앉아 있다",
    "fix_en": "Rotate the woman's torso to a 3/4 side angle. Maintain her face, attire, lighting, and the background playground.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경 놀이터와 아이들이 부드러운 아웃포커스 흔적이 아니라 좌우에서 뚜렷하게 보인다",
    "fix_en": "Blur the background playground and children significantly. Maintain the woman's sharpness, attire, face, bench, and lighting.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": false
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S41sh2"
  }
 },
 "S41sh7::cine": {
  "applied": true,
  "fingerprint": "b18151a7e364c8edfe532e6777faa98bd145dc18faa8bcfb57a3c8134ef7ac68",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S41sh7_sel.png",
  "source_sha256": "80aa979b69c16d71f7043075561094c4df8dbe465d34c045d44381426b188ef8",
  "file": "S41sh7_cine.png",
  "latency_ms": 11918
 },
 "S42sh3::signage": {
  "fp": "5c273073b56eecf4",
  "inscriptions": [
   {
    "surface_native": "초인종 버튼",
    "text_native": "호출",
    "reason_ko": "아파트 초인종 버튼에 흔히 인쇄되어 있는 '호출' 문구를 재현하여 현실감을 높이기 위함입니다."
   }
  ]
 },
 "era_assess::a6118298eb1decee": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 복도식 아파트 현관문 및 초인종",
    "search_terms_native": [
     "복도식 아파트 복도",
     "아파트 현관 방화문 디지털도어락",
     "아파트 초인종 인터폰"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성되어야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 초중반 한국 아파트의 특징적인 복도 구조, 철제 현관문, 디지털 도어락 및 초인종의 형태는 서구식 디자인과 크게 달라 고증이 필수적입니다."
   }
  ]
 },
 "era_ref::37d2273783207e96": {
  "subject": "2000년대 한국 복도식 아파트 현관문 및 초인종",
  "terms": [
   "복도식 아파트 복도",
   "아파트 현관 방화문 디지털도어락",
   "아파트 초인종 인터폰"
  ],
  "queries": [
   [
    "2000년대 한국 복도식 아파트 복도 현관 방화문 디지털도어락 초인종 인터폰",
    "2000년대 한국 아파트 현관문 초인종 인터폰 복도식 아파트"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://static.cdn.soomgo.com/upload/portfolio/5d392228-4850-4d9a-aff2-faa3d93020db.jpg?webp=1",
  "picked_reason_ko": "3번은 한국 복도식 아파트의 개별 세대 현관문과 벽면 초인종을 정면에서 선명하게 보여 주어 문체, 잠금장치, 마감재와 설치 관계를 가장 잘 읽을 수 있다.",
  "sha256": "a40fe776dccc3ed67baf285af26fb919277cb8034f155b743bc897e2d5053d5d",
  "file": "eraref_37d2273783207e96.png"
 },
 "S42sh3::bgfirst_bg": {
  "input_fingerprint": "8243c6c181bc9aa0",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용의 손과 초인종 버튼 사이를 막아서듯 덮은 나상혁의 손바닥 클로즈업\n\nLOCATION (lock): In the apartment building's long access corridor, directly outside the family's front door and doorbell.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height beside the apartment door controls, push into an extreme-close oblique insert across Na Sang-hyeok's open palm toward Seo Eui-yong's approaching fingers and the obscured doorbell behind it. Sang-hyeok's hand lies flat as a barrier in the center foreground while Eui-yong's fingers enter from the opposite edge, preserving enough of the door-control surround to make the intended press physically clear.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Na Sang-hyeok blocking hand in the middle-center of the frame, foreground; Seo Eui-yong pressing hand in the middle-left of the frame, foreground, reaches for covered doorbell button; doorbell button behind the palm in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (covered by Na Sang-hyeok's hand) — Its front face is angled toward the camera but partly hidden by the palm; used as The button is the blocked focal anchor immediately behind Sang-hyeok's palm; apartment door controls (visible around the contact area) — Their control-facing side is seen obliquely beside the hands; used as Provides a narrow scale and location reference around the hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime corridor illumination is treated neutrally with restrained coolness and documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.\n\nPERIOD REFERENCE — 2000년대 한국 복도식 아파트 현관문 및 초인종: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용의 손과 초인종 버튼 사이를 막아서듯 덮은 나상혁의 손바닥 클로즈업\n\nLOCATION (lock): In the apartment building's long access corridor, directly outside the family's front door and doorbell.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height beside the apartment door controls, push into an extreme-close oblique insert across Na Sang-hyeok's open palm toward Seo Eui-yong's approaching fingers and the obscured doorbell behind it. Sang-hyeok's hand lies flat as a barrier in the center foreground while Eui-yong's fingers enter from the opposite edge, preserving enough of the door-control surround to make the intended press physically clear.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Na Sang-hyeok blocking hand in the middle-center of the frame, foreground; Seo Eui-yong pressing hand in the middle-left of the frame, foreground, reaches for covered doorbell button; doorbell button behind the palm in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (covered by Na Sang-hyeok's hand) — Its front face is angled toward the camera but partly hidden by the palm; used as The button is the blocked focal anchor immediately behind Sang-hyeok's palm; apartment door controls (visible around the contact area) — Their control-facing side is seen obliquely beside the hands; used as Provides a narrow scale and location reference around the hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime corridor illumination is treated neutrally with restrained coolness and documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nThe THIRD attached image (STRUCTURE LOOK) is the identity source of the fixed structure at this location: its faces, openings, levels, materials and signage are truth. Where it conflicts with the LOCATION PHOTOGRAPH about the structure itself, the STRUCTURE LOOK wins; the photograph still governs the surroundings, time of day and lighting.\n\nPERIOD REFERENCE — 2000년대 한국 복도식 아파트 현관문 및 초인종: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S42sh3__bgfirst_bg.png",
  "asset_id": "e180fa2d-014c-4303-b1cc-8d97685e77ac",
  "input_asset_ids": [
   "72c26891-fde5-4a28-a307-032e596260b7",
   "f9f209a7-a9d3-41e3-baab-025df0d9e775",
   "a0b48e7f-ee0c-4fea-b260-8bf3490af68b"
  ],
  "era_research": {
   "subject": "2000년대 한국 복도식 아파트 현관문 및 초인종",
   "queries": [
    [
     "2000년대 한국 복도식 아파트 복도 현관 방화문 디지털도어락 초인종 인터폰",
     "2000년대 한국 아파트 현관문 초인종 인터폰 복도식 아파트"
    ]
   ],
   "picked_url": "https://static.cdn.soomgo.com/upload/portfolio/5d392228-4850-4d9a-aff2-faa3d93020db.jpg?webp=1",
   "sha256": "a40fe776dccc3ed67baf285af26fb919277cb8034f155b743bc897e2d5053d5d",
   "file": "eraref_37d2273783207e96.png"
  }
 },
 "S42sh3": {
  "input_fingerprint": "99195ee701c88fe7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 손과 초인종 버튼 사이를 막아서듯 덮은 나상혁의 손바닥 클로즈업\n\nLOCATION (lock): In the apartment building's long access corridor, directly outside the family's front door and doorbell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height beside the apartment door controls, push into an extreme-close oblique insert across Na Sang-hyeok's open palm toward Seo Eui-yong's approaching fingers and the obscured doorbell behind it. Sang-hyeok's hand lies flat as a barrier in the center foreground while Eui-yong's fingers enter from the opposite edge, preserving enough of the door-control surround to make the intended press physically clear.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Na Sang-hyeok blocking hand in the middle-center of the frame, foreground; Seo Eui-yong pressing hand in the middle-left of the frame, foreground, reaches for covered doorbell button; doorbell button behind the palm in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (covered by Na Sang-hyeok's hand) — Its front face is angled toward the camera but partly hidden by the palm; used as The button is the blocked focal anchor immediately behind Sang-hyeok's palm; apartment door controls (visible around the contact area) — Their control-facing side is seen obliquely beside the hands; used as Provides a narrow scale and location reference around the hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime corridor illumination is treated neutrally with restrained coolness and documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok's hand remains laid over the doorbell button, blocking Euiyong from pressing it directly.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 초인종 버튼: \"호출\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 손과 초인종 버튼 사이를 막아서듯 덮은 나상혁의 손바닥 클로즈업\n\nLOCATION (lock): In the apartment building's long access corridor, directly outside the family's front door and doorbell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height beside the apartment door controls, push into an extreme-close oblique insert across Na Sang-hyeok's open palm toward Seo Eui-yong's approaching fingers and the obscured doorbell behind it. Sang-hyeok's hand lies flat as a barrier in the center foreground while Eui-yong's fingers enter from the opposite edge, preserving enough of the door-control surround to make the intended press physically clear.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Na Sang-hyeok blocking hand in the middle-center of the frame, foreground; Seo Eui-yong pressing hand in the middle-left of the frame, foreground, reaches for covered doorbell button; doorbell button behind the palm in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (covered by Na Sang-hyeok's hand) — Its front face is angled toward the camera but partly hidden by the palm; used as The button is the blocked focal anchor immediately behind Sang-hyeok's palm; apartment door controls (visible around the contact area) — Their control-facing side is seen obliquely beside the hands; used as Provides a narrow scale and location reference around the hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime corridor illumination is treated neutrally with restrained coolness and documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok's hand remains laid over the doorbell button, blocking Euiyong from pressing it directly.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 초인종 버튼: \"호출\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 손과 초인종 버튼 사이를 막아서듯 덮은 나상혁의 손바닥 클로즈업\n\nLOCATION (lock): In the apartment building's long access corridor, directly outside the family's front door and doorbell. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height beside the apartment door controls, push into an extreme-close oblique insert across Na Sang-hyeok's open palm toward Seo Eui-yong's approaching fingers and the obscured doorbell behind it. Sang-hyeok's hand lies flat as a barrier in the center foreground while Eui-yong's fingers enter from the opposite edge, preserving enough of the door-control surround to make the intended press physically clear.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Na Sang-hyeok blocking hand in the middle-center of the frame, foreground; Seo Eui-yong pressing hand in the middle-left of the frame, foreground, reaches for covered doorbell button; doorbell button behind the palm in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (covered by Na Sang-hyeok's hand) — Its front face is angled toward the camera but partly hidden by the palm; used as The button is the blocked focal anchor immediately behind Sang-hyeok's palm; apartment door controls (visible around the contact area) — Their control-facing side is seen obliquely beside the hands; used as Provides a narrow scale and location reference around the hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime corridor illumination is treated neutrally with restrained coolness and documentary texture.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok's hand remains laid over the doorbell button, blocking Euiyong from pressing it directly.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 초인종 버튼: \"호출\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S42sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S42sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L06B01.png"
    },
    {
     "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_apartment_home_sel.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 복장과 역할이 뒤바뀌었으며, 손바닥이 초인종을 전혀 가리지 못해 핵심 연출 구현에 실패했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "역할이 뒤바뀐 채 손바닥이 아닌 손등을 보이고 있으며, 뻗은 손의 해부학적 오류가 뚜렷합니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "네이비 재킷의 손이 초인종을 향해 다가가고, 가죽 재킷의 손은 초인종 아래에서 손바닥을 위로 펼친 채 허공에 머물러 있음.",
      "built_space": "아파트 복도 벽면에 초인종이 정상적인 위치와 각도로 부착되어 있음.",
      "entities": "'호출' 텍스트는 정확하게 렌더링됨. 나상혁(네이비 재킷)과 서의용(가죽 재킷)의 역할이 반대로 적용되었으며, 어느 손도 초인종을 덮지 않음.",
      "hard_violations": [
       "인물 역할 및 복장 반전 (프롬프트의 행동 지시와 불일치)",
       "명시된 위치에 배치되지 않은 신체 (손바닥이 초인종을 가리지 않음)"
      ],
      "physics": "두 손 모두 팔에 연결되어 자세를 유지하고 있으며, 해부학적 구조가 정상적임."
     },
     {
      "label": "A",
      "direction": "네이비 재킷의 손이 초인종 쪽을 향해 뻗고, 가죽 재킷의 손은 초인종 패널 위에 손등을 보이며 밀착되어 있음.",
      "built_space": "아파트 복도 벽면 구조와 초인종 패널이 프레임 내에 안정적으로 자리 잡고 있음.",
      "entities": "'호출' 텍스트는 정확함. 두 인물의 복장이 반대로 적용되었고, 지시된 손바닥 대신 손등이 노출되었으며 버튼을 완전히 가리지 못함.",
      "hard_violations": [
       "인물 역할 및 복장 반전 (프롬프트의 행동 지시와 불일치)",
       "물리적으로 불가능한 인체 해부학 (네이비 재킷 손의 손가락 개수 및 형태 왜곡)"
      ],
      "physics": "가죽 재킷 손은 벽면에 지탱하고 있으나, 네이비 재킷 손은 관절과 손가락이 기형적으로 융합되어 허공에 뻗어 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 복장과 역할이 뒤바뀌었으며, 손바닥이 초인종을 전혀 가리지 못해 핵심 연출 구현에 실패했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "역할이 뒤바뀐 채 손바닥이 아닌 손등을 보이고 있으며, 뻗은 손의 해부학적 오류가 뚜렷합니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "네이비 재킷의 손이 초인종을 향해 다가가고, 가죽 재킷의 손은 초인종 아래에서 손바닥을 위로 펼친 채 허공에 머물러 있음.",
      "built_space": "아파트 복도 벽면에 초인종이 정상적인 위치와 각도로 부착되어 있음.",
      "entities": "'호출' 텍스트는 정확하게 렌더링됨. 나상혁(네이비 재킷)과 서의용(가죽 재킷)의 역할이 반대로 적용되었으며, 어느 손도 초인종을 덮지 않음.",
      "hard_violations": [
       "인물 역할 및 복장 반전 (프롬프트의 행동 지시와 불일치)",
       "명시된 위치에 배치되지 않은 신체 (손바닥이 초인종을 가리지 않음)"
      ],
      "physics": "두 손 모두 팔에 연결되어 자세를 유지하고 있으며, 해부학적 구조가 정상적임."
     },
     {
      "label": "A",
      "direction": "네이비 재킷의 손이 초인종 쪽을 향해 뻗고, 가죽 재킷의 손은 초인종 패널 위에 손등을 보이며 밀착되어 있음.",
      "built_space": "아파트 복도 벽면 구조와 초인종 패널이 프레임 내에 안정적으로 자리 잡고 있음.",
      "entities": "'호출' 텍스트는 정확함. 두 인물의 복장이 반대로 적용되었고, 지시된 손바닥 대신 손등이 노출되었으며 버튼을 완전히 가리지 못함.",
      "hard_violations": [
       "인물 역할 및 복장 반전 (프롬프트의 행동 지시와 불일치)",
       "물리적으로 불가능한 인체 해부학 (네이비 재킷 손의 손가락 개수 및 형태 왜곡)"
      ],
      "physics": "가죽 재킷 손은 벽면에 지탱하고 있으나, 네이비 재킷 손은 관절과 손가락이 기형적으로 융합되어 허공에 뻗어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "초인종을 가로막는 프롬프트의 핵심 액션이 완전히 누락되어 두 손이 무의미한 제스처를 취하고 있습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물들의 역할(의상)이 반대로 연출되었으며, 화면 좌측 손가락 끝에 허공에 뜬 정체불명의 선이 나타나는 치명적 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽 손(남색 소매)이 오른쪽 손(가죽 소매)의 펴진 손바닥을 향하고 있으며, 두 손 모두 초인종을 가리거나 누르려 하지 않음.",
      "built_space": "벽면에 금속 재질의 초인종 패널이 부착되어 있으며, 프레임 내에 다른 구조물은 보이지 않음.",
      "entities": "왼쪽 소매는 나상혁(남색 재킷, 줄무늬 셔츠), 오른쪽 소매는 서의용(가죽 재킷)의 의상과 일치함. 초인종에 '호출' 텍스트가 정확히 구현됨.",
      "hard_violations": [],
      "physics": "두 손은 초인종 아래 허공에서 포즈를 취하고 있으며 신체(팔)에 의해 지지됨."
     },
     {
      "label": "B",
      "direction": "오른쪽 손(가죽 소매)이 초인종 패널 상단을 덮어 누르고 있으며, 왼쪽 손(남색 소매)이 버튼을 향해 다가가고 있음.",
      "built_space": "벽면에 플라스틱 재질의 초인종 패널이 부착되어 있고, 배경으로 참조 사진과 유사한 아파트 복도와 난간이 보임.",
      "entities": "왼쪽 소매는 나상혁, 오른쪽 소매는 서의용과 일치하나, 프롬프트에서 지시한 역할(막는 자와 누르는 자)이 서로 뒤바뀜. '호출' 텍스트가 구현됨.",
      "hard_violations": [
       "왼쪽 손 집게손가락 끝부분 허공에 정체불명의 가로선이 생성됨 (invented object/leaked marker)"
      ],
      "physics": "오른쪽 손은 벽과 패널에 닿아 지지되고 있으나, 왼쪽 손가락 끝에 닿지 않은 채 허공에 떠 있는 선 형태의 인공물이 존재함."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "초인종을 가로막는 프롬프트의 핵심 액션이 완전히 누락되어 두 손이 무의미한 제스처를 취하고 있습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물들의 역할(의상)이 반대로 연출되었으며, 화면 좌측 손가락 끝에 허공에 뜬 정체불명의 선이 나타나는 치명적 오류가 있습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "왼쪽 손(남색 소매)이 오른쪽 손(가죽 소매)의 펴진 손바닥을 향하고 있으며, 두 손 모두 초인종을 가리거나 누르려 하지 않음.",
      "built_space": "벽면에 금속 재질의 초인종 패널이 부착되어 있으며, 프레임 내에 다른 구조물은 보이지 않음.",
      "entities": "왼쪽 소매는 나상혁(남색 재킷, 줄무늬 셔츠), 오른쪽 소매는 서의용(가죽 재킷)의 의상과 일치함. 초인종에 '호출' 텍스트가 정확히 구현됨.",
      "hard_violations": [],
      "physics": "두 손은 초인종 아래 허공에서 포즈를 취하고 있으며 신체(팔)에 의해 지지됨."
     },
     {
      "label": "A",
      "direction": "오른쪽 손(가죽 소매)이 초인종 패널 상단을 덮어 누르고 있으며, 왼쪽 손(남색 소매)이 버튼을 향해 다가가고 있음.",
      "built_space": "벽면에 플라스틱 재질의 초인종 패널이 부착되어 있고, 배경으로 참조 사진과 유사한 아파트 복도와 난간이 보임.",
      "entities": "왼쪽 소매는 나상혁, 오른쪽 소매는 서의용과 일치하나, 프롬프트에서 지시한 역할(막는 자와 누르는 자)이 서로 뒤바뀜. '호출' 텍스트가 구현됨.",
      "hard_violations": [
       "왼쪽 손 집게손가락 끝부분 허공에 정체불명의 가로선이 생성됨 (invented object/leaked marker)"
      ],
      "physics": "오른쪽 손은 벽과 패널에 닿아 지지되고 있으나, 왼쪽 손가락 끝에 닿지 않은 채 허공에 떠 있는 선 형태의 인공물이 존재함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 5,
     "B": 7
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "네이비 재킷의 손이 초인종을 향해 다가가고, 가죽 재킷의 손은 초인종 아래에서 손바닥을 위로 펼친 채 허공에 머물러 있음.",
    "built_space": "아파트 복도 벽면에 초인종이 정상적인 위치와 각도로 부착되어 있음.",
    "entities": "'호출' 텍스트는 정확하게 렌더링됨. 나상혁(네이비 재킷)과 서의용(가죽 재킷)의 역할이 반대로 적용되었으며, 어느 손도 초인종을 덮지 않음.",
    "hard_violations": [
     "인물 역할 및 복장 반전 (프롬프트의 행동 지시와 불일치)",
     "명시된 위치에 배치되지 않은 신체 (손바닥이 초인종을 가리지 않음)"
    ],
    "physics": "두 손 모두 팔에 연결되어 자세를 유지하고 있으며, 해부학적 구조가 정상적임."
   },
   {
    "label": "A",
    "direction": "네이비 재킷의 손이 초인종 쪽을 향해 뻗고, 가죽 재킷의 손은 초인종 패널 위에 손등을 보이며 밀착되어 있음.",
    "built_space": "아파트 복도 벽면 구조와 초인종 패널이 프레임 내에 안정적으로 자리 잡고 있음.",
    "entities": "'호출' 텍스트는 정확함. 두 인물의 복장이 반대로 적용되었고, 지시된 손바닥 대신 손등이 노출되었으며 버튼을 완전히 가리지 못함.",
    "hard_violations": [
     "인물 역할 및 복장 반전 (프롬프트의 행동 지시와 불일치)",
     "물리적으로 불가능한 인체 해부학 (네이비 재킷 손의 손가락 개수 및 형태 왜곡)"
    ],
    "physics": "가죽 재킷 손은 벽면에 지탱하고 있으나, 네이비 재킷 손은 관절과 손가락이 기형적으로 융합되어 허공에 뻗어 있음."
   }
  ],
  "totals": {
   "A": 5,
   "B": 7
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "인물의 복장과 역할이 뒤바뀌었으며, 손바닥이 초인종을 전혀 가리지 못해 핵심 연출 구현에 실패했습니다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "역할이 뒤바뀐 채 손바닥이 아닌 손등을 보이고 있으며, 뻗은 손의 해부학적 오류가 뚜렷합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L06B01.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_apartment_home_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "초인종을 누르려는 왼쪽 손이 나상혁의 의상을, 막으려는 아래쪽 손이 서의용의 의상을 입고 있어 인물의 역할과 의상이 뒤바뀜.",
     "fix_en": "Change the left sleeve to a black leather jacket over a grey hoodie, and the bottom sleeve to a navy jacket over a light blue shirt. Preserve the hand poses, doorbell panel, wall, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "손바닥이 초인종 버튼을 덮어 가려야 한다는 지시와 달리, 손이 버튼 아래에 위치하여 초인종 버튼이 완전히 노출되어 있음.",
     "fix_en": "Redraw the bottom palm higher so it completely covers the round doorbell button. Preserve the left hand, the characters' sleeves, the wall, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "아래쪽 손바닥의 왼쪽 가장자리에 비정상적인 살점과 손가락 관절 형태가 추가로 생성되어 해부학적으로 어색함.",
     "fix_en": "Remove the extra flesh on the left edge of the bottom palm to restore normal anatomy. Preserve both hands, sleeves, doorbell panel, lighting, and framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "초인종 패널이 로케이션 사진의 비디오 인터폰과 다른 형태임",
     "fix_en": "Replace the metal panel with a dark grey video intercom unit matching the location reference. Preserve the hands, sleeves, wall, lighting, and framing.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "초인종을 누르려는 왼쪽 손이 나상혁의 의상을, 막으려는 아래쪽 손이 서의용의 의상을 입고 있어 인물의 역할과 의상이 뒤바뀜.",
     "severity": "critical"
    },
    {
     "issue_ko": "손바닥이 초인종 버튼을 덮어 가려야 한다는 지시와 달리, 손이 버튼 아래에 위치하여 초인종 버튼이 완전히 노출되어 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "아래쪽 손바닥의 왼쪽 가장자리에 비정상적인 살점과 손가락 관절 형태가 추가로 생성되어 해부학적으로 어색함.",
     "severity": "major"
    },
    {
     "issue_ko": "막는 손바닥은 서의용 가죽재킷, 다가오는 손은 나상혁 남색재킷·남방으로 인물이 뒤바뀜",
     "severity": "critical"
    },
    {
     "issue_ko": "나상혁의 손바닥이 초인종 버튼을 덮지 않아 버튼이 가려지지 않고 완전히 노출됨",
     "severity": "major"
    },
    {
     "issue_ko": "초인종 패널이 로케이션 사진의 비디오 인터폰과 다른 형태임",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "아래쪽 손바닥의 왼쪽 가장자리에 비정상적인 살점과 손가락 관절 형태가 추가로 생성되어 해부학적으로 어색함.",
    "fix_en": "Remove the extra flesh on the left edge of the bottom palm to restore normal anatomy. Preserve both hands, sleeves, doorbell panel, lighting, and framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "초인종 패널이 로케이션 사진의 비디오 인터폰과 다른 형태임",
    "fix_en": "Replace the metal panel with a dark grey video intercom unit matching the location reference. Preserve the hands, sleeves, wall, lighting, and framing.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the left sleeve to a black leather jacket over a grey hoodie, and the bottom sleeve to a navy jacket over a light blue shirt. Preserve the hand poses, doorbell panel, wall, lighting, and framing.\n- Redraw the bottom palm higher so it completely covers the round doorbell button. Preserve the left hand, the characters' sleeves, the wall, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "클로즈업 샷 프레이밍과 '호출' 텍스트 구현은 정확하나, 소매 의상으로 볼 때 두 인물의 역할(누르는 손과 막는 손)이 반대로 연출되었으며 버튼을 덮어 가리는 동작에 실패함."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "요구된 인서트 클로즈업 프레이밍을 완전히 무시하고 전신 샷을 생성했으며, 배경이 수직으로 나뉜 콜라주 하드 위반으로 실격됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽 손(남색 소매)은 초인종 버튼을 향해 뻗고 있으며, 오른쪽 손(가죽 소매)은 손바닥을 위로 향한 채 허공을 향함.",
      "built_space": "벽면에 초인종 패널과 버튼이 정상적으로 부착되어 있음.",
      "entities": "남색 재킷 소매(나상혁), 가죽 재킷 소매(서의용), 초인종 버튼, '호출' 텍스트가 확인됨.",
      "hard_violations": [],
      "physics": "두 손은 프레임 밖의 팔에 연결되어 허공에 안정적으로 들려 있음."
     },
     {
      "label": "B",
      "direction": "두 인물 모두 카메라 렌즈를 정면으로 응시하고 있으며, 나상혁의 손은 초인종 옆 벽을 짚고 있음.",
      "built_space": "왼쪽은 실내 벽면, 오른쪽은 야외 아파트 외관으로 공간이 물리적으로 분리되어 있음.",
      "entities": "서의용과 나상혁의 전신 인물, 초인종 패널이 확인됨.",
      "hard_violations": [
       "콜라주/패널 (화면이 실내 단색 배경과 야외 아파트 배경으로 수직 분할됨)"
      ],
      "physics": "두 인물은 평평한 바닥에 서서 두 발로 체중을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "클로즈업 샷 프레이밍과 '호출' 텍스트 구현은 정확하나, 소매 의상으로 볼 때 두 인물의 역할(누르는 손과 막는 손)이 반대로 연출되었으며 버튼을 덮어 가리는 동작에 실패함."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "요구된 인서트 클로즈업 프레이밍을 완전히 무시하고 전신 샷을 생성했으며, 배경이 수직으로 나뉜 콜라주 하드 위반으로 실격됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽 손(남색 소매)은 초인종 버튼을 향해 뻗고 있으며, 오른쪽 손(가죽 소매)은 손바닥을 위로 향한 채 허공을 향함.",
      "built_space": "벽면에 초인종 패널과 버튼이 정상적으로 부착되어 있음.",
      "entities": "남색 재킷 소매(나상혁), 가죽 재킷 소매(서의용), 초인종 버튼, '호출' 텍스트가 확인됨.",
      "hard_violations": [],
      "physics": "두 손은 프레임 밖의 팔에 연결되어 허공에 안정적으로 들려 있음."
     },
     {
      "label": "B",
      "direction": "두 인물 모두 카메라 렌즈를 정면으로 응시하고 있으며, 나상혁의 손은 초인종 옆 벽을 짚고 있음.",
      "built_space": "왼쪽은 실내 벽면, 오른쪽은 야외 아파트 외관으로 공간이 물리적으로 분리되어 있음.",
      "entities": "서의용과 나상혁의 전신 인물, 초인종 패널이 확인됨.",
      "hard_violations": [
       "콜라주/패널 (화면이 실내 단색 배경과 야외 아파트 배경으로 수직 분할됨)"
      ],
      "physics": "두 인물은 평평한 바닥에 서서 두 발로 체중을 지탱하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "클로즈업 프레이밍과 '호출' 텍스트는 훌륭하게 구현했으나, 두 인물의 의상(역할)이 뒤바뀌었고 손바닥이 초인종을 가로막지 못해 스테이징 지시를 온전히 달성하지 못했습니다."
     },
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "요구된 인서트 클로즈업을 무시하고 전신 샷을 생성했으며, 화면이 콜라주 형태로 분할되는 치명적인 하드 위반이 발생하여 사용할 수 없습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 인물 모두 시선이 카메라를 향해 있으며, 나상혁의 오른손은 초인종 옆에서 허공을 향해 들려 있음.",
      "built_space": "화면 왼쪽은 빈 흰색 벽면이고 오른쪽은 야외 아파트 건물이 나타나는 구조로, 위치 레퍼런스의 복도 공간을 전혀 반영하지 않은 불가능한 기하학적 공간임.",
      "entities": "서의용과 나상혁의 전신이 보이며 의상은 레퍼런스와 일치함. 흰 벽에 초인종이 부착되어 있으나 '호출' 텍스트가 없음.",
      "hard_violations": [
       "콜라주/패널 (화면이 실내 흰 벽과 야외 아파트 외관으로 부자연스럽게 분할됨)",
       "물리적으로 불가능한 스테이징 (서로 다른 두 공간이 하나의 평면에 강제로 합성됨)"
      ],
      "physics": "두 인물은 바닥에 서 있으나 배경의 건물과 융합되지 못하고 비현실적으로 분리되어 있음."
     },
     {
      "label": "B",
      "direction": "왼쪽에서 들어온 손가락은 초인종 버튼을 향하고 있으나, 오른쪽의 펼친 손바닥은 버튼 앞을 가로막지 않고 그 아래에서 위를 향하고 있음.",
      "built_space": "아파트 현관문 가장자리와 벽면의 질감, 금속 초인종 패널이 위치 레퍼런스와 일치하는 클로즈업 구도로 묘사됨.",
      "entities": "초인종 패널 상단에 '호출' 텍스트가 명확하게 렌더링됨. 왼쪽의 뻗는 손은 나상혁의 의상(네이비 재킷과 셔츠)을, 오른쪽의 펼친 손은 서의용의 의상(가죽 재킷)을 입고 있어 샷 텍스트가 지시한 인물의 역할이 서로 뒤바뀌어 있음.",
      "hard_violations": [],
      "physics": "두 손 모두 프레임 밖의 팔에 의해 자연스럽게 지탱되며 허공에 떠 있지 않음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "클로즈업 프레이밍과 '호출' 텍스트는 훌륭하게 구현했으나, 두 인물의 의상(역할)이 뒤바뀌었고 손바닥이 초인종을 가로막지 못해 스테이징 지시를 온전히 달성하지 못했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "요구된 인서트 클로즈업을 무시하고 전신 샷을 생성했으며, 화면이 콜라주 형태로 분할되는 치명적인 하드 위반이 발생하여 사용할 수 없습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "두 인물 모두 시선이 카메라를 향해 있으며, 나상혁의 오른손은 초인종 옆에서 허공을 향해 들려 있음.",
      "built_space": "화면 왼쪽은 빈 흰색 벽면이고 오른쪽은 야외 아파트 건물이 나타나는 구조로, 위치 레퍼런스의 복도 공간을 전혀 반영하지 않은 불가능한 기하학적 공간임.",
      "entities": "서의용과 나상혁의 전신이 보이며 의상은 레퍼런스와 일치함. 흰 벽에 초인종이 부착되어 있으나 '호출' 텍스트가 없음.",
      "hard_violations": [
       "콜라주/패널 (화면이 실내 흰 벽과 야외 아파트 외관으로 부자연스럽게 분할됨)",
       "물리적으로 불가능한 스테이징 (서로 다른 두 공간이 하나의 평면에 강제로 합성됨)"
      ],
      "physics": "두 인물은 바닥에 서 있으나 배경의 건물과 융합되지 못하고 비현실적으로 분리되어 있음."
     },
     {
      "label": "A",
      "direction": "왼쪽에서 들어온 손가락은 초인종 버튼을 향하고 있으나, 오른쪽의 펼친 손바닥은 버튼 앞을 가로막지 않고 그 아래에서 위를 향하고 있음.",
      "built_space": "아파트 현관문 가장자리와 벽면의 질감, 금속 초인종 패널이 위치 레퍼런스와 일치하는 클로즈업 구도로 묘사됨.",
      "entities": "초인종 패널 상단에 '호출' 텍스트가 명확하게 렌더링됨. 왼쪽의 뻗는 손은 나상혁의 의상(네이비 재킷과 셔츠)을, 오른쪽의 펼친 손은 서의용의 의상(가죽 재킷)을 입고 있어 샷 텍스트가 지시한 인물의 역할이 서로 뒤바뀌어 있음.",
      "hard_violations": [],
      "physics": "두 손 모두 프레임 밖의 팔에 의해 자연스럽게 지탱되며 허공에 떠 있지 않음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 0
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S42sh3__bgfirst_bg.png",
   "bg_asset_id": "e180fa2d-014c-4303-b1cc-8d97685e77ac",
   "bg_record_key": "S42sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate",
   "seed_attached": true
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S42sh3::cine": {
  "applied": true,
  "fingerprint": "11f6c07102cd53a1248232f1d2148b6b49ff2dac0a4ad55b77725b06636479bb",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S42sh3_sel.png",
  "source_sha256": "fca47655c7209610a8490492543ffb162d5399f8da6c882f92910d45379d4716",
  "file": "S42sh3_cine.png",
  "latency_ms": 10034
 },
 "S42sh5::signage": {
  "fp": "b9683c85735b1f26",
  "inscriptions": [
   {
    "surface_native": "초인종 호출 버튼",
    "text_native": "호출",
    "reason_ko": "2001~2017년 한국 아파트 초인종/인터폰 기기에 흔히 표시되어 있는 '호출' 버튼의 사실성을 높이기 위함."
   }
  ]
 },
 "S42sh5": {
  "input_fingerprint": "9994f45e0041888d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 초인종을 가린 나상혁의 손등 위를 자신의 손가락으로 거침없이 꾹 누르고 있는 서의용의 억센 손\n\nLOCATION (lock): At the apartment front door's bell in the building's interior access corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Remain at hand height on the same side of the interaction axis and tighten only toward the contact point: Seo Eui-yong's firm finger bears down on the back of Na Sang-hyeok's hand, transmitting pressure to the concealed doorbell. Keep the overlapping hands centered with a slim border of the door-control panel visible, avoiding any change in angle while making the force of the press unmistakable.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: overlapping hands at contact point in the middle-center of the frame, foreground; concealed doorbell beneath the hands in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (being pressed through Na Sang-hyeok's hand) — Its front face lies directly beneath the overlapping hands and is only marginally visible; used as Remains the concealed mechanical destination of the downward force; door-control surround (partly visible) — A narrow portion of its outward-facing side remains visible around the hands; used as Frames the contact without allowing the hands to lose their real-world scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient corridor light preserves skin and surface detail with subdued contrast and a lightly documentary finish.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나상혁, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok's hand is still covering the bell while Euiyong presses down on the back of that hand to activate it.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 초인종 호출 버튼: \"호출\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 초인종을 가린 나상혁의 손등 위를 자신의 손가락으로 거침없이 꾹 누르고 있는 서의용의 억센 손\n\nLOCATION (lock): At the apartment front door's bell in the building's interior access corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Remain at hand height on the same side of the interaction axis and tighten only toward the contact point: Seo Eui-yong's firm finger bears down on the back of Na Sang-hyeok's hand, transmitting pressure to the concealed doorbell. Keep the overlapping hands centered with a slim border of the door-control panel visible, avoiding any change in angle while making the force of the press unmistakable.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: overlapping hands at contact point in the middle-center of the frame, foreground; concealed doorbell beneath the hands in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (being pressed through Na Sang-hyeok's hand) — Its front face lies directly beneath the overlapping hands and is only marginally visible; used as Remains the concealed mechanical destination of the downward force; door-control surround (partly visible) — A narrow portion of its outward-facing side remains visible around the hands; used as Frames the contact without allowing the hands to lose their real-world scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient corridor light preserves skin and surface detail with subdued contrast and a lightly documentary finish.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나상혁, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok's hand is still covering the bell while Euiyong presses down on the back of that hand to activate it.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 초인종 호출 버튼: \"호출\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 초인종을 가린 나상혁의 손등 위를 자신의 손가락으로 거침없이 꾹 누르고 있는 서의용의 억센 손\n\nLOCATION (lock): At the apartment front door's bell in the building's interior access corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Remain at hand height on the same side of the interaction axis and tighten only toward the contact point: Seo Eui-yong's firm finger bears down on the back of Na Sang-hyeok's hand, transmitting pressure to the concealed doorbell. Keep the overlapping hands centered with a slim border of the door-control panel visible, avoiding any change in angle while making the force of the press unmistakable.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: overlapping hands at contact point in the middle-center of the frame, foreground; concealed doorbell beneath the hands in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: doorbell button (being pressed through Na Sang-hyeok's hand) — Its front face lies directly beneath the overlapping hands and is only marginally visible; used as Remains the concealed mechanical destination of the downward force; door-control surround (partly visible) — A narrow portion of its outward-facing side remains visible around the hands; used as Frames the contact without allowing the hands to lose their real-world scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral ambient corridor light preserves skin and surface detail with subdued contrast and a lightly documentary finish.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나상혁, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sanghyeok's hand is still covering the bell while Euiyong presses down on the back of that hand to activate it.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 초인종 호출 버튼: \"호출\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "왼쪽에서 뻗어나온 손이 아래쪽에서 뻗어나온 손의 위를 누르고 있음.",
    "built_space": "'호출'이라고 적힌 금속 초인종 패널이 배경에 있음.",
    "entities": "왼쪽 팔은 가죽 재킷 소매에 남색 시보리와 줄무늬 셔츠가 섞인 잘못된 형태이며, 아래쪽 팔은 가죽 재킷 소매임.",
    "hard_violations": [
     "발명된 의상 (가죽 소매에 남색 시보리와 셔츠 소매가 결합됨)",
     "캐릭터 의상과 행동 역할 불일치"
    ],
    "physics": "두 손이 초인종 패널에 맞닿아 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "위쪽에서 뻗어나온 손이 초인종을 덮고 있는 아래쪽 손의 손등을 손가락으로 꾹 누르고 있음.",
    "built_space": "'호출'이라고 적힌 금속 초인종 패널이 배경에 올바르게 위치함.",
    "entities": "위쪽 손은 가죽 재킷 소매(서의용), 아래쪽 손은 남색 재킷과 줄무늬 셔츠 소매(나상혁)로 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "위쪽 손이 아래쪽 손등을 누르는 힘과, 아래쪽 손이 벽면 초인종에 닿아 지탱되는 물리적 상태가 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 8,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "서의용(가죽 재킷)의 손이 초인종을 가린 나상혁(남색 재킷)의 손등을 누르는 모습을 지시문과 레퍼런스에 맞게 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "두 캐릭터의 의상이 하나로 합쳐진 심각한 오류가 있으며, 인물별 의상과 행동의 연결이 지시문과 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 나상혁, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S42sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "서의용(가죽 재킷)의 손이 나상혁(네이비 재킷)의 손등을 눌러야 하나, 나상혁의 손바닥을 쥐고 있으며 서의용의 손가락이 초인종에 직접 닿아 있음.",
     "fix_en": "Redraw the interaction: the hand in the navy sleeve must lie flat against the panel, completely covering the doorbell button with its back facing the camera, while the hand in the leather sleeve rests on top, pressing one finger firmly onto the back of the navy sleeve's hand. Keep the sleeves, the '호출' text, the metal panel, the wall, and the ambient lighting exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 왼쪽 아래에서 뻗어나온 나상혁(네이비 재킷)의 손이 손바닥을 보이고 있으나, 엄지손가락이 오른쪽에 위치하여 왼팔에 오른손이 붙은 해부학적 오류가 있음.",
     "fix_en": "Redraw the hand extending from the navy sleeve to correct the reversed anatomy, turning it into a proper left hand that shows the back of the hand to the camera. Preserve the leather sleeve and its hand, the metal panel, the background wall, and the lighting.",
     "severity": "critical",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "서의용(가죽 재킷)의 손이 나상혁(네이비 재킷)의 손등을 눌러야 하나, 나상혁의 손바닥을 쥐고 있으며 서의용의 손가락이 초인종에 직접 닿아 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "나상혁(네이비 재킷)의 손이 초인종을 완전히 가리고 손등이 보여야 하나, 손바닥이 앞을 향한 채 서의용의 손 아래에 깔려 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 샷에서는 네이비 재킷의 손이 위쪽(손등 보임), 가죽 재킷의 손이 아래쪽(손바닥 보임)이었으나, 현재 샷에서는 두 소매의 위치와 손의 방향이 완전히 뒤바뀜.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 왼쪽 아래에서 뻗어나온 나상혁(네이비 재킷)의 손이 손바닥을 보이고 있으나, 엄지손가락이 오른쪽에 위치하여 왼팔에 오른손이 붙은 해부학적 오류가 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "초인종 버튼이 나상혁의 손에 가려지지 않고 패널 상단에 완전히 드러나 있으며, 위쪽 가죽 소매 손가락이 버튼을 직접 누르고 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용(가죽 소매)의 손이 손등을 손가락으로 누르지 않고 아래쪽 손을 통째로 감싸 쥐고 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the interaction: the hand in the navy sleeve must lie flat against the panel, completely covering the doorbell button with its back facing the camera, while the hand in the leather sleeve rests on top, pressing one finger firmly onto the back of the navy sleeve's hand. Keep the sleeves, the '호출' text, the metal panel, the wall, and the ambient lighting exactly as they are.\n- Redraw the hand extending from the navy sleeve to correct the reversed anatomy, turning it into a proper left hand that shows the back of the hand to the camera. Preserve the leather sleeve and its hand, the metal panel, the background wall, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "서의용(가죽 재킷)이 나상혁(남색 재킷)의 손등을 꾹 누르는 텍스트의 동작을 정확한 역할과 의상 매칭을 통해 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "누르는 손과 초인종을 가린 손의 인물(의상) 역할이 반대로 뒤바뀌었으며, 하단에 불필요한 세 번째 팔이 등장하는 치명적인 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "위쪽의 가죽 재킷을 입은 손이 아래쪽의 남색 재킷을 입은 손등을 초인종 방향으로 강하게 누르고 있음.",
      "built_space": "벽면에 부착된 초인종 패널과 '호출' 버튼이 올바른 위치와 비율로 배치되어 있으며, 손에 의해 버튼이 자연스럽게 가려짐.",
      "entities": "서의용의 손(가죽 재킷)이 누르는 역할을, 나상혁의 손(남색 재킷과 줄무늬 셔츠)이 초인종을 가리는 역할을 텍스트에 맞게 정확히 수행함. '호출' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "두 손의 겹침과 눌리는 압력이 해부학적으로 자연스러우며, 벽과 초인종 패널에 의해 안정적으로 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "위쪽 남색 소매의 손가락이 아래쪽 가죽 소매의 손등을 가리키거나 누르고 있음.",
      "built_space": "벽면과 초인종 패널, '호출' 텍스트가 정상적으로 존재함.",
      "entities": "나상혁(남색 재킷)이 서의용(가죽 재킷)의 손을 누르고 있어 프롬프트가 지시한 인물의 역할이 완전히 반대로 적용됨.",
      "hard_violations": [
       "신체 구조 오류 및 중복: 화면 좌측 하단에 남색 재킷을 입은 세 번째 팔이 추가로 생성됨.",
       "연출 오류: 텍스트가 지정한 인물 간의 역할(누르는 주체와 눌리는 대상)이 정반대로 묘사됨."
      ],
      "physics": "손의 위치와 접촉은 물리적으로 가능해 보이나, 세 번째 팔이 등장하여 전체적인 신체 구조가 성립하지 않음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "서의용(가죽 재킷)이 나상혁(남색 재킷)의 손등을 꾹 누르는 텍스트의 동작을 정확한 역할과 의상 매칭을 통해 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "누르는 손과 초인종을 가린 손의 인물(의상) 역할이 반대로 뒤바뀌었으며, 하단에 불필요한 세 번째 팔이 등장하는 치명적인 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "위쪽의 가죽 재킷을 입은 손이 아래쪽의 남색 재킷을 입은 손등을 초인종 방향으로 강하게 누르고 있음.",
      "built_space": "벽면에 부착된 초인종 패널과 '호출' 버튼이 올바른 위치와 비율로 배치되어 있으며, 손에 의해 버튼이 자연스럽게 가려짐.",
      "entities": "서의용의 손(가죽 재킷)이 누르는 역할을, 나상혁의 손(남색 재킷과 줄무늬 셔츠)이 초인종을 가리는 역할을 텍스트에 맞게 정확히 수행함. '호출' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "두 손의 겹침과 눌리는 압력이 해부학적으로 자연스러우며, 벽과 초인종 패널에 의해 안정적으로 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "위쪽 남색 소매의 손가락이 아래쪽 가죽 소매의 손등을 가리키거나 누르고 있음.",
      "built_space": "벽면과 초인종 패널, '호출' 텍스트가 정상적으로 존재함.",
      "entities": "나상혁(남색 재킷)이 서의용(가죽 재킷)의 손을 누르고 있어 프롬프트가 지시한 인물의 역할이 완전히 반대로 적용됨.",
      "hard_violations": [
       "신체 구조 오류 및 중복: 화면 좌측 하단에 남색 재킷을 입은 세 번째 팔이 추가로 생성됨.",
       "연출 오류: 텍스트가 지정한 인물 간의 역할(누르는 주체와 눌리는 대상)이 정반대로 묘사됨."
      ],
      "physics": "손의 위치와 접촉은 물리적으로 가능해 보이나, 세 번째 팔이 등장하여 전체적인 신체 구조가 성립하지 않음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "캐릭터의 의상과 행동(가죽 재킷을 입은 서의용이 남색 재킷을 입은 나상혁의 손을 누름)이 프롬프트와 완벽히 일치하며, '호출' 텍스트와 클로즈업 구도까지 정확히 구현된 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "두 인물의 행동 역할이 반대로 적용되었으며, 화면 좌측 하단에 세 번째 팔이 등장하는 치명적인 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "서의용의 손(가죽 재킷)이 초인종을 가린 나상혁의 손(남색 재킷)등 위를 정확한 궤적으로 누르고 있음.",
      "built_space": "아파트 복도의 초인종 패널이 프레임 중앙 배경에 위치하며, 지정된 '호출' 텍스트가 바른 방향으로 정확히 렌더링됨.",
      "entities": "가죽 재킷을 입은 서의용의 손과 남색 재킷 및 줄무늬 셔츠를 입은 나상혁의 손 의상이 레퍼런스와 완벽히 일치함.",
      "hard_violations": [],
      "physics": "위에서 아래로 누르는 물리적인 힘과 손이 겹쳐진 상태의 접촉면, 무게감이 자연스럽게 지탱 및 묘사됨."
     },
     {
      "label": "A",
      "direction": "남색 재킷을 입은 손이 가죽 재킷을 입은 손을 누르고 있어 지시된 방향 및 행동 주체가 완전히 반대로 묘사됨.",
      "built_space": "벽면에 부착된 초인종 패널과 '호출' 텍스트는 올바른 위치에 렌더링됨.",
      "entities": "나상혁(남색 재킷)과 서의용(가죽 재킷)의 의상 텍스처는 구현되었으나 행동 주체가 뒤바뀌었으며, 불필요한 팔이 추가로 묘사됨.",
      "hard_violations": [
       "화면 좌측 하단에 남색 재킷을 입은 세 번째 팔이 등장함 (추가된 신체 부위)",
       "서의용과 나상혁의 행동과 의상 매칭이 완전히 반대로 적용됨 (역할 반전)"
      ],
      "physics": "손가락이 손등을 누르는 접촉 자체는 있으나, 화면에 등장하는 팔의 개수와 배치가 물리적으로 불가능함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "캐릭터의 의상과 행동(가죽 재킷을 입은 서의용이 남색 재킷을 입은 나상혁의 손을 누름)이 프롬프트와 완벽히 일치하며, '호출' 텍스트와 클로즈업 구도까지 정확히 구현된 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "두 인물의 행동 역할이 반대로 적용되었으며, 화면 좌측 하단에 세 번째 팔이 등장하는 치명적인 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서의용의 손(가죽 재킷)이 초인종을 가린 나상혁의 손(남색 재킷)등 위를 정확한 궤적으로 누르고 있음.",
      "built_space": "아파트 복도의 초인종 패널이 프레임 중앙 배경에 위치하며, 지정된 '호출' 텍스트가 바른 방향으로 정확히 렌더링됨.",
      "entities": "가죽 재킷을 입은 서의용의 손과 남색 재킷 및 줄무늬 셔츠를 입은 나상혁의 손 의상이 레퍼런스와 완벽히 일치함.",
      "hard_violations": [],
      "physics": "위에서 아래로 누르는 물리적인 힘과 손이 겹쳐진 상태의 접촉면, 무게감이 자연스럽게 지탱 및 묘사됨."
     },
     {
      "label": "B",
      "direction": "남색 재킷을 입은 손이 가죽 재킷을 입은 손을 누르고 있어 지시된 방향 및 행동 주체가 완전히 반대로 묘사됨.",
      "built_space": "벽면에 부착된 초인종 패널과 '호출' 텍스트는 올바른 위치에 렌더링됨.",
      "entities": "나상혁(남색 재킷)과 서의용(가죽 재킷)의 의상 텍스처는 구현되었으나 행동 주체가 뒤바뀌었으며, 불필요한 팔이 추가로 묘사됨.",
      "hard_violations": [
       "화면 좌측 하단에 남색 재킷을 입은 세 번째 팔이 등장함 (추가된 신체 부위)",
       "서의용과 나상혁의 행동과 의상 매칭이 완전히 반대로 적용됨 (역할 반전)"
      ],
      "physics": "손가락이 손등을 누르는 접촉 자체는 있으나, 화면에 등장하는 팔의 개수와 배치가 물리적으로 불가능함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S42sh3"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S42sh5::cine": {
  "applied": true,
  "fingerprint": "c7c6266281279c8a993feb051c0aa21876bcd34691209c52e370ab37f670d9fd",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S42sh5_sel.png",
  "source_sha256": "81f9e6da980d0ecbe0f1c76abbb12dc2f682880c6663cf7169a700673b6a1d8e",
  "file": "S42sh5_cine.png",
  "latency_ms": 10003
 },
 "S43sh1::signage": {
  "fp": "5ec71e3080f98083",
  "inscriptions": []
 },
 "era_assess::aa350e72a4f3d918": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 한국 아파트 거실 및 좌식 테이블",
    "search_terms_native": [
     "2000년대 한국 아파트 거실",
     "가정집 거실 좌식 탁자",
     "한국 인테리어 2000년대 아파트"
    ],
    "language_lock_native": "이 검색은 오직 한국어로만 수행되어야 하며, 다른 언어로 번역하거나 추가적인 영어 검색어를 포함해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 가정집의 장판 바닥, 전형적인 아파트 거실 창호 구조, 그리고 독특한 좌식 테이블 문화는 일반적인 현대 서구식 거실 이미지와 확연히 다릅니다."
   }
  ]
 },
 "era_ref::94e105b2ac3d4672": {
  "subject": "2000년대-2010년대 한국 아파트 거실 및 좌식 테이블",
  "terms": [
   "2000년대 한국 아파트 거실",
   "가정집 거실 좌식 탁자",
   "한국 인테리어 2000년대 아파트"
  ],
  "queries": [
   [
    "2000년대 한국 아파트 거실 좌식 탁자 가정집 인테리어",
    "2010년대 한국 아파트 거실 좌식 테이블 인테리어"
   ],
   [
    "2000년대 한국 아파트 거실 좌식 탁자 사진",
    "2000년대 가정집 거실 인테리어 좌식 테이블",
    "2010년대 초반 한국 아파트 거실 사진 좌탁"
   ]
  ],
  "candidates": 1,
  "picked_index": 1,
  "picked_url": "https://thumbnews.nateimg.co.kr/view610/news.nateimg.co.kr/orgImg/do/2020/02/05/99538776.3.jpg",
  "picked_reason_ko": "2000년대 한국의 일반적인 아파트 거실 분위기와 소파 앞 좌식 테이블의 형태·재료·사용 방식을 충분히 읽을 수 있는 유일한 후보다.",
  "sha256": "c50ab8cdd04e86255dc2a989049284ca7b0d5ad96d4509529c05943696d5a1d0",
  "file": "eraref_94e105b2ac3d4672.png"
 },
 "S43sh1::bgfirst_bg": {
  "input_fingerprint": "84143295bb0970f3",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 뜨거운 김이 오르는 커피잔이 놓인 테이블을 사이에 두고 마주 앉은 서의용, 나상혁, 민정의 전경\n\nLOCATION (lock): At the sitting-room table inside the family home, with hot coffee and an open laptop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated chest height on the open side of the table, hold a wide three-quarter view with the steaming coffee cup small in the lower foreground and the three participants arranged around the table beyond it. Seo Eui-yong angles his torso toward the room while glancing aside, Na Sang-hyeok bends toward the open laptop, and Min-jeong sits opposite them with her attention fixed on Eui-yong, making the interview relationship legible without forming a frontal lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Seo Eui-yong on the investigators' side in the middle-left of the frame, midground; Na Sang-hyeok beside the laptop in the middle-center of the frame, midground; Min-jeong opposite the investigators in the middle-right of the frame, midground, looks toward Seo Eui-yong across the table.\n- KEY BACKGROUND ELEMENTS: coffee cup (hot steam rising) — Seen at a three-quarter angle from slightly above, between the camera and the seated group; used as Anchors the near foreground without exceeding a small portion of the frame; its rising steam creates a quiet visual measure of the pause; table (coffee cup and laptop placed on it) — Its open-side edge angles across the lower frame toward the participants; used as Organizes the three-way interview and separates foreground from the seated figures; open laptop (open) — The screen side faces Na Sang-hyeok and is seen obliquely by the camera; no specific screen content is established; used as Keeps Sang-hyeok's procedural role visible within the establishing composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is kept subdued and neutral-to-cool, with the visible steam softly separated from the restrained interior tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국 아파트 거실 및 좌식 테이블: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 뜨거운 김이 오르는 커피잔이 놓인 테이블을 사이에 두고 마주 앉은 서의용, 나상혁, 민정의 전경\n\nLOCATION (lock): At the sitting-room table inside the family home, with hot coffee and an open laptop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At seated chest height on the open side of the table, hold a wide three-quarter view with the steaming coffee cup small in the lower foreground and the three participants arranged around the table beyond it. Seo Eui-yong angles his torso toward the room while glancing aside, Na Sang-hyeok bends toward the open laptop, and Min-jeong sits opposite them with her attention fixed on Eui-yong, making the interview relationship legible without forming a frontal lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Seo Eui-yong on the investigators' side in the middle-left of the frame, midground; Na Sang-hyeok beside the laptop in the middle-center of the frame, midground; Min-jeong opposite the investigators in the middle-right of the frame, midground, looks toward Seo Eui-yong across the table.\n- KEY BACKGROUND ELEMENTS: coffee cup (hot steam rising) — Seen at a three-quarter angle from slightly above, between the camera and the seated group; used as Anchors the near foreground without exceeding a small portion of the frame; its rising steam creates a quiet visual measure of the pause; table (coffee cup and laptop placed on it) — Its open-side edge angles across the lower frame toward the participants; used as Organizes the three-way interview and separates foreground from the seated figures; open laptop (open) — The screen side faces Na Sang-hyeok and is seen obliquely by the camera; no specific screen content is established; used as Keeps Sang-hyeok's procedural role visible within the establishing composition.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is kept subdued and neutral-to-cool, with the visible steam softly separated from the restrained interior tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국 아파트 거실 및 좌식 테이블: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S43sh1__bgfirst_bg.png",
  "asset_id": "eb3bac51-c07b-4025-9b87-b32cb09b025e",
  "input_asset_ids": [
   "19dc5eea-b1e4-4ac2-8f1f-142a3d84245d",
   "28f58b4d-6aee-4a72-8712-8c78b526a68a"
  ],
  "era_research": {
   "subject": "2000년대-2010년대 한국 아파트 거실 및 좌식 테이블",
   "queries": [
    [
     "2000년대 한국 아파트 거실 좌식 탁자 가정집 인테리어",
     "2010년대 한국 아파트 거실 좌식 테이블 인테리어"
    ],
    [
     "2000년대 한국 아파트 거실 좌식 탁자 사진",
     "2000년대 가정집 거실 인테리어 좌식 테이블",
     "2010년대 초반 한국 아파트 거실 사진 좌탁"
    ]
   ],
   "picked_url": "https://thumbnews.nateimg.co.kr/view610/news.nateimg.co.kr/orgImg/do/2020/02/05/99538776.3.jpg",
   "sha256": "c50ab8cdd04e86255dc2a989049284ca7b0d5ad96d4509529c05943696d5a1d0",
   "file": "eraref_94e105b2ac3d4672.png"
  }
 },
 "S43sh1": {
  "input_fingerprint": "5b365b06661b9667",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 뜨거운 김이 오르는 커피잔이 놓인 테이블을 사이에 두고 마주 앉은 서의용, 나상혁, 민정의 전경\n\nLOCATION (lock): At the sitting-room table inside the family home, with hot coffee and an open laptop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height on the open side of the table, hold a wide three-quarter view with the steaming coffee cup small in the lower foreground and the three participants arranged around the table beyond it. Seo Eui-yong angles his torso toward the room while glancing aside, Na Sang-hyeok bends toward the open laptop, and Min-jeong sits opposite them with her attention fixed on Eui-yong, making the interview relationship legible without forming a frontal lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Seo Eui-yong on the investigators' side in the middle-left of the frame, midground; Na Sang-hyeok beside the laptop in the middle-center of the frame, midground; Min-jeong opposite the investigators in the middle-right of the frame, midground, looks toward Seo Eui-yong across the table.\n- KEY BACKGROUND ELEMENTS: coffee cup (hot steam rising) — Seen at a three-quarter angle from slightly above, between the camera and the seated group; used as Anchors the near foreground without exceeding a small portion of the frame; its rising steam creates a quiet visual measure of the pause; table (coffee cup and laptop placed on it) — Its open-side edge angles across the lower frame toward the participants; used as Organizes the three-way interview and separates foreground from the seated figures; open laptop (open) — The screen side faces Na Sang-hyeok and is seen obliquely by the camera; no specific screen content is established; used as Keeps Sang-hyeok's procedural role visible within the establishing composition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is kept subdued and neutral-to-cool, with the visible steam softly separated from the restrained interior tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steaming coffee cups remain on the table, and Sanghyeok's laptop stays open during the interview with Minjeong.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 뜨거운 김이 오르는 커피잔이 놓인 테이블을 사이에 두고 마주 앉은 서의용, 나상혁, 민정의 전경\n\nLOCATION (lock): At the sitting-room table inside the family home, with hot coffee and an open laptop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height on the open side of the table, hold a wide three-quarter view with the steaming coffee cup small in the lower foreground and the three participants arranged around the table beyond it. Seo Eui-yong angles his torso toward the room while glancing aside, Na Sang-hyeok bends toward the open laptop, and Min-jeong sits opposite them with her attention fixed on Eui-yong, making the interview relationship legible without forming a frontal lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Seo Eui-yong on the investigators' side in the middle-left of the frame, midground; Na Sang-hyeok beside the laptop in the middle-center of the frame, midground; Min-jeong opposite the investigators in the middle-right of the frame, midground, looks toward Seo Eui-yong across the table.\n- KEY BACKGROUND ELEMENTS: coffee cup (hot steam rising) — Seen at a three-quarter angle from slightly above, between the camera and the seated group; used as Anchors the near foreground without exceeding a small portion of the frame; its rising steam creates a quiet visual measure of the pause; table (coffee cup and laptop placed on it) — Its open-side edge angles across the lower frame toward the participants; used as Organizes the three-way interview and separates foreground from the seated figures; open laptop (open) — The screen side faces Na Sang-hyeok and is seen obliquely by the camera; no specific screen content is established; used as Keeps Sang-hyeok's procedural role visible within the establishing composition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is kept subdued and neutral-to-cool, with the visible steam softly separated from the restrained interior tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steaming coffee cups remain on the table, and Sanghyeok's laptop stays open during the interview with Minjeong.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 뜨거운 김이 오르는 커피잔이 놓인 테이블을 사이에 두고 마주 앉은 서의용, 나상혁, 민정의 전경\n\nLOCATION (lock): At the sitting-room table inside the family home, with hot coffee and an open laptop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height on the open side of the table, hold a wide three-quarter view with the steaming coffee cup small in the lower foreground and the three participants arranged around the table beyond it. Seo Eui-yong angles his torso toward the room while glancing aside, Na Sang-hyeok bends toward the open laptop, and Min-jeong sits opposite them with her attention fixed on Eui-yong, making the interview relationship legible without forming a frontal lineup.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Seo Eui-yong on the investigators' side in the middle-left of the frame, midground; Na Sang-hyeok beside the laptop in the middle-center of the frame, midground; Min-jeong opposite the investigators in the middle-right of the frame, midground, looks toward Seo Eui-yong across the table.\n- KEY BACKGROUND ELEMENTS: coffee cup (hot steam rising) — Seen at a three-quarter angle from slightly above, between the camera and the seated group; used as Anchors the near foreground without exceeding a small portion of the frame; its rising steam creates a quiet visual measure of the pause; table (coffee cup and laptop placed on it) — Its open-side edge angles across the lower frame toward the participants; used as Organizes the three-way interview and separates foreground from the seated figures; open laptop (open) — The screen side faces Na Sang-hyeok and is seen obliquely by the camera; no specific screen content is established; used as Keeps Sang-hyeok's procedural role visible within the establishing composition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light is kept subdued and neutral-to-cool, with the visible steam softly separated from the restrained interior tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The steaming coffee cups remain on the table, and Sanghyeok's laptop stays open during the interview with Minjeong.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S43sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S43sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 민정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:803682>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L04B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 민정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:803682>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 500,
      "verdict_ko": "인물의 위치가 프롬프트와 반대로 배치되었고, 지정된 장소 참조와 전혀 다른 임의의 거실을 배경으로 생성하여 위치 조건을 위반했습니다.  ★위반: [gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜) / [gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함 / [gemini-pro] 장소 참조 이미지 위반 (구조와 배경이 전혀 다름) / [openrouter:x-ai/grok-4.6] 민정이 조사관들 맞은편이 아니라 같은 소파 쪽에 앉아 스테이징이 정한 좌석을 위반한다."
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "인물의 위치가 프롬프트와 반대로 배치되었으며, 지정된 장소를 완전히 무시하고 스튜디오 배경으로 렌더링하여 심각한 위반입니다.  ★위반: [gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜) / [gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함 / [gemini-pro] 장소 참조 이미지 위반 (공간적 맥락이 없는 스튜디오 배경)"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 2.0,
      "B": 1.5
     },
     "adjusted": {
      "A": 1.25,
      "B": 0.5
     },
     "violations": {
      "B": [
       "[gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜)",
       "[gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함",
       "[gemini-pro] 장소 참조 이미지 위반 (구조와 배경이 전혀 다름)",
       "[openrouter:x-ai/grok-4.6] 민정이 조사관들 맞은편이 아니라 같은 소파 쪽에 앉아 스테이징이 정한 좌석을 위반한다."
      ],
      "A": [
       "[gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜)",
       "[gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함",
       "[gemini-pro] 장소 참조 이미지 위반 (공간적 맥락이 없는 스튜디오 배경)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 500,
      "verdict_ko": "인물의 위치가 프롬프트와 반대로 배치되었고, 지정된 장소 참조와 전혀 다른 임의의 거실을 배경으로 생성하여 위치 조건을 위반했습니다.  ★위반: [gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜) / [gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함 / [gemini-pro] 장소 참조 이미지 위반 (구조와 배경이 전혀 다름) / [openrouter:x-ai/grok-4.6] 민정이 조사관들 맞은편이 아니라 같은 소파 쪽에 앉아 스테이징이 정한 좌석을 위반한다."
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "인물의 위치가 프롬프트와 반대로 배치되었으며, 지정된 장소를 완전히 무시하고 스튜디오 배경으로 렌더링하여 심각한 위반입니다.  ★위반: [gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜) / [gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함 / [gemini-pro] 장소 참조 이미지 위반 (공간적 맥락이 없는 스튜디오 배경)"
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 로케이션 레퍼런스를 무시한 임의의 거실을 묘사했으며, 상혁과 의용의 위치 및 역할(노트북 조작)이 프롬프트와 반대로 적용됨."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "로케이션 공간을 아예 누락한 흰색 배경이며, 인물들의 자리 배치와 행동이 프롬프트의 지시와 전혀 일치하지 않음."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "상혁과 민정은 의용과 노트북을 향하고, 의용은 노트북 화면을 응시함. 의용이 방을 향하고 상혁이 노트북을 조작해야 하는 프롬프트 지시와 다름.",
      "built_space": "소파와 책장이 있는 서구식 거실 구조로, 제시된 한국식 좌식 로케이션 레퍼런스를 완전히 무시함.",
      "entities": "왼쪽에 상혁, 중앙에 의용, 오른쪽에 민정이 위치함. 프롬프트의 배치(왼쪽 의용, 중앙 상혁)와 어긋남. 김이 나는 커피잔과 노트북은 존재함.",
      "hard_violations": [],
      "physics": "인물들은 의자와 소파에 앉아 지탱되고, 사물은 테이블 위에 정상적으로 놓여 있음."
     },
     {
      "label": "B",
      "direction": "상혁과 민정은 의용을, 의용은 노트북을 바라봄. 프롬프트에 명시된 각 인물의 시선 및 행동 지시와 불일치함.",
      "built_space": "배경이 없는 흰색 스튜디오 공간으로, 반드시 따라야 할 로케이션 레퍼런스를 완전히 누락함.",
      "entities": "왼쪽에 상혁, 중앙에 의용, 오른쪽에 민정이 배치되어 프롬프트와 반대임. 노트북을 조작하는 인물도 상혁이 아닌 의용임.",
      "hard_violations": [],
      "physics": "인물들은 의자에 정상적으로 앉아 있고, 사물 역시 테이블의 지지를 받음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 로케이션 레퍼런스를 무시한 임의의 거실을 묘사했으며, 상혁과 의용의 위치 및 역할(노트북 조작)이 프롬프트와 반대로 적용됨."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "로케이션 공간을 아예 누락한 흰색 배경이며, 인물들의 자리 배치와 행동이 프롬프트의 지시와 전혀 일치하지 않음."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "상혁과 민정은 의용과 노트북을 향하고, 의용은 노트북 화면을 응시함. 의용이 방을 향하고 상혁이 노트북을 조작해야 하는 프롬프트 지시와 다름.",
      "built_space": "소파와 책장이 있는 서구식 거실 구조로, 제시된 한국식 좌식 로케이션 레퍼런스를 완전히 무시함.",
      "entities": "왼쪽에 상혁, 중앙에 의용, 오른쪽에 민정이 위치함. 프롬프트의 배치(왼쪽 의용, 중앙 상혁)와 어긋남. 김이 나는 커피잔과 노트북은 존재함.",
      "hard_violations": [],
      "physics": "인물들은 의자와 소파에 앉아 지탱되고, 사물은 테이블 위에 정상적으로 놓여 있음."
     },
     {
      "label": "A",
      "direction": "상혁과 민정은 의용을, 의용은 노트북을 바라봄. 프롬프트에 명시된 각 인물의 시선 및 행동 지시와 불일치함.",
      "built_space": "배경이 없는 흰색 스튜디오 공간으로, 반드시 따라야 할 로케이션 레퍼런스를 완전히 누락함.",
      "entities": "왼쪽에 상혁, 중앙에 의용, 오른쪽에 민정이 배치되어 프롬프트와 반대임. 노트북을 조작하는 인물도 상혁이 아닌 의용임.",
      "hard_violations": [],
      "physics": "인물들은 의자에 정상적으로 앉아 있고, 사물 역시 테이블의 지지를 받음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1252,
     "B": 503
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "initial_roll_all_fail": true,
  "totals": {
   "A": 1252,
   "B": 503
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 500,
    "verdict_ko": "인물의 위치가 프롬프트와 반대로 배치되었고, 지정된 장소 참조와 전혀 다른 임의의 거실을 배경으로 생성하여 위치 조건을 위반했습니다.  ★위반: [gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜) / [gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함 / [gemini-pro] 장소 참조 이미지 위반 (구조와 배경이 전혀 다름) / [openrouter:x-ai/grok-4.6] 민정이 조사관들 맞은편이 아니라 같은 소파 쪽에 앉아 스테이징이 정한 좌석을 위반한다."
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "인물의 위치가 프롬프트와 반대로 배치되었으며, 지정된 장소를 완전히 무시하고 스튜디오 배경으로 렌더링하여 심각한 위반입니다.  ★위반: [gemini-pro] 지정된 인물 배치 위반 (나상혁과 서의용의 좌/중앙 위치가 뒤바뀜) / [gemini-pro] 노트북을 나상혁이 아닌 서의용이 사용함 / [gemini-pro] 장소 참조 이미지 위반 (공간적 맥락이 없는 스튜디오 배경)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L04B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 민정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:803682>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "지정된 배경 이미지(가정집 거실)가 완전히 무시되고 단색 스튜디오 배경으로 렌더링되었습니다.",
     "fix_en": "Replace the solid white background with the specified family home interior, featuring a kitchen and hallway behind the characters to match the reference. Keep the three people, their clothing and positions, the table, the laptop, the mug, the lighting, and the framing exactly as they are.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "프롬프트의 지시(서의용이 왼쪽, 나상혁이 중앙 노트북 앞)와 달리 서의용(비니 쓴 남성)과 나상혁(네이비 재킷 남성)의 위치가 서로 뒤바뀌었습니다.",
     "fix_en": "Redraw the left man to be Seo Eui-yong wearing a beanie and leather jacket, and redraw the center man to be Na Sang-hyeok wearing a navy bomber jacket and light shirt. Keep their overall poses, the woman on the right, the table, the laptop, the mug, the framing, and the current background exactly as they are.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "민정이 서의용이 아니라 왼쪽 나상혁을 바라보고 있어 시선 관계가 틀렸다.",
     "fix_en": "Adjust the woman's head and eye direction so she looks toward the man in the center instead of the man on the left. Maintain the identities and positions of all three people, their clothing, the table, the props, the framing, and the background exactly as they are.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": false
    },
    {
     "issue_ko": "테이블·컵·노트북이 지정된 어두운 나무 테이블·흰 머그·은색 노트북과 다르다.",
     "fix_en": "Change the coffee mug to plain white, the laptop to metallic silver, and the table surface to dark reddish-brown wood. Keep the three people, their clothing, their positions, the mug's steam, the framing, the lighting, and the background exactly as they are.",
     "severity": "major",
     "observation_index": 6,
     "needs_regeneration": false
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "지정된 배경 이미지(가정집 거실)가 완전히 무시되고 단색 스튜디오 배경으로 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "프롬프트의 지시(서의용이 왼쪽, 나상혁이 중앙 노트북 앞)와 달리 서의용(비니 쓴 남성)과 나상혁(네이비 재킷 남성)의 위치가 서로 뒤바뀌었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경이 잠긴 거실·부엌 실내가 아니라 흰 스튜디오 벽으로 바뀌어 장소가 완전히 사라졌다.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용과 나상혁의 자리·신원 의상이 바뀌어 왼쪽은 나상혁, 가운데는 서의용이 앉아 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "노트북 화면이 나상혁(왼쪽)이 아니라 가운데 서의용 쪽을 향하고 그가 노트북을 보고 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "민정이 서의용이 아니라 왼쪽 나상혁을 바라보고 있어 시선 관계가 틀렸다.",
     "severity": "major"
    },
    {
     "issue_ko": "테이블·컵·노트북이 지정된 어두운 나무 테이블·흰 머그·은색 노트북과 다르다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "민정이 서의용이 아니라 왼쪽 나상혁을 바라보고 있어 시선 관계가 틀렸다.",
    "fix_en": "Adjust the woman's head and eye direction so she looks toward the man in the center instead of the man on the left. Maintain the identities and positions of all three people, their clothing, the table, the props, the framing, and the background exactly as they are.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": false
   },
   {
    "issue_ko": "테이블·컵·노트북이 지정된 어두운 나무 테이블·흰 머그·은색 노트북과 다르다.",
    "fix_en": "Change the coffee mug to plain white, the laptop to metallic silver, and the table surface to dark reddish-brown wood. Keep the three people, their clothing, their positions, the mug's steam, the framing, the lighting, and the background exactly as they are.",
    "severity": "major",
    "observation_index": 6,
    "needs_regeneration": false
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 6,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the solid white background with the specified family home interior, featuring a kitchen and hallway behind the characters to match the reference. Keep the three people, their clothing and positions, the table, the laptop, the mug, the lighting, and the framing exactly as they are.\n- Redraw the left man to be Seo Eui-yong wearing a beanie and leather jacket, and redraw the center man to be Na Sang-hyeok wearing a navy bomber jacket and light shirt. Keep their overall poses, the woman on the right, the table, the laptop, the mug, the framing, and the current background exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "지정된 실내 배경을 완전히 누락하고 빈 스튜디오로 처리했으며, 서의용과 나상혁의 좌석 위치를 뒤바꾸는 치명적인 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "원본 배경을 무시하고 임의의 주방을 창조했으며, 마주 앉아야 할 인물들을 모두 세워둔 채 카메라 정면을 응시하게 만든 심각한 위반이 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽의 나상혁은 가운데 서의용을 향하고, 서의용은 랩탑을 보며, 오른쪽 민정은 서의용을 바라봄.",
      "built_space": "요구된 레퍼런스 실내 공간이 완전히 사라지고 하얀 무배경에 테이블과 의자만 덩그러니 놓여 있음.",
      "entities": "나상혁(왼쪽), 서의용(가운데), 민정(오른쪽)의 외형과 의상은 레퍼런스와 일치하나, 서의용과 나상혁의 위치가 프롬프트 지시와 반대로 배치됨. 김이 나는 커피잔과 열린 랩탑 존재.",
      "hard_violations": [
       "지정된 장소/배경 완전 누락 및 무배경(하얀 캔버스) 처리",
       "서의용과 나상혁의 좌석 위치가 프롬프트 및 스케치 지시와 반대로 뒤바뀜"
      ],
      "physics": "세 인물 모두 의자에 안정적으로 앉아 있으며 손과 팔의 지지 상태는 자연스러움."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 렌즈를 정면으로 꼿꼿이 응시하고 있음.",
      "built_space": "레퍼런스와 전혀 일치하지 않는 임의의 현대식 주방 배경이 렌더링됨.",
      "entities": "서의용(왼쪽), 나상혁(가운데), 민정(오른쪽)의 외형은 레퍼런스와 일치. 커피잔과 랩탑이 테이블에 있음.",
      "hard_violations": [
       "지정된 장소/배경을 무시하고 완전히 다른 장소(임의의 주방) 생성",
       "마주 앉아야 하는 지시를 무시하고 세 인물 모두 서 있음",
       "자연스러운 연기 지시를 위반하고 카메라를 정면 응시함"
      ],
      "physics": "세 사람 모두 테이블 뒤 바닥에 서서 체중을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "지정된 실내 배경을 완전히 누락하고 빈 스튜디오로 처리했으며, 서의용과 나상혁의 좌석 위치를 뒤바꾸는 치명적인 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "원본 배경을 무시하고 임의의 주방을 창조했으며, 마주 앉아야 할 인물들을 모두 세워둔 채 카메라 정면을 응시하게 만든 심각한 위반이 있습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽의 나상혁은 가운데 서의용을 향하고, 서의용은 랩탑을 보며, 오른쪽 민정은 서의용을 바라봄.",
      "built_space": "요구된 레퍼런스 실내 공간이 완전히 사라지고 하얀 무배경에 테이블과 의자만 덩그러니 놓여 있음.",
      "entities": "나상혁(왼쪽), 서의용(가운데), 민정(오른쪽)의 외형과 의상은 레퍼런스와 일치하나, 서의용과 나상혁의 위치가 프롬프트 지시와 반대로 배치됨. 김이 나는 커피잔과 열린 랩탑 존재.",
      "hard_violations": [
       "지정된 장소/배경 완전 누락 및 무배경(하얀 캔버스) 처리",
       "서의용과 나상혁의 좌석 위치가 프롬프트 및 스케치 지시와 반대로 뒤바뀜"
      ],
      "physics": "세 인물 모두 의자에 안정적으로 앉아 있으며 손과 팔의 지지 상태는 자연스러움."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 렌즈를 정면으로 꼿꼿이 응시하고 있음.",
      "built_space": "레퍼런스와 전혀 일치하지 않는 임의의 현대식 주방 배경이 렌더링됨.",
      "entities": "서의용(왼쪽), 나상혁(가운데), 민정(오른쪽)의 외형은 레퍼런스와 일치. 커피잔과 랩탑이 테이블에 있음.",
      "hard_violations": [
       "지정된 장소/배경을 무시하고 완전히 다른 장소(임의의 주방) 생성",
       "마주 앉아야 하는 지시를 무시하고 세 인물 모두 서 있음",
       "자연스러운 연기 지시를 위반하고 카메라를 정면 응시함"
      ],
      "physics": "세 사람 모두 테이블 뒤 바닥에 서서 체중을 지탱하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "지시된 앉은 자세와 상호작용하는 시선 구도는 어느 정도 반영했으나, 지정된 필수 배경을 완전히 누락하고 두 남성 캐릭터의 위치를 뒤바꾸어 최종 사용이 불가합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "필수 배경을 무시하고 임의의 주방을 생성했으며, 프롬프트에서 금지한 '정면을 응시하며 일렬로 선 구도'를 연출하여 완전히 실패한 결과물입니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "좌측의 나상혁과 우측의 민정은 중앙의 서의용을 바라보고 있으며, 중앙의 서의용은 노트북 화면을 응시하고 있음.",
      "built_space": "제공된 배경 레퍼런스가 완전히 배제되고 텅 빈 회색 스튜디오 배경으로 렌더링됨. 세 명의 인물이 테이블 주위에 앉아 있음.",
      "entities": "민정, 서의용, 나상혁의 외모와 의상은 레퍼런스와 일치하나, 텍스트가 지시한 서의용(좌측)과 나상혁(중앙)의 위치가 서로 뒤바뀌어 있음.",
      "hard_violations": [
       "지정된 레퍼런스 배경(Location lock) 완전히 삭제 및 스튜디오 배경으로 대체",
       "서의용과 나상혁의 지정된 배치 위치 오류"
      ],
      "physics": "인물들은 의자에 앉아 체중을 지지하고 있으며, 컵과 노트북은 테이블 위에 안정적으로 놓여 있음. 노트북은 열린 상태로 유지됨."
     },
     {
      "label": "A",
      "direction": "세 인물 모두 카메라 정면을 똑바로 응시하고 있음. 상호작용하는 시선이 부재함.",
      "built_space": "제공된 레퍼런스와 전혀 다른 구조의 현대식 주방이 배경으로 등장함. 인물들은 의자 없이 테이블 뒤에 서 있음.",
      "entities": "민정, 서의용, 나상혁의 외형 및 의상은 지정된 캐릭터 레퍼런스를 잘 반영함.",
      "hard_violations": [
       "지정된 레퍼런스 배경 무시 및 새로운 배경 창조",
       "인물들이 앉아있지 않고 서 있는 자세(Staging 위반)",
       "프롬프트가 금지한 정면 일렬 배치(frontal lineup) 및 렌즈 응시"
      ],
      "physics": "인물들은 두 발로 바닥에 서서 몸을 지탱하고 있으며, 커피잔과 닫혀있는 듯한 노트북이 테이블에 얹혀 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 앉은 자세와 상호작용하는 시선 구도는 어느 정도 반영했으나, 지정된 필수 배경을 완전히 누락하고 두 남성 캐릭터의 위치를 뒤바꾸어 최종 사용이 불가합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "필수 배경을 무시하고 임의의 주방을 생성했으며, 프롬프트에서 금지한 '정면을 응시하며 일렬로 선 구도'를 연출하여 완전히 실패한 결과물입니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "좌측의 나상혁과 우측의 민정은 중앙의 서의용을 바라보고 있으며, 중앙의 서의용은 노트북 화면을 응시하고 있음.",
      "built_space": "제공된 배경 레퍼런스가 완전히 배제되고 텅 빈 회색 스튜디오 배경으로 렌더링됨. 세 명의 인물이 테이블 주위에 앉아 있음.",
      "entities": "민정, 서의용, 나상혁의 외모와 의상은 레퍼런스와 일치하나, 텍스트가 지시한 서의용(좌측)과 나상혁(중앙)의 위치가 서로 뒤바뀌어 있음.",
      "hard_violations": [
       "지정된 레퍼런스 배경(Location lock) 완전히 삭제 및 스튜디오 배경으로 대체",
       "서의용과 나상혁의 지정된 배치 위치 오류"
      ],
      "physics": "인물들은 의자에 앉아 체중을 지지하고 있으며, 컵과 노트북은 테이블 위에 안정적으로 놓여 있음. 노트북은 열린 상태로 유지됨."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 카메라 정면을 똑바로 응시하고 있음. 상호작용하는 시선이 부재함.",
      "built_space": "제공된 레퍼런스와 전혀 다른 구조의 현대식 주방이 배경으로 등장함. 인물들은 의자 없이 테이블 뒤에 서 있음.",
      "entities": "민정, 서의용, 나상혁의 외형 및 의상은 지정된 캐릭터 레퍼런스를 잘 반영함.",
      "hard_violations": [
       "지정된 레퍼런스 배경 무시 및 새로운 배경 창조",
       "인물들이 앉아있지 않고 서 있는 자세(Staging 위반)",
       "프롬프트가 금지한 정면 일렬 배치(frontal lineup) 및 렌즈 응시"
      ],
      "physics": "인물들은 두 발로 바닥에 서서 몸을 지탱하고 있으며, 커피잔과 닫혀있는 듯한 노트북이 테이블에 얹혀 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S43sh1__bgfirst_bg.png",
   "bg_asset_id": "eb3bac51-c07b-4025-9b87-b32cb09b025e",
   "bg_record_key": "S43sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S43sh1::cine": {
  "applied": true,
  "fingerprint": "8510a7aeac1700ed0fe5bed2747bb6505c6ae6654efa1bd4b3cec25cbc25f535",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S43sh1_sel.png",
  "source_sha256": "10c632acad5412dfd6769817c2859730d647ce5e4331fda196f9e1cf5eb83c80",
  "file": "S43sh1_cine.png",
  "latency_ms": 12123
 },
 "S43sh12::signage": {
  "fp": "4e012930b6d2baf7",
  "inscriptions": [
   {
    "surface_native": "폴라로이드 사진 하단 여백",
    "text_native": "우리 아빠랑 ♥ (2003. 5)",
    "reason_ko": "폴라로이드 사진 하단 여백에 삐뚤빼뚤하게 적힌 손글씨는 가족의 추억이 담긴 소중한 소품임을 사실적으로 드러내 줍니다."
   }
  ]
 },
 "S43sh12": {
  "input_fingerprint": "d588af9a12311788",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 손에 들린, 환하게 웃고 있는 김선영과 선영의 아버지의 얼굴이 찍힌 폴라로이드 사진 표면 클로즈업\n\nLOCATION (lock): At the sitting-room table inside the family home, beside the opened box of keepsakes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From very close above Seo Eui-yong's hands, end the push at a near-perpendicular angle to the Polaroid while keeping the photograph to roughly the central two-fifths of the frame, bordered by his fingers and traces of the keepsake box for scale. Within the photograph, Kim Seon-yeong and her father hold each other closely and smile brightly; their faces are the sharp visual focus while the live hand-held context remains visible around the print.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Polaroid held flat beneath the camera in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Polaroid photograph (held in Seo Eui-yong's hands) — Its image-bearing front face is nearly perpendicular to the lens and shows Kim Seon-yeong embracing her smiling father; used as Primary content surface carrying the smiling portrait while remaining bordered by hands and surrounding context; keepsake box (open with notebooks, diary, photographs, and letters inside) — Only portions around its opening are visible beyond the photograph and hands; used as Provides peripheral evidence that the photograph came from the stored belongings rather than appearing as a freestanding image.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light renders the physical print and hands with subdued neutral-to-cool surrounding tones while preserving the photographed smiles clearly.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same modest home interior, table area, daylight, and the detective's seated position. Exclude the three faces and coffee cups from the insert, showing only the polaroid held in his hand.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The keepsake box remains open with Seonyeong's notebooks, diary, photographs, and letters inside, while Euiyong holds the Polaroid of Seonyeong embracing her father.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 폴라로이드 사진 하단 여백: \"우리 아빠랑 ♥ (2003. 5)\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 손에 들린, 환하게 웃고 있는 김선영과 선영의 아버지의 얼굴이 찍힌 폴라로이드 사진 표면 클로즈업\n\nLOCATION (lock): At the sitting-room table inside the family home, beside the opened box of keepsakes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From very close above Seo Eui-yong's hands, end the push at a near-perpendicular angle to the Polaroid while keeping the photograph to roughly the central two-fifths of the frame, bordered by his fingers and traces of the keepsake box for scale. Within the photograph, Kim Seon-yeong and her father hold each other closely and smile brightly; their faces are the sharp visual focus while the live hand-held context remains visible around the print.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Polaroid held flat beneath the camera in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Polaroid photograph (held in Seo Eui-yong's hands) — Its image-bearing front face is nearly perpendicular to the lens and shows Kim Seon-yeong embracing her smiling father; used as Primary content surface carrying the smiling portrait while remaining bordered by hands and surrounding context; keepsake box (open with notebooks, diary, photographs, and letters inside) — Only portions around its opening are visible beyond the photograph and hands; used as Provides peripheral evidence that the photograph came from the stored belongings rather than appearing as a freestanding image.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light renders the physical print and hands with subdued neutral-to-cool surrounding tones while preserving the photographed smiles clearly.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same modest home interior, table area, daylight, and the detective's seated position. Exclude the three faces and coffee cups from the insert, showing only the polaroid held in his hand.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The keepsake box remains open with Seonyeong's notebooks, diary, photographs, and letters inside, while Euiyong holds the Polaroid of Seonyeong embracing her father.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 폴라로이드 사진 하단 여백: \"우리 아빠랑 ♥ (2003. 5)\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 손에 들린, 환하게 웃고 있는 김선영과 선영의 아버지의 얼굴이 찍힌 폴라로이드 사진 표면 클로즈업\n\nLOCATION (lock): At the sitting-room table inside the family home, beside the opened box of keepsakes. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From very close above Seo Eui-yong's hands, end the push at a near-perpendicular angle to the Polaroid while keeping the photograph to roughly the central two-fifths of the frame, bordered by his fingers and traces of the keepsake box for scale. Within the photograph, Kim Seon-yeong and her father hold each other closely and smile brightly; their faces are the sharp visual focus while the live hand-held context remains visible around the print.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: Polaroid held flat beneath the camera in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Polaroid photograph (held in Seo Eui-yong's hands) — Its image-bearing front face is nearly perpendicular to the lens and shows Kim Seon-yeong embracing her smiling father; used as Primary content surface carrying the smiling portrait while remaining bordered by hands and surrounding context; keepsake box (open with notebooks, diary, photographs, and letters inside) — Only portions around its opening are visible beyond the photograph and hands; used as Provides peripheral evidence that the photograph came from the stored belongings rather than appearing as a freestanding image.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light renders the physical print and hands with subdued neutral-to-cool surrounding tones while preserving the photographed smiles clearly.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same modest home interior, table area, daylight, and the detective's seated position. Exclude the three faces and coffee cups from the insert, showing only the polaroid held in his hand.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The keepsake box remains open with Seonyeong's notebooks, diary, photographs, and letters inside, while Euiyong holds the Polaroid of Seonyeong embracing her father.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 폴라로이드 사진 하단 여백: \"우리 아빠랑 ♥ (2003. 5)\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 수직에 가까운 각도로 두 손에 들린 폴라로이드 사진을 내려다보고 있습니다.",
    "built_space": "나무 테이블 위에 종이 상자가 놓여 있고, 상자 안에는 낡은 책과 사진들이 들어 있습니다.",
    "entities": "서의용의 양손(가죽 재킷 소매 착용)이 폴라로이드 사진을 들고 있습니다. 사진 속 김선영과 아버지의 모습이 레퍼런스와 일치하며, 하단 여백에 \"우리 아빠랑 ♥ (2003. 5)\"라는 텍스트가 정확히 적혀 있습니다.",
    "hard_violations": [],
    "physics": "폴라로이드 사진은 두 손에 의해 단단히 쥐어져 지탱되고 있으며, 상자는 테이블 표면 위에 놓여 있습니다."
   },
   {
    "label": "B",
    "direction": "카메라는 수직에 가까운 각도로 두 손에 들린 폴라로이드 사진을 내려다보고 있습니다.",
    "built_space": "나무 테이블 위에 상자가 있으며, 상자 안팎으로 책, 편지, 사진들이 흩어져 있습니다.",
    "entities": "서의용의 양손이 사진을 들고 있습니다. 사진의 피사체는 레퍼런스와 일치하나 하단 텍스트의 띄어쓰기(\"2003.5\")가 약간 다릅니다. 상자 안의 책 표지에 지시되지 않은 영문(\"FBOI...\")이 적혀 있습니다.",
    "hard_violations": [
     "배경의 책 표지에 프롬프트에서 금지한 '지시되지 않은 읽을 수 있는 텍스트(영문 알파벳)'가 추가됨"
    ],
    "physics": "사진은 손에 들려 있고, 주변의 소품과 상자는 테이블 위에 중력에 맞게 놓여 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 10,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 10,
    "verdict_ko": "프롬프트가 요구한 인서트 클로즈업 구도와 피사체(가죽 재킷 소매, 보관함)를 정확히 구현했으며, 폴라로이드 안의 인물과 지정된 텍스트(\"우리 아빠랑 ♥ (2003. 5)\")를 오차 없이 완벽하게 표현했습니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "폴라로이드의 인물과 형태는 훌륭하나, 배경에 있는 책에 지시되지 않은 읽을 수 있는 영문 알파벳(\"FBOI...\")이 포함되어 '다른 텍스트 추가 금지' 지침을 어겼고 지정된 텍스트의 띄어쓰기 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S43sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "PROP REFERENCE — 김선영과 선영의 아버지의 폴라로이드 사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:885953>"
   }
  ],
  "critique": {
   "issues": [],
   "observer_observations": [],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 0,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_skipped": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S43sh1"
  }
 },
 "S43sh12::cine": {
  "applied": true,
  "fingerprint": "7dcbf10fb43290da0ce045176665e667d433468661846d56c6adac6faec8397b",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S43sh12_sel.png",
  "source_sha256": "cd90dbe2a1cc0752a5b90b7eedd48751e050a6510936c95c4a939dc513f92b9f",
  "file": "S43sh12_cine.png",
  "latency_ms": 9570
 },
 "S44sh2::signage": {
  "fp": "3fd5c06dba62b41c",
  "inscriptions": [
   {
    "surface_native": "동화책 표지",
    "text_native": "숲속의 친구들",
    "reason_ko": "주인공이 유치원 교실에서 들고 있는 동화책에 자연스러운 한국어 제목을 부여하여 장면의 상황적 맥락을 살립니다."
   }
  ]
 },
 "era_assess::f1a908b3274e892e": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 대한민국의 유치원 교실 내부",
    "search_terms_native": [
     "유치원 교실 내부",
     "어린이집 교실",
     "유치원 언어영역"
    ],
    "language_lock_native": "검색어는 오직 한국어로만 작성되어야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대와 2010년대 한국 유치원의 바닥 난방 매트, 낮은 목재 가구, 한글 벽보 및 교구 배치 형태는 서구식 교실과 뚜렷이 구분됩니다."
   }
  ]
 },
 "era_ref::31aaf6e61be90839": {
  "subject": "2000년대-2010년대 대한민국의 유치원 교실 내부",
  "terms": [
   "유치원 교실 내부",
   "어린이집 교실",
   "유치원 언어영역"
  ],
  "queries": [
   [
    "2000년대 2010년대 대한민국 유치원 교실 내부 어린이집 교실 언어영역",
    "옛날 유치원 언어영역 교실 내부 대한민국"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://saeonbit.com/assets/images/kakaotalk-20220718-201405967.jpg",
  "picked_reason_ko": "2번은 2001~2017년 대한민국의 일상적인 유치원 교실을 넓고 선명하게 보여 주며, 아동용 책상·의자, 낮은 교구장, 영역별 학습 자료와 실내 마감까지 가장 쉽게 파악할 수 있다.",
  "sha256": "187fc99194c3c125e91a6e94b79da46358c95a57045b0f6d467165ab717896cb",
  "file": "eraref_31aaf6e61be90839.png"
 },
 "S44sh2::bgfirst_bg": {
  "input_fingerprint": "51b65345020095ff",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 동화책 너머 허공을 향해 초점 잃은 시선을 던진 채 멍하니 굳어버린 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): Inside a kindergarten classroom, in the story-reading area with the children.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the children's side of the open storybook, finish the approach just below Lee Mi-gyeong's eye height in a tight oblique three-quarter close-up. Retain the book's upper edge as a soft lower-foreground line while her hands stay suspended in the reading posture, her face rigid and her unfocused gaze passing over the pages into empty space beyond them.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: open storybook upper edge in the lower-center of the frame, foreground; Lee Mi-gyeong isolated in close-up in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: open storybook (open in Lee Mi-gyeong's hands) — The open page faces Lee Mi-gyeong and the children; the camera sees only the upper portion obliquely, with no specific page content established; used as Its upper edge remains in the near foreground as the interrupted task from which her attention has drifted; children's side of the reading area (children are being read to); used as Kept outside the tight facial focus, establishing the direction from which the book was being read.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light holds her face in restrained neutral-to-cool tones with moderate contrast and no subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 대한민국의 유치원 교실 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 동화책 너머 허공을 향해 초점 잃은 시선을 던진 채 멍하니 굳어버린 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): Inside a kindergarten classroom, in the story-reading area with the children.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the children's side of the open storybook, finish the approach just below Lee Mi-gyeong's eye height in a tight oblique three-quarter close-up. Retain the book's upper edge as a soft lower-foreground line while her hands stay suspended in the reading posture, her face rigid and her unfocused gaze passing over the pages into empty space beyond them.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: open storybook upper edge in the lower-center of the frame, foreground; Lee Mi-gyeong isolated in close-up in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: open storybook (open in Lee Mi-gyeong's hands) — The open page faces Lee Mi-gyeong and the children; the camera sees only the upper portion obliquely, with no specific page content established; used as Its upper edge remains in the near foreground as the interrupted task from which her attention has drifted; children's side of the reading area (children are being read to); used as Kept outside the tight facial focus, establishing the direction from which the book was being read.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light holds her face in restrained neutral-to-cool tones with moderate contrast and no subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 대한민국의 유치원 교실 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S44sh2__bgfirst_bg.png",
  "asset_id": "fe3f2d7a-86ce-4c2b-9f8f-296c78610b98",
  "input_asset_ids": [
   "b5743136-bfb8-40b8-8d32-7afa1433510a",
   "789e0f09-95c5-44f5-9185-0115583e88e6"
  ],
  "era_research": {
   "subject": "2000년대-2010년대 대한민국의 유치원 교실 내부",
   "queries": [
    [
     "2000년대 2010년대 대한민국 유치원 교실 내부 어린이집 교실 언어영역",
     "옛날 유치원 언어영역 교실 내부 대한민국"
    ]
   ],
   "picked_url": "https://saeonbit.com/assets/images/kakaotalk-20220718-201405967.jpg",
   "sha256": "187fc99194c3c125e91a6e94b79da46358c95a57045b0f6d467165ab717896cb",
   "file": "eraref_31aaf6e61be90839.png"
  }
 },
 "S44sh2": {
  "input_fingerprint": "c8661eb80f75a2c3",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 동화책 너머 허공을 향해 초점 잃은 시선을 던진 채 멍하니 굳어버린 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): Inside a kindergarten classroom, in the story-reading area with the children. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the children's side of the open storybook, finish the approach just below Lee Mi-gyeong's eye height in a tight oblique three-quarter close-up. Retain the book's upper edge as a soft lower-foreground line while her hands stay suspended in the reading posture, her face rigid and her unfocused gaze passing over the pages into empty space beyond them.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: open storybook upper edge in the lower-center of the frame, foreground; Lee Mi-gyeong isolated in close-up in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: open storybook (open in Lee Mi-gyeong's hands) — The open page faces Lee Mi-gyeong and the children; the camera sees only the upper portion obliquely, with no specific page content established; used as Its upper edge remains in the near foreground as the interrupted task from which her attention has drifted; children's side of the reading area (children are being read to); used as Kept outside the tight facial focus, establishing the direction from which the book was being read.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light holds her face in restrained neutral-to-cool tones with moderate contrast and no subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 동화책 표지: \"숲속의 친구들\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 동화책 너머 허공을 향해 초점 잃은 시선을 던진 채 멍하니 굳어버린 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): Inside a kindergarten classroom, in the story-reading area with the children. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the children's side of the open storybook, finish the approach just below Lee Mi-gyeong's eye height in a tight oblique three-quarter close-up. Retain the book's upper edge as a soft lower-foreground line while her hands stay suspended in the reading posture, her face rigid and her unfocused gaze passing over the pages into empty space beyond them.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: open storybook upper edge in the lower-center of the frame, foreground; Lee Mi-gyeong isolated in close-up in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: open storybook (open in Lee Mi-gyeong's hands) — The open page faces Lee Mi-gyeong and the children; the camera sees only the upper portion obliquely, with no specific page content established; used as Its upper edge remains in the near foreground as the interrupted task from which her attention has drifted; children's side of the reading area (children are being read to); used as Kept outside the tight facial focus, establishing the direction from which the book was being read.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light holds her face in restrained neutral-to-cool tones with moderate contrast and no subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 동화책 표지: \"숲속의 친구들\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 동화책 너머 허공을 향해 초점 잃은 시선을 던진 채 멍하니 굳어버린 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): Inside a kindergarten classroom, in the story-reading area with the children. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the children's side of the open storybook, finish the approach just below Lee Mi-gyeong's eye height in a tight oblique three-quarter close-up. Retain the book's upper edge as a soft lower-foreground line while her hands stay suspended in the reading posture, her face rigid and her unfocused gaze passing over the pages into empty space beyond them.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: open storybook upper edge in the lower-center of the frame, foreground; Lee Mi-gyeong isolated in close-up in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: open storybook (open in Lee Mi-gyeong's hands) — The open page faces Lee Mi-gyeong and the children; the camera sees only the upper portion obliquely, with no specific page content established; used as Its upper edge remains in the near foreground as the interrupted task from which her attention has drifted; children's side of the reading area (children are being read to); used as Kept outside the tight facial focus, establishing the direction from which the book was being read.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light holds her face in restrained neutral-to-cool tones with moderate contrast and no subjective visual distortion.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 동화책 표지: \"숲속의 친구들\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S44sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S44sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L48B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "시선이 허공이 아닌 책을 향하고 명찰에 텍스트가 유출되는 오류가 있으나, A에 비해 정상적인 인체 표현을 유지했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "눈동자가 하얗게 변한 심각한 해부학적 오류가 있으며, 이는 프롬프트의 지시사항을 정면으로 위반한 결과입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선이 정면을 향하고 있으나, 눈동자가 하얗게 지워져 정확한 타겟을 특정할 수 없음.",
      "built_space": "창문, 교구장, 우측 싱크대 등 레퍼런스의 유치원 교실 공간 구조와 일치함.",
      "entities": "이미경의 의상(분홍색 가디건, 흰 블라우스)과 외모가 일치함. 책에 '숲속의 친구들' 텍스트가 명확히 보임.",
      "hard_violations": [
       "해부학적으로 불가능한 신체 (지시사항을 위반하여 눈동자 없이 하얗게 렌더링된 눈)"
      ],
      "physics": "동화책이 공중에 떠 있지 않고 아래에서 지지되고 있으나(프레임상 손은 잘림), 자연스럽게 위치함."
     },
     {
      "label": "B",
      "direction": "시선이 허공을 향하지 않고 아래쪽 동화책을 바라보고 있음.",
      "built_space": "교구장과 배경 창문이 레퍼런스의 교실 환경과 일치하게 배치됨.",
      "entities": "이미경의 의상과 외모 일치. 동화책에 '숲속의 친구들' 텍스트가 있음. 가슴에 요구되지 않은 명찰이 있음.",
      "hard_violations": [
       "텍스트 유출 (가슴 명찰에 프롬프트 텍스트인 '숲속의'가 잘못 유출되어 표기됨)"
      ],
      "physics": "양손이 동화책의 양쪽을 안정적으로 잡고 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "시선이 허공이 아닌 책을 향하고 명찰에 텍스트가 유출되는 오류가 있으나, A에 비해 정상적인 인체 표현을 유지했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "눈동자가 하얗게 변한 심각한 해부학적 오류가 있으며, 이는 프롬프트의 지시사항을 정면으로 위반한 결과입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 정면을 향하고 있으나, 눈동자가 하얗게 지워져 정확한 타겟을 특정할 수 없음.",
      "built_space": "창문, 교구장, 우측 싱크대 등 레퍼런스의 유치원 교실 공간 구조와 일치함.",
      "entities": "이미경의 의상(분홍색 가디건, 흰 블라우스)과 외모가 일치함. 책에 '숲속의 친구들' 텍스트가 명확히 보임.",
      "hard_violations": [
       "해부학적으로 불가능한 신체 (지시사항을 위반하여 눈동자 없이 하얗게 렌더링된 눈)"
      ],
      "physics": "동화책이 공중에 떠 있지 않고 아래에서 지지되고 있으나(프레임상 손은 잘림), 자연스럽게 위치함."
     },
     {
      "label": "B",
      "direction": "시선이 허공을 향하지 않고 아래쪽 동화책을 바라보고 있음.",
      "built_space": "교구장과 배경 창문이 레퍼런스의 교실 환경과 일치하게 배치됨.",
      "entities": "이미경의 의상과 외모 일치. 동화책에 '숲속의 친구들' 텍스트가 있음. 가슴에 요구되지 않은 명찰이 있음.",
      "hard_violations": [
       "텍스트 유출 (가슴 명찰에 프롬프트 텍스트인 '숲속의'가 잘못 유출되어 표기됨)"
      ],
      "physics": "양손이 동화책의 양쪽을 안정적으로 잡고 지지하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "시선이 책을 향하고 이름표에 불필요한 문자가 있으나, 요구된 3/4 측면 구도를 상대적으로 잘 따랐습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "정면 구도를 사용하였으며, 명시적으로 금지된 해부학적 눈 변형(하얀 눈동자)을 묘사하여 지시를 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선이 허공이 아닌 아래쪽 동화책 페이지를 명확히 향하고 있음.",
      "built_space": "참조 이미지와 일치하는 유치원 교실 배경(창문, 교구장)이 올바르게 묘사됨.",
      "entities": "이미경의 인물 및 의상(분홍색 가디건, 흰색 셔츠)은 일치하나, 가슴에 지시되지 않은 텍스트가 적힌 이름표가 추가됨.",
      "hard_violations": [
       "지시되지 않은 임의의 텍스트(이름표) 생성"
      ],
      "physics": "양손으로 동화책을 양옆에서 잡고 안정적으로 지탱함."
     },
     {
      "label": "B",
      "direction": "정면을 향하고 있으나, 동공이 없어 정확한 시선 방향이나 대상을 확인할 수 없음.",
      "built_space": "참조 이미지의 유치원 교실 배경(교구장, 싱크대 등)이 위치에 맞게 배치됨.",
      "entities": "인물의 의상과 동화책 표지의 '숲속의 친구들' 텍스트는 일치하나, 정상적인 인간의 눈이 아님.",
      "hard_violations": [
       "물리적으로 불가능한 해부학적 구조 (홍채와 동공이 없는 하얀 눈)"
      ],
      "physics": "양손으로 펼쳐진 동화책의 아래쪽 가장자리를 받치고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "시선이 책을 향하고 이름표에 불필요한 문자가 있으나, 요구된 3/4 측면 구도를 상대적으로 잘 따랐습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "정면 구도를 사용하였으며, 명시적으로 금지된 해부학적 눈 변형(하얀 눈동자)을 묘사하여 지시를 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선이 허공이 아닌 아래쪽 동화책 페이지를 명확히 향하고 있음.",
      "built_space": "참조 이미지와 일치하는 유치원 교실 배경(창문, 교구장)이 올바르게 묘사됨.",
      "entities": "이미경의 인물 및 의상(분홍색 가디건, 흰색 셔츠)은 일치하나, 가슴에 지시되지 않은 텍스트가 적힌 이름표가 추가됨.",
      "hard_violations": [
       "지시되지 않은 임의의 텍스트(이름표) 생성"
      ],
      "physics": "양손으로 동화책을 양옆에서 잡고 안정적으로 지탱함."
     },
     {
      "label": "A",
      "direction": "정면을 향하고 있으나, 동공이 없어 정확한 시선 방향이나 대상을 확인할 수 없음.",
      "built_space": "참조 이미지의 유치원 교실 배경(교구장, 싱크대 등)이 위치에 맞게 배치됨.",
      "entities": "인물의 의상과 동화책 표지의 '숲속의 친구들' 텍스트는 일치하나, 정상적인 인간의 눈이 아님.",
      "hard_violations": [
       "물리적으로 불가능한 해부학적 구조 (홍채와 동공이 없는 하얀 눈)"
      ],
      "physics": "양손으로 펼쳐진 동화책의 아래쪽 가장자리를 받치고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 5,
     "B": 8
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선이 정면을 향하고 있으나, 눈동자가 하얗게 지워져 정확한 타겟을 특정할 수 없음.",
    "built_space": "창문, 교구장, 우측 싱크대 등 레퍼런스의 유치원 교실 공간 구조와 일치함.",
    "entities": "이미경의 의상(분홍색 가디건, 흰 블라우스)과 외모가 일치함. 책에 '숲속의 친구들' 텍스트가 명확히 보임.",
    "hard_violations": [
     "해부학적으로 불가능한 신체 (지시사항을 위반하여 눈동자 없이 하얗게 렌더링된 눈)"
    ],
    "physics": "동화책이 공중에 떠 있지 않고 아래에서 지지되고 있으나(프레임상 손은 잘림), 자연스럽게 위치함."
   },
   {
    "label": "B",
    "direction": "시선이 허공을 향하지 않고 아래쪽 동화책을 바라보고 있음.",
    "built_space": "교구장과 배경 창문이 레퍼런스의 교실 환경과 일치하게 배치됨.",
    "entities": "이미경의 의상과 외모 일치. 동화책에 '숲속의 친구들' 텍스트가 있음. 가슴에 요구되지 않은 명찰이 있음.",
    "hard_violations": [
     "텍스트 유출 (가슴 명찰에 프롬프트 텍스트인 '숲속의'가 잘못 유출되어 표기됨)"
    ],
    "physics": "양손이 동화책의 양쪽을 안정적으로 잡고 지지하고 있음."
   }
  ],
  "totals": {
   "A": 5,
   "B": 8
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "시선이 허공이 아닌 책을 향하고 명찰에 텍스트가 유출되는 오류가 있으나, A에 비해 정상적인 인체 표현을 유지했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "눈동자가 하얗게 변한 심각한 해부학적 오류가 있으며, 이는 프롬프트의 지시사항을 정면으로 위반한 결과입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L48B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "동화책 너머 허공을 향해야 할 시선이 책 페이지를 직접 내려다보고 있음.",
     "fix_en": "Redirect her pupils to look straight ahead and slightly up into the empty space beyond the book, creating a blank, unfocused gaze; strictly preserve her face, hairstyle, pink cardigan, the open book in her hands, and the classroom background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "명찰에 지시되지 않은 훼손된 텍스트('숲속의 전...5학')가 임의로 생성됨.",
     "fix_en": "Erase all garbled black text from the white rectangular name tag on her pink cardigan, leaving the white surface blank; preserve the name tag's shape, her clothing, facial features, the book, and the classroom background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "구체적인 페이지 내용이 보이지 않아야 한다는 카메라 지시와 달리, 펼쳐진 책의 내부와 삽화가 프레임에 넓게 노출됨.",
     "fix_en": "Reposition the book lower so only its upper edge is visible in the foreground, hiding the internal pages and illustrations.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "'숲속의 친구들' 텍스트가 동화책 표지가 아닌 펼쳐진 책의 내부 우측 페이지에 인쇄되어 있음.",
     "fix_en": "Remove the '숲속의 친구들' text from the right page of the open book.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "펼쳐진 책의 좌측 페이지에 훼손되고 의미를 알 수 없는 문자들이 임의로 생성됨.",
     "fix_en": "Remove the garbled text marks from the left page of the book.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "동화책 너머 허공을 향해야 할 시선이 책 페이지를 직접 내려다보고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "명찰에 지시되지 않은 훼손된 텍스트('숲속의 전...5학')가 임의로 생성됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "구체적인 페이지 내용이 보이지 않아야 한다는 카메라 지시와 달리, 펼쳐진 책의 내부와 삽화가 프레임에 넓게 노출됨.",
     "severity": "major"
    },
    {
     "issue_ko": "'숲속의 친구들' 텍스트가 동화책 표지가 아닌 펼쳐진 책의 내부 우측 페이지에 인쇄되어 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "펼쳐진 책의 좌측 페이지에 훼손되고 의미를 알 수 없는 문자들이 임의로 생성됨.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경이 동화책 너머 허공이 아니라 페이지를 또렷이 내려다보고 있어 초점 잃은 굳은 얼굴이 아님",
     "severity": "critical"
    },
    {
     "issue_ko": "펼친 속페이지에 표지 제목 「숲속의 친구들」과 그림이 구체적으로 선명히 보임",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 가슴 이름표 글자가 「숲속의 조래능악」처럼 깨져 있음",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 5,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "구체적인 페이지 내용이 보이지 않아야 한다는 카메라 지시와 달리, 펼쳐진 책의 내부와 삽화가 프레임에 넓게 노출됨.",
    "fix_en": "Reposition the book lower so only its upper edge is visible in the foreground, hiding the internal pages and illustrations.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "'숲속의 친구들' 텍스트가 동화책 표지가 아닌 펼쳐진 책의 내부 우측 페이지에 인쇄되어 있음.",
    "fix_en": "Remove the '숲속의 친구들' text from the right page of the open book.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "펼쳐진 책의 좌측 페이지에 훼손되고 의미를 알 수 없는 문자들이 임의로 생성됨.",
    "fix_en": "Remove the garbled text marks from the left page of the book.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redirect her pupils to look straight ahead and slightly up into the empty space beyond the book, creating a blank, unfocused gaze; strictly preserve her face, hairstyle, pink cardigan, the open book in her hands, and the classroom background.\n- Erase all garbled black text from the white rectangular name tag on her pink cardigan, leaving the white surface blank; preserve the name tag's shape, her clothing, facial features, the book, and the classroom background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "동화책 너머 허공을 향한 초점 잃은 시선과 멍한 표정을 지문 그대로 완벽하게 묘사하였으며 텍스트와 레퍼런스 반영도 훌륭합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "시선이 책을 향해 있어 '허공을 향한 시선'이라는 핵심 연출을 놓쳤으며, 레퍼런스에 없는 정체불명의 텍스트 명찰을 생성했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선이 동화책 페이지를 향해 아래로 떨어져 있어, 동화책 너머 허공을 보라는 지시와 어긋남.",
      "built_space": "유치원 교실 배경이 레퍼런스 사진의 창문 및 교구장 구조와 맞게 배치됨.",
      "entities": "이미경의 얼굴, 헤어, 의상은 레퍼런스와 일치하나, 가슴에 지시되지 않은 글씨가 적힌 명찰이 임의로 추가됨. 책 표지 텍스트 '숲속의 친구들'은 정확함.",
      "hard_violations": [
       "지시되지 않은 텍스트(가슴 명찰) 무단 생성"
      ],
      "physics": "양손이 동화책의 하단을 잡아 무게를 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "시선이 책에 머물지 않고 앞쪽 허공을 향해 멍하니 던져진 상태로 타겟 연출을 정확히 구현함.",
      "built_space": "유치원 교실 내부의 창문과 수납장이 올바른 스케일과 원근감으로 배치됨.",
      "entities": "이미경의 외모와 착장(노란색 배지 포함)이 레퍼런스와 완벽히 일치하며 불필요한 추가 요소가 없음. 책 표지의 '숲속의 친구들' 텍스트도 선명하고 정확함.",
      "hard_violations": [],
      "physics": "양손이 동화책 양쪽 하단 모서리를 안정적으로 쥐고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "동화책 너머 허공을 향한 초점 잃은 시선과 멍한 표정을 지문 그대로 완벽하게 묘사하였으며 텍스트와 레퍼런스 반영도 훌륭합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "시선이 책을 향해 있어 '허공을 향한 시선'이라는 핵심 연출을 놓쳤으며, 레퍼런스에 없는 정체불명의 텍스트 명찰을 생성했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 동화책 페이지를 향해 아래로 떨어져 있어, 동화책 너머 허공을 보라는 지시와 어긋남.",
      "built_space": "유치원 교실 배경이 레퍼런스 사진의 창문 및 교구장 구조와 맞게 배치됨.",
      "entities": "이미경의 얼굴, 헤어, 의상은 레퍼런스와 일치하나, 가슴에 지시되지 않은 글씨가 적힌 명찰이 임의로 추가됨. 책 표지 텍스트 '숲속의 친구들'은 정확함.",
      "hard_violations": [
       "지시되지 않은 텍스트(가슴 명찰) 무단 생성"
      ],
      "physics": "양손이 동화책의 하단을 잡아 무게를 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "시선이 책에 머물지 않고 앞쪽 허공을 향해 멍하니 던져진 상태로 타겟 연출을 정확히 구현함.",
      "built_space": "유치원 교실 내부의 창문과 수납장이 올바른 스케일과 원근감으로 배치됨.",
      "entities": "이미경의 외모와 착장(노란색 배지 포함)이 레퍼런스와 완벽히 일치하며 불필요한 추가 요소가 없음. 책 표지의 '숲속의 친구들' 텍스트도 선명하고 정확함.",
      "hard_violations": [],
      "physics": "양손이 동화책 양쪽 하단 모서리를 안정적으로 쥐고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "동화책 너머 허공을 응시하는 초점 잃은 시선을 잘 표현했으며, 지시된 텍스트 외에 불필요한 문자를 생성하지 않아 프롬프트를 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "시선이 허공이 아닌 동화책으로 향해 있어 연출 의도를 벗어났으며, 명찰에 요구되지 않은 임의의 텍스트가 누출되어 감점되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경의 시선은 동화책의 내용을 향하지 않고 그 너머 우측 하단의 허공을 초점 없이 응시하고 있음.",
      "built_space": "흐리게 처리된 배경에 참조 이미지와 일치하는 유치원 교실의 창문과 원목 교구장들이 배치되어 있음.",
      "entities": "이미경의 얼굴과 복장(분홍색 가디건, 흰색 리본 블라우스, 노란색 배지)이 참조와 일치함. 동화책 펼쳐진 면에 '숲속의 친구들'이라는 지정된 텍스트가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "두 손이 펼쳐진 동화책의 양쪽 가장자리를 자연스럽게 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "이미경의 시선이 동화책의 페이지 내용을 직접적으로 내려다보고 있어 '허공을 향한 시선'이라는 지시를 충족하지 못함.",
      "built_space": "흐리게 처리된 배경에 유치원 교실의 원목 수납장과 창문이 배치되어 있음.",
      "entities": "이미경의 얼굴과 복장이 참조와 일치하며 동화책에 '숲속의 친구들'이 적혀 있으나, 명찰에 지시되지 않은 문자가 포함됨.",
      "hard_violations": [
       "명찰 부근에 프롬프트에 없는 불필요한 텍스트('숲속의...')가 누출 및 발명됨 (leaked/invented text)."
      ],
      "physics": "두 손이 동화책을 안정적으로 쥐고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "동화책 너머 허공을 응시하는 초점 잃은 시선을 잘 표현했으며, 지시된 텍스트 외에 불필요한 문자를 생성하지 않아 프롬프트를 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "시선이 허공이 아닌 동화책으로 향해 있어 연출 의도를 벗어났으며, 명찰에 요구되지 않은 임의의 텍스트가 누출되어 감점되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "이미경의 시선은 동화책의 내용을 향하지 않고 그 너머 우측 하단의 허공을 초점 없이 응시하고 있음.",
      "built_space": "흐리게 처리된 배경에 참조 이미지와 일치하는 유치원 교실의 창문과 원목 교구장들이 배치되어 있음.",
      "entities": "이미경의 얼굴과 복장(분홍색 가디건, 흰색 리본 블라우스, 노란색 배지)이 참조와 일치함. 동화책 펼쳐진 면에 '숲속의 친구들'이라는 지정된 텍스트가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "두 손이 펼쳐진 동화책의 양쪽 가장자리를 자연스럽게 지지하고 있음."
     },
     {
      "label": "A",
      "direction": "이미경의 시선이 동화책의 페이지 내용을 직접적으로 내려다보고 있어 '허공을 향한 시선'이라는 지시를 충족하지 못함.",
      "built_space": "흐리게 처리된 배경에 유치원 교실의 원목 수납장과 창문이 배치되어 있음.",
      "entities": "이미경의 얼굴과 복장이 참조와 일치하며 동화책에 '숲속의 친구들'이 적혀 있으나, 명찰에 지시되지 않은 문자가 포함됨.",
      "hard_violations": [
       "명찰 부근에 프롬프트에 없는 불필요한 텍스트('숲속의...')가 누출 및 발명됨 (leaked/invented text)."
      ],
      "physics": "두 손이 동화책을 안정적으로 쥐고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 19
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S44sh2__bgfirst_bg.png",
   "bg_asset_id": "fe3f2d7a-86ce-4c2b-9f8f-296c78610b98",
   "bg_record_key": "S44sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S44sh2::cine": {
  "applied": true,
  "fingerprint": "cac6104f222581b714a1d56c046772dd6fa96bc4f18604d2882630a9d67dff04",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S44sh2_sel.png",
  "source_sha256": "b31f8f1f50bdb4420f7cd32044b2f2af1b5df959a348aba066bae747a2eecfdf",
  "file": "S44sh2_cine.png",
  "latency_ms": 10673
 },
 "S45sh3::signage": {
  "fp": "c349fe40999272e9",
  "inscriptions": []
 },
 "era_assess::7e4e97d1dfc4c9cf": {
  "subjects": [
   {
    "subject_native": "2000년대 대한민국 경찰서 건물 및 옥상",
    "search_terms_native": [
     "경찰서 건물 옥상",
     "파출소 외관 2000년대",
     "치안센터 건물",
     "경찰서 참수리 마크"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 경찰서 및 파출소의 특유의 건물 색감, 경찰 로고(참수리 마크), 옥상 및 안전 난간 구조는 일반적인 서구식 경찰서나 최신식 빌딩과 달라 철저한 고증이 필요합니다."
   }
  ]
 },
 "era_ref::a2823d78640ca6ad": {
  "subject": "2000년대 대한민국 경찰서 건물 및 옥상",
  "terms": [
   "경찰서 건물 옥상",
   "파출소 외관 2000년대",
   "치안센터 건물",
   "경찰서 참수리 마크"
  ],
  "queries": [
   [
    "2000년대 대한민국 경찰서 건물 옥상 참수리 마크",
    "2000년대 대한민국 파출소 치안센터 건물 외관"
   ],
   [
    "2004년 경찰서 청사 외관 경찰 참수리 마크",
    "2000년대 파출소 전경 경찰 참수리 간판"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://file2.nocutnews.co.kr/newsroom/image/2022/11/14/202211141825043443_0.jpg",
  "picked_reason_ko": "사진 1은 2000년대 대한민국의 일상적인 소규모 경찰 시설을 정면에서 선명하게 보여 주며, 벽돌 외장·창호·경찰 표지·옥상 가장자리와 설비를 함께 읽기 가장 좋다.",
  "sha256": "f1d541164532154893c66e087b0cb46c162878a71d896da7863250af56fff0b7",
  "file": "eraref_a2823d78640ca6ad.png"
 },
 "S45sh3::bgfirst_bg": {
  "input_fingerprint": "bbd0062dc270ba1a",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 푸른 논을 응시하며 씁쓸한 표정으로 입꼬리를 내린 서의용의 옆얼굴\n\nLOCATION (lock): At the police-station rooftop railing, overlooking the surrounding green fields at sunset.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside Seo Eui-yong, end the lateral approach close to his open profile and offset from his gaze axis, leaving a narrow strip of the green rice fields as look space beyond his lowered mouth. He lets his weight settle into the railing, jaw slackened into a bitter expression, while his attention remains fixed across the fields rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Seo Eui-yong in profile in the middle-left of the frame, foreground, looks toward rice fields below the rooftop; green rice fields in his look space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rooftop railing (Seo Eui-yong is leaning against it) — Its length recedes beside Eui-yong toward the edge of frame; used as Supports Eui-yong's leaned posture and leads the composition toward the field-side look space; green rice fields (spread below the rooftop with wind passing over them); used as Remains as a narrow but legible background destination for his gaze.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light is rendered with subdued neutral-to-cool field tones, naturalistic falloff, and moderate contrast on Eui-yong's profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 경찰서 건물 및 옥상: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 푸른 논을 응시하며 씁쓸한 표정으로 입꼬리를 내린 서의용의 옆얼굴\n\nLOCATION (lock): At the police-station rooftop railing, overlooking the surrounding green fields at sunset.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside Seo Eui-yong, end the lateral approach close to his open profile and offset from his gaze axis, leaving a narrow strip of the green rice fields as look space beyond his lowered mouth. He lets his weight settle into the railing, jaw slackened into a bitter expression, while his attention remains fixed across the fields rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Seo Eui-yong in profile in the middle-left of the frame, foreground, looks toward rice fields below the rooftop; green rice fields in his look space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rooftop railing (Seo Eui-yong is leaning against it) — Its length recedes beside Eui-yong toward the edge of frame; used as Supports Eui-yong's leaned posture and leads the composition toward the field-side look space; green rice fields (spread below the rooftop with wind passing over them); used as Remains as a narrow but legible background destination for his gaze.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light is rendered with subdued neutral-to-cool field tones, naturalistic falloff, and moderate contrast on Eui-yong's profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 경찰서 건물 및 옥상: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S45sh3__bgfirst_bg.png",
  "asset_id": "e98bcfcd-6112-47b8-aa9a-dd526d955c8d",
  "input_asset_ids": [
   "31dceee3-aef6-4043-afb5-47701676295c",
   "acbb5cf1-d0df-44af-8c4f-74d5e5eb61ab"
  ],
  "era_research": {
   "subject": "2000년대 대한민국 경찰서 건물 및 옥상",
   "queries": [
    [
     "2000년대 대한민국 경찰서 건물 옥상 참수리 마크",
     "2000년대 대한민국 파출소 치안센터 건물 외관"
    ],
    [
     "2004년 경찰서 청사 외관 경찰 참수리 마크",
     "2000년대 파출소 전경 경찰 참수리 간판"
    ]
   ],
   "picked_url": "https://file2.nocutnews.co.kr/newsroom/image/2022/11/14/202211141825043443_0.jpg",
   "sha256": "f1d541164532154893c66e087b0cb46c162878a71d896da7863250af56fff0b7",
   "file": "eraref_a2823d78640ca6ad.png"
  }
 },
 "S45sh3": {
  "input_fingerprint": "ffa0286463c1bb43",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 푸른 논을 응시하며 씁쓸한 표정으로 입꼬리를 내린 서의용의 옆얼굴\n\nLOCATION (lock): At the police-station rooftop railing, overlooking the surrounding green fields at sunset. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height beside Seo Eui-yong, end the lateral approach close to his open profile and offset from his gaze axis, leaving a narrow strip of the green rice fields as look space beyond his lowered mouth. He lets his weight settle into the railing, jaw slackened into a bitter expression, while his attention remains fixed across the fields rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Seo Eui-yong in profile in the middle-left of the frame, foreground, looks toward rice fields below the rooftop; green rice fields in his look space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rooftop railing (Seo Eui-yong is leaning against it) — Its length recedes beside Eui-yong toward the edge of frame; used as Supports Eui-yong's leaned posture and leads the composition toward the field-side look space; green rice fields (spread below the rooftop with wind passing over them); used as Remains as a narrow but legible background destination for his gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light is rendered with subdued neutral-to-cool field tones, naturalistic falloff, and moderate contrast on Eui-yong's profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 푸른 논을 응시하며 씁쓸한 표정으로 입꼬리를 내린 서의용의 옆얼굴\n\nLOCATION (lock): At the police-station rooftop railing, overlooking the surrounding green fields at sunset. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height beside Seo Eui-yong, end the lateral approach close to his open profile and offset from his gaze axis, leaving a narrow strip of the green rice fields as look space beyond his lowered mouth. He lets his weight settle into the railing, jaw slackened into a bitter expression, while his attention remains fixed across the fields rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Seo Eui-yong in profile in the middle-left of the frame, foreground, looks toward rice fields below the rooftop; green rice fields in his look space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rooftop railing (Seo Eui-yong is leaning against it) — Its length recedes beside Eui-yong toward the edge of frame; used as Supports Eui-yong's leaned posture and leads the composition toward the field-side look space; green rice fields (spread below the rooftop with wind passing over them); used as Remains as a narrow but legible background destination for his gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light is rendered with subdued neutral-to-cool field tones, naturalistic falloff, and moderate contrast on Eui-yong's profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 푸른 논을 응시하며 씁쓸한 표정으로 입꼬리를 내린 서의용의 옆얼굴\n\nLOCATION (lock): At the police-station rooftop railing, overlooking the surrounding green fields at sunset. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height beside Seo Eui-yong, end the lateral approach close to his open profile and offset from his gaze axis, leaving a narrow strip of the green rice fields as look space beyond his lowered mouth. He lets his weight settle into the railing, jaw slackened into a bitter expression, while his attention remains fixed across the fields rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: Seo Eui-yong in profile in the middle-left of the frame, foreground, looks toward rice fields below the rooftop; green rice fields in his look space in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: rooftop railing (Seo Eui-yong is leaning against it) — Its length recedes beside Eui-yong toward the edge of frame; used as Supports Eui-yong's leaned posture and leads the composition toward the field-side look space; green rice fields (spread below the rooftop with wind passing over them); used as Remains as a narrow but legible background destination for his gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Dusk light is rendered with subdued neutral-to-cool field tones, naturalistic falloff, and moderate contrast on Eui-yong's profile.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S45sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S45sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L49B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 833,
      "verdict_ko": "프레이밍(클로즈업) 지시를 어기고 배경에 레퍼런스에 없는 건물을 생성하는 오류가 있으나, 난간에 기댄 포즈는 요구사항을 충족함.  ★위반: [gemini-pro] 레퍼런스에 없는 건물 임의 생성 / [openrouter:x-ai/grok-4.6] 위치 참조에 없는 벽돌 경찰서 건물과 간판을 발명함"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "프레이밍과 기댄 자세 지시를 모두 무시했으며, 화면 우측 허공에 워터마크 형태의 시각적 결함이 존재함.  ★위반: [gemini-pro] 화면 우측 허공에 워터마크 및 마커 흔적으로 보이는 시각적 결함 노출"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.333,
      "B": 1.75
     },
     "adjusted": {
      "A": 0.833,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[gemini-pro] 레퍼런스에 없는 건물 임의 생성",
       "[openrouter:x-ai/grok-4.6] 위치 참조에 없는 벽돌 경찰서 건물과 간판을 발명함"
      ],
      "B": [
       "[gemini-pro] 화면 우측 허공에 워터마크 및 마커 흔적으로 보이는 시각적 결함 노출"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 833,
      "verdict_ko": "프레이밍(클로즈업) 지시를 어기고 배경에 레퍼런스에 없는 건물을 생성하는 오류가 있으나, 난간에 기댄 포즈는 요구사항을 충족함.  ★위반: [gemini-pro] 레퍼런스에 없는 건물 임의 생성 / [openrouter:x-ai/grok-4.6] 위치 참조에 없는 벽돌 경찰서 건물과 간판을 발명함"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "프레이밍과 기댄 자세 지시를 모두 무시했으며, 화면 우측 허공에 워터마크 형태의 시각적 결함이 존재함.  ★위반: [gemini-pro] 화면 우측 허공에 워터마크 및 마커 흔적으로 보이는 시각적 결함 노출"
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1083,
      "verdict_ko": "캐릭터가 난간에 체중을 싣고 기대어 있는 물리적 자세를 정확히 연출했으며, 시선과 표정 역시 프롬프트의 요구사항에 잘 부합합니다.  ★위반: [openrouter:x-ai/grok-4.6] 위치 사진에 없는 벽돌 경찰서 건물과 간판을 발명함"
     },
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "캐릭터가 난간에 기대지 않고 단순 직립해 있어 핵심 동작 지시를 누락했으며, 우측 배경 논 위에 정체불명의 렌더링 얼룩이 존재합니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.571,
      "B": 1.083
     },
     "violations": {
      "B": [
       "[openrouter:x-ai/grok-4.6] 위치 사진에 없는 벽돌 경찰서 건물과 간판을 발명함"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1083,
      "verdict_ko": "캐릭터가 난간에 체중을 싣고 기대어 있는 물리적 자세를 정확히 연출했으며, 시선과 표정 역시 프롬프트의 요구사항에 잘 부합합니다.  ★위반: [openrouter:x-ai/grok-4.6] 위치 사진에 없는 벽돌 경찰서 건물과 간판을 발명함"
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "캐릭터가 난간에 기대지 않고 단순 직립해 있어 핵심 동작 지시를 누락했으며, 우측 배경 논 위에 정체불명의 렌더링 얼룩이 존재합니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1916,
     "B": 3071
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 1916,
   "B": 3071
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 833,
    "verdict_ko": "프레이밍(클로즈업) 지시를 어기고 배경에 레퍼런스에 없는 건물을 생성하는 오류가 있으나, 난간에 기댄 포즈는 요구사항을 충족함.  ★위반: [gemini-pro] 레퍼런스에 없는 건물 임의 생성 / [openrouter:x-ai/grok-4.6] 위치 참조에 없는 벽돌 경찰서 건물과 간판을 발명함"
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "프레이밍과 기댄 자세 지시를 모두 무시했으며, 화면 우측 허공에 워터마크 형태의 시각적 결함이 존재함.  ★위반: [gemini-pro] 화면 우측 허공에 워터마크 및 마커 흔적으로 보이는 시각적 결함 노출"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L49B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 우측 배경의 논 위 허공에 정체불명의 얇고 구불구불한 흰색 선(그래픽 오류 및 아티팩트)들이 떠 있음.",
     "fix_en": "Remove the thin, squiggly white graphical artifacts floating in the air above the green rice fields on the right, replacing them with clear dusk atmosphere. Maintain the character Seo Eui-yong exactly as he is (face, beanie, leather jacket, posture), the rooftop railing, the sunset lighting, and the overall composition.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "난간에 체중을 싣고 기대어(leaning) 있다는 지시와 달리, 인물이 난간에 기대지 않고 똑바로 서 있음.",
     "fix_en": "Adjust the character's posture so his weight visibly settles into the rooftop railing beside him, rather than standing straight up. Maintain his clothing, the location's architecture, lighting, and current framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "턱에 힘을 빼고 입꼬리를 내린(jaw slackened) 씁쓸한 표정이라는 묘사와 달리 입을 굳게 다물고 있음.",
     "fix_en": "Relax the character's jaw and slightly lower the corners of his mouth to form a slackened, bitter expression. Maintain his facial features, skin texture, gaze direction, and the existing lighting.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "얼굴 위주의 클로즈업(close-up) 샷 지시와 달리, 인물의 허리띠 부근까지 드러나는 미디엄 샷으로 렌더링됨.",
     "fix_en": "Crop or scale the framing to a proper close-up that focuses on the character's face and shoulders.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "중좌 전경의 서의용이 옆얼굴이 아니라 양쪽 눈이 보이는 3/4 얼굴이다",
     "fix_en": "Turn the character's head further away from the lens into a strict open profile so only the near side of his face is visible. Maintain his identity, beanie, jacket, lighting, and the background fields.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 우측 배경의 논 위 허공에 정체불명의 얇고 구불구불한 흰색 선(그래픽 오류 및 아티팩트)들이 떠 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "난간에 체중을 싣고 기대어(leaning) 있다는 지시와 달리, 인물이 난간에 기대지 않고 똑바로 서 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "턱에 힘을 빼고 입꼬리를 내린(jaw slackened) 씁쓸한 표정이라는 묘사와 달리 입을 굳게 다물고 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 위주의 클로즈업(close-up) 샷 지시와 달리, 인물의 허리띠 부근까지 드러나는 미디엄 샷으로 렌더링됨.",
     "severity": "major"
    },
    {
     "issue_ko": "중좌 전경의 서의용이 옆얼굴이 아니라 양쪽 눈이 보이는 3/4 얼굴이다",
     "severity": "major"
    },
    {
     "issue_ko": "난간에 체중을 기대지 않고 양손을 주머니에 넣은 채 서 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "난간에 체중을 싣고 기대어(leaning) 있다는 지시와 달리, 인물이 난간에 기대지 않고 똑바로 서 있음.",
    "fix_en": "Adjust the character's posture so his weight visibly settles into the rooftop railing beside him, rather than standing straight up. Maintain his clothing, the location's architecture, lighting, and current framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "턱에 힘을 빼고 입꼬리를 내린(jaw slackened) 씁쓸한 표정이라는 묘사와 달리 입을 굳게 다물고 있음.",
    "fix_en": "Relax the character's jaw and slightly lower the corners of his mouth to form a slackened, bitter expression. Maintain his facial features, skin texture, gaze direction, and the existing lighting.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "얼굴 위주의 클로즈업(close-up) 샷 지시와 달리, 인물의 허리띠 부근까지 드러나는 미디엄 샷으로 렌더링됨.",
    "fix_en": "Crop or scale the framing to a proper close-up that focuses on the character's face and shoulders.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "중좌 전경의 서의용이 옆얼굴이 아니라 양쪽 눈이 보이는 3/4 얼굴이다",
    "fix_en": "Turn the character's head further away from the lens into a strict open profile so only the near side of his face is visible. Maintain his identity, beanie, jacket, lighting, and the background fields.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the thin, squiggly white graphical artifacts floating in the air above the green rice fields on the right, replacing them with clear dusk atmosphere. Maintain the character Seo Eui-yong exactly as he is (face, beanie, leather jacket, posture), the rooftop railing, the sunset lighting, and the overall composition.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "프롬프트의 인물 및 배경 설정을 잘 따랐으나, 프레임이 클로즈업보다는 미디엄 샷에 가깝고 난간에 기대는 자세가 다소 약하게 표현되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "우측 배경과 난간 위로 설명할 수 없는 기하학적 형태의 선형 아티팩트가 떠 있어 렌더링 품질이 크게 떨어집니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선이 우측 논을 향하고 있습니다.",
      "built_space": "콘크리트 난간과 금속 파이프가 있는 옥상에 위치해 있으며 구조가 자연스럽습니다.",
      "entities": "캐릭터 레퍼런스와 일치하는 서의용(비니, 가죽 재킷, 배지 등)이 보이며 배경은 푸른 논입니다.",
      "hard_violations": [],
      "physics": "바닥에 서 있으며 난간 옆에 위치해 무게 중심이 유지되고 있습니다."
     },
     {
      "label": "A",
      "direction": "시선이 우측 논을 향하고 있습니다.",
      "built_space": "콘크리트 난간과 금속 파이프가 있는 옥상에 위치해 있습니다.",
      "entities": "서의용의 외모와 의상은 레퍼런스와 일치합니다.",
      "hard_violations": [
       "화면 우측 허공과 난간 위로 기하학적인 그래픽 아티팩트(투명한 선형 자국)가 떠 있음"
      ],
      "physics": "바닥에 서 있으나 화면 우측에 물리적으로 불가능한 부유하는 흔적들이 존재합니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "프롬프트의 인물 및 배경 설정을 잘 따랐으나, 프레임이 클로즈업보다는 미디엄 샷에 가깝고 난간에 기대는 자세가 다소 약하게 표현되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "우측 배경과 난간 위로 설명할 수 없는 기하학적 형태의 선형 아티팩트가 떠 있어 렌더링 품질이 크게 떨어집니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선이 우측 논을 향하고 있습니다.",
      "built_space": "콘크리트 난간과 금속 파이프가 있는 옥상에 위치해 있으며 구조가 자연스럽습니다.",
      "entities": "캐릭터 레퍼런스와 일치하는 서의용(비니, 가죽 재킷, 배지 등)이 보이며 배경은 푸른 논입니다.",
      "hard_violations": [],
      "physics": "바닥에 서 있으며 난간 옆에 위치해 무게 중심이 유지되고 있습니다."
     },
     {
      "label": "A",
      "direction": "시선이 우측 논을 향하고 있습니다.",
      "built_space": "콘크리트 난간과 금속 파이프가 있는 옥상에 위치해 있습니다.",
      "entities": "서의용의 외모와 의상은 레퍼런스와 일치합니다.",
      "hard_violations": [
       "화면 우측 허공과 난간 위로 기하학적인 그래픽 아티팩트(투명한 선형 자국)가 떠 있음"
      ],
      "physics": "바닥에 서 있으나 화면 우측에 물리적으로 불가능한 부유하는 흔적들이 존재합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "레퍼런스와 프롬프트의 구도, 인물의 씁쓸한 표정을 정확히 재현했으며 시각적 결함이 없습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "구도와 묘사는 우수하나, 화면 우측에 편집 마커 자국이 누출되는 하드 위반이 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 화면 우측의 논밭을 자연스럽게 향함.",
      "built_space": "옥상 금속 난간 옆에 위치하며 구조물의 투시와 배치가 정상적임.",
      "entities": "서의용의 외양, 복장(비니, 가죽 재킷, 배지) 및 논밭 배경이 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "바닥을 딛고 난간에 의지해 서 있는 안정적인 자세임."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 화면 우측의 논밭을 향함.",
      "built_space": "옥상 난간과 공간 배치가 자연스러움.",
      "entities": "인물과 배경 요소는 일치하나, 우측에 불필요한 선형 아티팩트가 있음.",
      "hard_violations": [
       "화면 우측 허공 및 배경 위에 브러시/마커 자국이 누출됨 (leaked markers)"
      ],
      "physics": "인물은 바닥을 딛고 안정적으로 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "레퍼런스와 프롬프트의 구도, 인물의 씁쓸한 표정을 정확히 재현했으며 시각적 결함이 없습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "구도와 묘사는 우수하나, 화면 우측에 편집 마커 자국이 누출되는 하드 위반이 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물의 시선이 화면 우측의 논밭을 자연스럽게 향함.",
      "built_space": "옥상 금속 난간 옆에 위치하며 구조물의 투시와 배치가 정상적임.",
      "entities": "서의용의 외양, 복장(비니, 가죽 재킷, 배지) 및 논밭 배경이 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "바닥을 딛고 난간에 의지해 서 있는 안정적인 자세임."
     },
     {
      "label": "A",
      "direction": "인물의 시선이 화면 우측의 논밭을 향함.",
      "built_space": "옥상 난간과 공간 배치가 자연스러움.",
      "entities": "인물과 배경 요소는 일치하나, 우측에 불필요한 선형 아티팩트가 있음.",
      "hard_violations": [
       "화면 우측 허공 및 배경 위에 브러시/마커 자국이 누출됨 (leaked markers)"
      ],
      "physics": "인물은 바닥을 딛고 안정적으로 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S45sh3__bgfirst_bg.png",
   "bg_asset_id": "e98bcfcd-6112-47b8-aa9a-dd526d955c8d",
   "bg_record_key": "S45sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S45sh3::cine": {
  "applied": true,
  "fingerprint": "38eadab9cd5622a2976d9759a585a5ee0219f5b1f6399c5ec71a042253b8c87f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S45sh3_sel.png",
  "source_sha256": "5061db934359b573db2c7b2f26a435ed22b113146de65abc2c3f0f6452df4c26",
  "file": "S45sh3_cine.png",
  "latency_ms": 13110
 },
 "S45sh8::signage": {
  "fp": "4c866b2fe6d96ae6",
  "inscriptions": []
 },
 "S45sh8": {
  "input_fingerprint": "7e5d75cbe9e2f2a8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 바람에 흔들리는 옥상 아래 푸른 논을 배경으로 난간 곁에 나란히 선 서의용과 강주철의 전경\n\nLOCATION (lock): Along the open rooftop railing of the police station, above the wind-swept fields. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane rise and pullback from inside the rooftop, settling into a slightly high, wide rear three-quarter view with 서의용 on the left and 강주철 on the right beside the railing. The increased camera distance is the sole emphasis: both men remain prominent across the lower foreground while the blue rice fields occupy the depth beyond them, and 서의용 leans toward the view as 강주철 watches him.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 옥상 난간 (Both men are positioned beside it) — Its inside face and receding top edge are seen from the rooftop side; used as Runs across the lower foreground as a shared spatial anchor for the two men; 푸른 논 (Visible below with wind moving across it); used as Provides the expansive background beneath the rooftop and carries the final visual beat.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light appropriate to the stated sunset, with a subdued naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 바람에 흔들리는 옥상 아래 푸른 논을 배경으로 난간 곁에 나란히 선 서의용과 강주철의 전경\n\nLOCATION (lock): Along the open rooftop railing of the police station, above the wind-swept fields. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane rise and pullback from inside the rooftop, settling into a slightly high, wide rear three-quarter view with 서의용 on the left and 강주철 on the right beside the railing. The increased camera distance is the sole emphasis: both men remain prominent across the lower foreground while the blue rice fields occupy the depth beyond them, and 서의용 leans toward the view as 강주철 watches him.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 옥상 난간 (Both men are positioned beside it) — Its inside face and receding top edge are seen from the rooftop side; used as Runs across the lower foreground as a shared spatial anchor for the two men; 푸른 논 (Visible below with wind moving across it); used as Provides the expansive background beneath the rooftop and carries the final visual beat.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light appropriate to the stated sunset, with a subdued naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 바람에 흔들리는 옥상 아래 푸른 논을 배경으로 난간 곁에 나란히 선 서의용과 강주철의 전경\n\nLOCATION (lock): Along the open rooftop railing of the police station, above the wind-swept fields. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane rise and pullback from inside the rooftop, settling into a slightly high, wide rear three-quarter view with 서의용 on the left and 강주철 on the right beside the railing. The increased camera distance is the sole emphasis: both men remain prominent across the lower foreground while the blue rice fields occupy the depth beyond them, and 서의용 leans toward the view as 강주철 watches him.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 옥상 난간 (Both men are positioned beside it) — Its inside face and receding top edge are seen from the rooftop side; used as Runs across the lower foreground as a shared spatial anchor for the two men; 푸른 논 (Visible below with wind moving across it); used as Provides the expansive background beneath the rooftop and carries the final visual beat.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient light appropriate to the stated sunset, with a subdued naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "두 인물이 풍경이 아닌 서로를 마주보고 서 있음.",
    "built_space": "옥상 난간 옆에 두 사람이 서 있으며, 카메라는 측면에서 인물들을 잡고 있음.",
    "entities": "서의용과 강주철의 의상(재킷, 비니, 배낭 등) 및 외모가 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "두 사람 모두 두 발로 지면에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "서의용은 난간 쪽을 향해 몸을 기울이고 있으며, 강주철은 서의용을 바라보고 있음.",
    "built_space": "카메라가 두 사람의 뒤쪽에서 약간 높은 앵글로 난간과 배경의 논을 조망함.",
    "entities": "서의용과 강주철의 의상과 소품(강주철의 배낭 포함)이 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "서의용은 양팔을 난간에 기대어 체중을 싣고 있으며, 강주철은 지면에 서서 균형을 유지함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "요청된 후면 3/4 뷰의 카메라 앵글을 정확히 반영했으며, 서의용이 난간에 기대고 강주철이 그를 지켜보는 자세 연출을 충실히 구현함."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "지시된 후면 3/4 뷰가 아닌 측면 앵글로 촬영되었으며, 서의용이 풍경 쪽으로 기대는 동작이 누락되어 프롬프트의 카메라 및 자세 연출 지침을 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S45sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "오른쪽 강주철의 왼쪽 어깨에 배낭 끈이 전혀 없으나, 배낭 본체가 등에 똑바로 매달려 있어 물리적으로 불가능한 지지 형태를 보입니다.",
     "fix_en": "Draw a green backpack strap going over Kang Joo-chul's left shoulder, connecting to the top of the backpack so it is physically supported on both sides. Preserve the two men, their poses, their facial features, their existing clothing, the backpack body, the railing, the rice fields, and the sunset lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 장면(Previous Shot Still)에 의상이 완전히 고정(Lock)되었음에도, 왼쪽 서의용의 목 주변과 회색 티셔츠 위로 보여야 할 경찰 배지 체인이 누락되었습니다.",
     "fix_en": "Add a silver ball chain for a police badge around Seo Eui-yong's neck, visible on his gray t-shirt beneath his jacket. Preserve the two men, their poses, their facial features, their existing clothing, the backpack, the railing, the rice fields, and the sunset lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "왼쪽 서의용이 전망 쪽으로 기울이지 않고 오른쪽 강주철을 바라보고 있다.",
     "fix_en": "Turn Seo Eui-yong's head forward to look out over the fields instead of at Kang Joo-chul. Preserve the two men's bodies and positions, their facial features, their clothing, the backpack, the railing, the rice fields, and the sunset lighting.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "오른쪽 강주철의 왼쪽 어깨에 배낭 끈이 전혀 없으나, 배낭 본체가 등에 똑바로 매달려 있어 물리적으로 불가능한 지지 형태를 보입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 장면(Previous Shot Still)에 의상이 완전히 고정(Lock)되었음에도, 왼쪽 서의용의 목 주변과 회색 티셔츠 위로 보여야 할 경찰 배지 체인이 누락되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 서의용이 전망 쪽으로 기울이지 않고 오른쪽 강주철을 바라보고 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 장면(Previous Shot Still)에 의상이 완전히 고정(Lock)되었음에도, 왼쪽 서의용의 목 주변과 회색 티셔츠 위로 보여야 할 경찰 배지 체인이 누락되었습니다.",
    "fix_en": "Add a silver ball chain for a police badge around Seo Eui-yong's neck, visible on his gray t-shirt beneath his jacket. Preserve the two men, their poses, their facial features, their existing clothing, the backpack, the railing, the rice fields, and the sunset lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "왼쪽 서의용이 전망 쪽으로 기울이지 않고 오른쪽 강주철을 바라보고 있다.",
    "fix_en": "Turn Seo Eui-yong's head forward to look out over the fields instead of at Kang Joo-chul. Preserve the two men's bodies and positions, their facial features, their clothing, the backpack, the railing, the rice fields, and the sunset lighting.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Draw a green backpack strap going over Kang Joo-chul's left shoulder, connecting to the top of the backpack so it is physically supported on both sides. Preserve the two men, their poses, their facial features, their existing clothing, the backpack body, the railing, the rice fields, and the sunset lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 두 인물의 위치(왼쪽 서의용, 오른쪽 강주철), 후면 3/4 앵글, 그리고 서의용을 바라보는 강주철의 시선과 자세를 프롬프트에 맞게 훌륭히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "필수 인물인 강주철이 완전히 누락되었으며, 서의용이 강주철의 소품인 배낭을 잘못 메고 있어 프롬프트의 지시를 심각하게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "서의용은 난간에 기대어 앞의 풍경을 향하고, 강주철은 서의용을 주시함.",
      "built_space": "옥상 난간이 화면 하단을 가로지르고, 배경에 푸른 논이 넓게 배치됨.",
      "entities": "서의용(왼쪽)과 강주철(오른쪽) 모두 지정된 레퍼런스의 외모와 의상을 정확히 착용하고 등장함.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥에 서 있으며, 서의용의 팔은 난간에 자연스럽게 지지됨."
     },
     {
      "label": "B",
      "direction": "서의용이 화면 오른쪽 먼 곳을 응시함.",
      "built_space": "옥상 난간과 배경의 논이 이전 샷과 동일하게 유지됨.",
      "entities": "서의용 한 명만 등장하며, 나타나지 않은 강주철의 배낭을 메고 있음.",
      "hard_violations": [
       "명시된 필수 인물(강주철) 누락",
       "지정되지 않은 소품(강주철의 배낭)을 서의용에게 적용함"
      ],
      "physics": "인물이 바닥에 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 두 인물의 위치(왼쪽 서의용, 오른쪽 강주철), 후면 3/4 앵글, 그리고 서의용을 바라보는 강주철의 시선과 자세를 프롬프트에 맞게 훌륭히 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "필수 인물인 강주철이 완전히 누락되었으며, 서의용이 강주철의 소품인 배낭을 잘못 메고 있어 프롬프트의 지시를 심각하게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서의용은 난간에 기대어 앞의 풍경을 향하고, 강주철은 서의용을 주시함.",
      "built_space": "옥상 난간이 화면 하단을 가로지르고, 배경에 푸른 논이 넓게 배치됨.",
      "entities": "서의용(왼쪽)과 강주철(오른쪽) 모두 지정된 레퍼런스의 외모와 의상을 정확히 착용하고 등장함.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥에 서 있으며, 서의용의 팔은 난간에 자연스럽게 지지됨."
     },
     {
      "label": "B",
      "direction": "서의용이 화면 오른쪽 먼 곳을 응시함.",
      "built_space": "옥상 난간과 배경의 논이 이전 샷과 동일하게 유지됨.",
      "entities": "서의용 한 명만 등장하며, 나타나지 않은 강주철의 배낭을 메고 있음.",
      "hard_violations": [
       "명시된 필수 인물(강주철) 누락",
       "지정되지 않은 소품(강주철의 배낭)을 서의용에게 적용함"
      ],
      "physics": "인물이 바닥에 서 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 풀백 카메라 워크를 통해 두 인물을 와이드 샷으로 정확히 담아냈으며, 지정된 복장과 시선 처리(강주철이 서의용을 바라봄)를 성공적으로 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "필수 등장 인물인 강주철이 누락되었으며, 와이드 샷 프레이밍 요구를 무시하고 서의용에게 임의로 배낭을 추가해 이전 컷과의 연속성을 훼손했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "남자가 화면 오른쪽 먼 곳을 응시하고 있습니다.",
      "built_space": "옥상 바닥과 난간이 보이며, 이전 컷과 거의 동일한 카메라 위치입니다.",
      "entities": "서의용 한 명만 등장하며, 이전 컷에 없던 녹색 배낭 끈이 어깨에 추가되었습니다. 강주철은 존재하지 않습니다.",
      "hard_violations": [
       "프롬프트에 명시된 인물(강주철) 누락",
       "지정된 카메라 프레이밍(풀백 와이드 샷) 실패",
       "고정(LOCKED)된 인물 외형 임의 변경(배낭 끈 추가)"
      ],
      "physics": "두 발로 옥상 바닥에 안정적으로 서 있습니다."
     },
     {
      "label": "B",
      "direction": "왼쪽 남자는 난간 밖을 향해 몸을 기울이고, 오른쪽 남자는 왼쪽 남자를 주시하고 있습니다.",
      "built_space": "약간 높은 후면 3/4 각도에서 옥상과 난간이 넓게 조망되며, 두 인물이 난간 곁에 올바르게 배치되었습니다.",
      "entities": "왼쪽은 서의용(검은 비니, 가죽 재킷), 오른쪽은 강주철(회색 비니, 오렌지/회색 재킷, 배낭)로 두 명 모두 레퍼런스와 일치합니다.",
      "hard_violations": [],
      "physics": "두 사람 모두 옥상 바닥을 딛고 있으며, 왼쪽 남자는 난간에 팔을 올리고 체중을 지탱하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 풀백 카메라 워크를 통해 두 인물을 와이드 샷으로 정확히 담아냈으며, 지정된 복장과 시선 처리(강주철이 서의용을 바라봄)를 성공적으로 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "필수 등장 인물인 강주철이 누락되었으며, 와이드 샷 프레이밍 요구를 무시하고 서의용에게 임의로 배낭을 추가해 이전 컷과의 연속성을 훼손했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "남자가 화면 오른쪽 먼 곳을 응시하고 있습니다.",
      "built_space": "옥상 바닥과 난간이 보이며, 이전 컷과 거의 동일한 카메라 위치입니다.",
      "entities": "서의용 한 명만 등장하며, 이전 컷에 없던 녹색 배낭 끈이 어깨에 추가되었습니다. 강주철은 존재하지 않습니다.",
      "hard_violations": [
       "프롬프트에 명시된 인물(강주철) 누락",
       "지정된 카메라 프레이밍(풀백 와이드 샷) 실패",
       "고정(LOCKED)된 인물 외형 임의 변경(배낭 끈 추가)"
      ],
      "physics": "두 발로 옥상 바닥에 안정적으로 서 있습니다."
     },
     {
      "label": "A",
      "direction": "왼쪽 남자는 난간 밖을 향해 몸을 기울이고, 오른쪽 남자는 왼쪽 남자를 주시하고 있습니다.",
      "built_space": "약간 높은 후면 3/4 각도에서 옥상과 난간이 넓게 조망되며, 두 인물이 난간 곁에 올바르게 배치되었습니다.",
      "entities": "왼쪽은 서의용(검은 비니, 가죽 재킷), 오른쪽은 강주철(회색 비니, 오렌지/회색 재킷, 배낭)로 두 명 모두 레퍼런스와 일치합니다.",
      "hard_violations": [],
      "physics": "두 사람 모두 옥상 바닥을 딛고 있으며, 왼쪽 남자는 난간에 팔을 올리고 체중을 지탱하고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S45sh3"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S45sh8::cine": {
  "applied": true,
  "fingerprint": "085af757ce380b98d7807f15feaaa0dc3eb5bb1a9f000cfba6926dfb8b293225",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S45sh8_sel.png",
  "source_sha256": "de2a610294f64b41e79dd3b7b108ea4c2f124c5bac4cbcfc4d7a8809e4becc72",
  "file": "S45sh8_cine.png",
  "latency_ms": 9950
 },
 "S46sh5::signage": {
  "fp": "ea8879ebb4fd4336",
  "inscriptions": []
 },
 "S46sh5": {
  "input_fingerprint": "b7a6022ba174924f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 목을 팔로 꽉 감싼 채 환하게 웃고 있는 전택수의 상체\n\nLOCATION (lock): In the sofa seating area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track beside the sofa at upper-torso height and tighten into an oblique close view across 서의용 toward 전택수, whose smiling upper body occupies most of the frame while his bent arm encloses 서의용's neck. Camera distance is the single emphasized change; 서의용's head and neck remain at the near lower edge as the spatial anchor, with both men absorbed in the playful struggle rather than the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 소파 (서의용 is positioned at its edge) — The seat and backrest recede obliquely behind 서의용; used as Retains a narrow strip of office context behind the intertwined figures; 책상 (Visible in the office background) — Only an oblique portion of the desk faces the camera in the background; used as Provides subdued office depth without competing with the headlock.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime office, kept moderate and unsensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa area, institutional daylight, and both officers' clothing. Exclude the earlier seated group and show only the playful headlock beside the sofas.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 목을 팔로 꽉 감싼 채 환하게 웃고 있는 전택수의 상체\n\nLOCATION (lock): In the sofa seating area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track beside the sofa at upper-torso height and tighten into an oblique close view across 서의용 toward 전택수, whose smiling upper body occupies most of the frame while his bent arm encloses 서의용's neck. Camera distance is the single emphasized change; 서의용's head and neck remain at the near lower edge as the spatial anchor, with both men absorbed in the playful struggle rather than the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 소파 (서의용 is positioned at its edge) — The seat and backrest recede obliquely behind 서의용; used as Retains a narrow strip of office context behind the intertwined figures; 책상 (Visible in the office background) — Only an oblique portion of the desk faces the camera in the background; used as Provides subdued office depth without competing with the headlock.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime office, kept moderate and unsensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa area, institutional daylight, and both officers' clothing. Exclude the earlier seated group and show only the playful headlock beside the sofas.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 목을 팔로 꽉 감싼 채 환하게 웃고 있는 전택수의 상체\n\nLOCATION (lock): In the sofa seating area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track beside the sofa at upper-torso height and tighten into an oblique close view across 서의용 toward 전택수, whose smiling upper body occupies most of the frame while his bent arm encloses 서의용's neck. Camera distance is the single emphasized change; 서의용's head and neck remain at the near lower edge as the spatial anchor, with both men absorbed in the playful struggle rather than the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 소파 (서의용 is positioned at its edge) — The seat and backrest recede obliquely behind 서의용; used as Retains a narrow strip of office context behind the intertwined figures; 책상 (Visible in the office background) — Only an oblique portion of the desk faces the camera in the background; used as Provides subdued office depth without competing with the headlock.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime office, kept moderate and unsensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa area, institutional daylight, and both officers' clothing. Exclude the earlier seated group and show only the playful headlock beside the sofas.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수는 시선을 화면 밖으로 둔 채 밝게 웃고 있으며, 서의용 역시 카메라를 의식하지 않고 몸싸움에 몰두하고 있음.",
    "built_space": "어두운 가죽 소파에 앉아 있으며, 배경에는 이전 샷과 동일한 '정의구현' 액자와 책상의 일부가 심도를 가지며 배치됨.",
    "entities": "전택수(모자와 남색 재킷)와 서의용(비니와 가죽 재킷) 모두 캐릭터 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "전택수의 오른팔이 서의용의 목을 단단히 감싸고 있으며, 이를 풀려는 서의용의 왼손 그립과 지지 상태가 물리적으로 자연스러움."
   },
   {
    "label": "B",
    "direction": "전택수는 서의용을 내려다보며 웃고 있고, 서의용은 위를 향해 전택수를 바라봄.",
    "built_space": "이전 샷과 동일한 사무실 배경(소파, 액자, 캐비닛)이 올바르게 구현됨.",
    "entities": "전택수는 레퍼런스와 일치하지만, 헤드락에 걸린 인물은 서의용(Ref 3)이 아닌 이전 샷 스틸컷의 인물로 대체됨.",
    "hard_violations": [
     "목을 감싼 전택수의 오른손 구조가 해부학적으로 불가능함 (오른팔이 감아들어왔으나 엄지손가락이 아래를 향해 있음)"
    ],
    "physics": "헤드락을 거는 팔의 위치는 맞으나, 손가락과 엄지의 방향이 뼈 구조상 불가능하게 왜곡되어 얹혀 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 8,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "제시된 앵글과 두 인물의 캐릭터 레퍼런스를 정확히 반영했으며, 장난스러운 헤드락 자세를 해부학적 오류 없이 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "서의용의 캐릭터 레퍼런스를 무시하고 다른 인물을 배치했으며, 목을 감싼 손의 형태가 물리적으로 불가능한 구조를 보입니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S39sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트의 지시에 따라 전택수의 복장과 머리 스타일은 이전 컷 스틸(모자 없음, 하늘색 셔츠, 민무늬 남색 점퍼)과 동일하게 고정되어야 하나, 이를 무시하고 캐릭터 레퍼런스의 복장(수사 모자, 체크 셔츠, 패치 달린 점퍼)으로 렌더링됨.",
     "fix_en": "Change the upper man's clothing to a light blue collared shirt and plain navy bomber jacket, and remove his cap to show neatly parted black hair. Preserve the men's faces, poses, the lower man's clothing, the sofa, background, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "서의용의 가슴 앞에서 헤드락을 건 팔의 손목을 잡고 있는 손이 전택수의 체크무늬 셔츠 소매를 입고 있으나, 전택수의 등 뒤로 넘어간 왼쪽 어깨의 위치상 해당 손이 앞으로 올 수 없어 해부학적으로 불가능한 세 번째 손처럼 묘사됨.",
     "fix_en": "Remove the anatomically impossible third hand with the checkered sleeve grabbing the wrist in the foreground, replacing it with the lower man's leather jacket and hoodie. Preserve the men's faces, their exact intertwined pose, the remaining clothing, the sofa, lighting, and framing.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "클로즈업이 아니며 서의용 머리·목이 하단 가장자리 앵커가 아니라 화면 중앙에 있고 소파와 책상이 넓게 보인다.",
     "fix_en": "Crop the image tighter to anchor the lower man's head at the bottom edge, turning the shot into a close-up. Preserve the characters, their expressions, poses, clothing, lighting, and the remaining background.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "두 사람이 장난에 몰입하지 않고 카메라를 의식한 포즈로, 전택수가 렌즈 쪽으로 환하게 웃고 있다.",
     "fix_en": "Redirect the upper man's eyes to look down at the lower man instead of at the lens. Preserve his smile, both faces, poses, clothing, the sofa, background, lighting, and framing.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트의 지시에 따라 전택수의 복장과 머리 스타일은 이전 컷 스틸(모자 없음, 하늘색 셔츠, 민무늬 남색 점퍼)과 동일하게 고정되어야 하나, 이를 무시하고 캐릭터 레퍼런스의 복장(수사 모자, 체크 셔츠, 패치 달린 점퍼)으로 렌더링됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용의 가슴 앞에서 헤드락을 건 팔의 손목을 잡고 있는 손이 전택수의 체크무늬 셔츠 소매를 입고 있으나, 전택수의 등 뒤로 넘어간 왼쪽 어깨의 위치상 해당 손이 앞으로 올 수 없어 해부학적으로 불가능한 세 번째 손처럼 묘사됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수가 이전 스틸에 고정된 네이비 봄버·연청 셔츠·모자 없는 머리가 아니라 수사 모자와 SAURS 점퍼를 입고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "클로즈업이 아니며 서의용 머리·목이 하단 가장자리 앵커가 아니라 화면 중앙에 있고 소파와 책상이 넓게 보인다.",
     "severity": "major"
    },
    {
     "issue_ko": "두 사람이 장난에 몰입하지 않고 카메라를 의식한 포즈로, 전택수가 렌즈 쪽으로 환하게 웃고 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "클로즈업이 아니며 서의용 머리·목이 하단 가장자리 앵커가 아니라 화면 중앙에 있고 소파와 책상이 넓게 보인다.",
    "fix_en": "Crop the image tighter to anchor the lower man's head at the bottom edge, turning the shot into a close-up. Preserve the characters, their expressions, poses, clothing, lighting, and the remaining background.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "두 사람이 장난에 몰입하지 않고 카메라를 의식한 포즈로, 전택수가 렌즈 쪽으로 환하게 웃고 있다.",
    "fix_en": "Redirect the upper man's eyes to look down at the lower man instead of at the lens. Preserve his smile, both faces, poses, clothing, the sofa, background, lighting, and framing.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the upper man's clothing to a light blue collared shirt and plain navy bomber jacket, and remove his cap to show neatly parted black hair. Preserve the men's faces, poses, the lower man's clothing, the sofa, background, lighting, and framing.\n- Remove the anatomically impossible third hand with the checkered sleeve grabbing the wrist in the foreground, replacing it with the lower man's leather jacket and hoodie. Preserve the men's faces, their exact intertwined pose, the remaining clothing, the sofa, lighting, and framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "샷 텍스트가 요구한 두 인물의 역동적인 헤드록 액션과 밀착된 프레이밍을 훌륭하게 구현했으나, 전택수의 복장을 이전 컷 스틸로 고정(Lock)하지 않고 레퍼런스의 모자와 점퍼를 입힌 점이 유일한 아쉬움입니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "헤드록 액션과 두 인물 구도를 완전히 무시한 채 1인 구도를 답습했으며, 전택수의 얼굴에 서의용의 옷을 입히는 치명적인 정체성 훼손을 범해 최하점을 받았습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수는 정면 왼쪽을 향해 환하게 웃고 있으며, 서의용은 시선이 위를 향한 채 헤드록에 걸려 있음.",
      "built_space": "사무실 가죽 소파에 두 사람이 얽혀 앉아 있으며, 배경에는 책상 일부와 '정의구현' 서예 액자가 정확하게 배치됨.",
      "entities": "전택수와 서의용이 모두 명확히 묘사됨. 서의용은 비니와 가죽 재킷을 입고 있으나, 전택수는 이전 컷 스틸의 복장(모자 없음) 대신 캐릭터 레퍼런스의 수사 모자와 점퍼를 착용함.",
      "hard_violations": [],
      "physics": "전택수의 오른팔이 서의용의 목을 단단히 감싸며 체중을 싣고 있으며, 서의용의 왼손이 그 팔을 붙잡고 저항하는 자세가 물리적으로 매우 자연스럽게 지지되고 있음."
     },
     {
      "label": "B",
      "direction": "단일 인물이 화면 바깥 왼쪽을 무표정하게 응시하고 있음.",
      "built_space": "가죽 소파에 앉아 있으며 뒤쪽으로 캐비닛과 '정의구현' 액자가 있는 이전 컷과 동일한 공간 구성임.",
      "entities": "전택수의 얼굴을 한 인물 한 명만 등장하며, 서의용의 가죽 재킷과 후드를 입고 있어 두 인물의 정체성과 복장이 섞여버림.",
      "hard_violations": [],
      "physics": "소파에 기댄 채 양팔을 다리 위에 올리고 하중을 받으며 안정적으로 앉아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "샷 텍스트가 요구한 두 인물의 역동적인 헤드록 액션과 밀착된 프레이밍을 훌륭하게 구현했으나, 전택수의 복장을 이전 컷 스틸로 고정(Lock)하지 않고 레퍼런스의 모자와 점퍼를 입힌 점이 유일한 아쉬움입니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "헤드록 액션과 두 인물 구도를 완전히 무시한 채 1인 구도를 답습했으며, 전택수의 얼굴에 서의용의 옷을 입히는 치명적인 정체성 훼손을 범해 최하점을 받았습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수는 정면 왼쪽을 향해 환하게 웃고 있으며, 서의용은 시선이 위를 향한 채 헤드록에 걸려 있음.",
      "built_space": "사무실 가죽 소파에 두 사람이 얽혀 앉아 있으며, 배경에는 책상 일부와 '정의구현' 서예 액자가 정확하게 배치됨.",
      "entities": "전택수와 서의용이 모두 명확히 묘사됨. 서의용은 비니와 가죽 재킷을 입고 있으나, 전택수는 이전 컷 스틸의 복장(모자 없음) 대신 캐릭터 레퍼런스의 수사 모자와 점퍼를 착용함.",
      "hard_violations": [],
      "physics": "전택수의 오른팔이 서의용의 목을 단단히 감싸며 체중을 싣고 있으며, 서의용의 왼손이 그 팔을 붙잡고 저항하는 자세가 물리적으로 매우 자연스럽게 지지되고 있음."
     },
     {
      "label": "B",
      "direction": "단일 인물이 화면 바깥 왼쪽을 무표정하게 응시하고 있음.",
      "built_space": "가죽 소파에 앉아 있으며 뒤쪽으로 캐비닛과 '정의구현' 액자가 있는 이전 컷과 동일한 공간 구성임.",
      "entities": "전택수의 얼굴을 한 인물 한 명만 등장하며, 서의용의 가죽 재킷과 후드를 입고 있어 두 인물의 정체성과 복장이 섞여버림.",
      "hard_violations": [],
      "physics": "소파에 기댄 채 양팔을 다리 위에 올리고 하중을 받으며 안정적으로 앉아 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "샷 텍스트에 명시된 헤드락 액션과 환하게 웃는 표정 등 핵심 연출을 훌륭하게 구현했으나, 전택수의 복장을 이전 컷 스틸이 아닌 캐릭터 레퍼런스에서 가져온 점이 아쉽습니다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "프롬프트의 핵심인 헤드락 연출과 서의용의 등장을 완전히 누락한 채 이전 컷 스틸의 구도만 답습했으며, 인물의 의상마저 잘못 적용되어 요구사항을 전혀 충족하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "전택수는 화면 오른쪽 상단을 향해 환하게 웃으며 시선을 두고 있고, 목이 졸린 서의용은 시선이 위를 향하고 있음.",
      "built_space": "사무실 소파에 두 사람이 겹쳐 앉아 있으며, 배경으로 '정의구현' 액자와 책상 등 사무실 환경이 적절한 원근감과 함께 보임.",
      "entities": "전택수와 서의용이 모두 등장함. 서의용은 레퍼런스(비니, 가죽 재킷)와 일치하지만, 전택수는 록(lock)이 걸린 이전 컷 스틸의 복장(파란 셔츠와 점퍼) 대신 캐릭터 레퍼런스의 복장(수사 모자, 경찰 점퍼)을 입고 있어 의상 조건에 위배됨.",
      "hard_violations": [],
      "physics": "전택수의 구부린 팔이 서의용의 목을 단단히 감싸고 있으며, 서의용의 손이 방어를 위해 전택수의 팔을 잡고 있는 등 물리적인 자세와 접촉이 자연스럽게 지탱됨."
     },
     {
      "label": "A",
      "direction": "남자가 화면 왼쪽 앞에 위치한 인물(어깨 일부만 보임)을 향해 진지한 표정으로 시선을 향하고 있음.",
      "built_space": "사무실 소파에 앉아 있으며 뒤쪽에 수납장과 '정의구현' 액자가 배치되어 있음.",
      "entities": "전택수(얼굴은 이전 컷 스틸과 일치함) 한 명만 프레임에 제대로 잡혀 있으며, 서의용의 의상(가죽 재킷과 후드)을 입고 있음. 서의용은 등장하지 않음.",
      "hard_violations": [],
      "physics": "남자가 소파에 기대어 앉아 양팔을 다리 쪽과 무릎 위에 안정적으로 올려두고 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "샷 텍스트에 명시된 헤드락 액션과 환하게 웃는 표정 등 핵심 연출을 훌륭하게 구현했으나, 전택수의 복장을 이전 컷 스틸이 아닌 캐릭터 레퍼런스에서 가져온 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "프롬프트의 핵심인 헤드락 연출과 서의용의 등장을 완전히 누락한 채 이전 컷 스틸의 구도만 답습했으며, 인물의 의상마저 잘못 적용되어 요구사항을 전혀 충족하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수는 화면 오른쪽 상단을 향해 환하게 웃으며 시선을 두고 있고, 목이 졸린 서의용은 시선이 위를 향하고 있음.",
      "built_space": "사무실 소파에 두 사람이 겹쳐 앉아 있으며, 배경으로 '정의구현' 액자와 책상 등 사무실 환경이 적절한 원근감과 함께 보임.",
      "entities": "전택수와 서의용이 모두 등장함. 서의용은 레퍼런스(비니, 가죽 재킷)와 일치하지만, 전택수는 록(lock)이 걸린 이전 컷 스틸의 복장(파란 셔츠와 점퍼) 대신 캐릭터 레퍼런스의 복장(수사 모자, 경찰 점퍼)을 입고 있어 의상 조건에 위배됨.",
      "hard_violations": [],
      "physics": "전택수의 구부린 팔이 서의용의 목을 단단히 감싸고 있으며, 서의용의 손이 방어를 위해 전택수의 팔을 잡고 있는 등 물리적인 자세와 접촉이 자연스럽게 지탱됨."
     },
     {
      "label": "B",
      "direction": "남자가 화면 왼쪽 앞에 위치한 인물(어깨 일부만 보임)을 향해 진지한 표정으로 시선을 향하고 있음.",
      "built_space": "사무실 소파에 앉아 있으며 뒤쪽에 수납장과 '정의구현' 액자가 배치되어 있음.",
      "entities": "전택수(얼굴은 이전 컷 스틸과 일치함) 한 명만 프레임에 제대로 잡혀 있으며, 서의용의 의상(가죽 재킷과 후드)을 입고 있음. 서의용은 등장하지 않음.",
      "hard_violations": [],
      "physics": "남자가 소파에 기대어 앉아 양팔을 다리 쪽과 무릎 위에 안정적으로 올려두고 지탱하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S39sh4"
  }
 },
 "S46sh5::cine": {
  "applied": true,
  "fingerprint": "28808799e7fcdf3978900e9ff8c6d91307267ca44cdb51c06adfde9d79eb37dd",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S46sh5_sel.png",
  "source_sha256": "78d3a4fde6af81f456672e6ed9dde5efa9099385a9a0425e14a1f733b27dd9d3",
  "file": "S46sh5_cine.png",
  "latency_ms": 10385
 },
 "S46sh8::signage": {
  "fp": "47842664e3f268a6",
  "inscriptions": []
 },
 "S46sh8": {
  "input_fingerprint": "16320a4d1e91812e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용을 향해 신뢰가 담긴 따뜻한 미소를 짓는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): Beside the sofas inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close behind and slightly beside 서의용's shoulder at seated eye height, finish the dolly-in across the conversational axis on 전택수's face, leaving 서의용 as a soft near-edge shoulder and preserving a small amount of look space toward him. The emphasized change is 전택수's gaze: his attention settles directly on 서의용 as the broad smile resolves into warmth and trust.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 소파 (전택수 is seated on it) — The backrest angles away behind 전택수; used as Supports 전택수's seated posture while falling mostly outside the narrow focus.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office ambience with restrained contrast and gentle facial modeling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office lighting, sofa background, and the senior officer's clothing. Exclude the headlock and the younger officer from the close frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용을 향해 신뢰가 담긴 따뜻한 미소를 짓는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): Beside the sofas inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close behind and slightly beside 서의용's shoulder at seated eye height, finish the dolly-in across the conversational axis on 전택수's face, leaving 서의용 as a soft near-edge shoulder and preserving a small amount of look space toward him. The emphasized change is 전택수's gaze: his attention settles directly on 서의용 as the broad smile resolves into warmth and trust.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 소파 (전택수 is seated on it) — The backrest angles away behind 전택수; used as Supports 전택수's seated posture while falling mostly outside the narrow focus.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office ambience with restrained contrast and gentle facial modeling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office lighting, sofa background, and the senior officer's clothing. Exclude the headlock and the younger officer from the close frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용을 향해 신뢰가 담긴 따뜻한 미소를 짓는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): Beside the sofas inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close behind and slightly beside 서의용's shoulder at seated eye height, finish the dolly-in across the conversational axis on 전택수's face, leaving 서의용 as a soft near-edge shoulder and preserving a small amount of look space toward him. The emphasized change is 전택수's gaze: his attention settles directly on 서의용 as the broad smile resolves into warmth and trust.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 소파 (전택수 is seated on it) — The backrest angles away behind 전택수; used as Supports 전택수's seated posture while falling mostly outside the narrow focus.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office ambience with restrained contrast and gentle facial modeling.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office lighting, sofa background, and the senior officer's clothing. Exclude the headlock and the younger officer from the close frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수의 시선이 프레임 전경에 걸쳐 있는 서의용을 정확히 향하고 있음.",
    "built_space": "사무실 내부. 전택수가 가죽 소파에 앉아 있으며, 배경에 소파 등받이, 목재 수납장, 서예 액자가 위치함.",
    "entities": "전택수는 참조 이미지와 일치하는 남색 재킷, 체크 셔츠, '수사' 문구가 적힌 모자를 착용하고 있음. 서의용은 가죽 재킷을 입은 어깨와 뒷모습으로 올바르게 표현됨.",
    "hard_violations": [],
    "physics": "전택수는 소파에 체중을 싣고 자연스럽게 앉아 자세를 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "전택수의 시선이 프레임 앞쪽의 서의용을 향하고 있음.",
    "built_space": "사무실 내부. 소파에 앉아 있는 구도이며, 배경의 액자와 가구 배치가 이전 샷의 환경과 일치함.",
    "entities": "전경의 서의용은 어깨만 걸쳐서 보이나, 전택수는 이전 샷에서 고정된 필수 의상인 모자를 착용하지 않고 있음.",
    "hard_violations": [],
    "physics": "소파에 기대어 앉아 있는 자세로 무게 중심과 지지가 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "프롬프트가 요구한 오버 더 숄더 구도와 전택수의 따뜻한 미소를 잘 연출했으며, 이전 샷에 고정된 의상(모자 포함)을 정확히 유지했습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "구도와 표정 연출은 좋으나, 이전 샷에서 반드시 유지해야 하는 전택수의 모자가 누락되어 의상 고정 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S46sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 전경에 걸쳐진 서의용의 어깨가 이전 컷 참고 이미지에서 배제된 인물의 갈색 가죽 재킷을 그대로 입고 있어 의상 복사 금지 지시를 위반했습니다.",
     "fix_en": "Replace the foreground character's brown leather jacket with a generic dark grey wool suit jacket. Preserve the character's position, the background office, and the smiling man in the navy police jacket.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "전택수의 남색 재킷 가슴 부위에 참고 이미지에는 존재하지 않는 가로형 은색 반사 띠가 임의로 추가되었습니다.",
     "fix_en": "Remove the horizontal silver reflective strips across the chest of the navy jacket, making the fabric solid navy blue. Preserve the jacket's overall shape, the patches, the man's face, and the foreground character.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "전택수의 왼쪽 팔에 부착된 방패 모양 패치의 글자가 'POURS'로 잘못 표기되어 있습니다.",
     "fix_en": "Smudge the text on the shield-shaped patch on the left arm so that it is entirely illegible. Preserve the jacket's design, the man's face, and the foreground character.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "서의용은 가장자리의 부드러운 어깨만 보여야 하는데 머리 뒷면 전체가 크게 들어와 있다",
     "fix_en": "Move the camera in to reduce the foreground character to just a soft, out-of-focus shoulder at the left edge. Preserve the characters' identities and the office background.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전택수 얼굴 클로즈업이 아니라 상반신과 상대가 함께 보이는 오버숄더 미디엄 샷이다",
     "fix_en": "Push the camera in to a true close-up of the smiling man's face, excluding his lower torso and most of the foreground character. Preserve the lighting, the office background, and the smiling expression.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 전경에 걸쳐진 서의용의 어깨가 이전 컷 참고 이미지에서 배제된 인물의 갈색 가죽 재킷을 그대로 입고 있어 의상 복사 금지 지시를 위반했습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수의 남색 재킷 가슴 부위에 참고 이미지에는 존재하지 않는 가로형 은색 반사 띠가 임의로 추가되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수의 왼쪽 팔에 부착된 방패 모양 패치의 글자가 'POURS'로 잘못 표기되어 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "왼쪽 전경에 제외해야 할 이전 샷의 젊은 수사관(가죽 재킷과 뒷머리)이 그대로 등장한다",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용은 가장자리의 부드러운 어깨만 보여야 하는데 머리 뒷면 전체가 크게 들어와 있다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수 얼굴 클로즈업이 아니라 상반신과 상대가 함께 보이는 오버숄더 미디엄 샷이다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "전택수의 남색 재킷 가슴 부위에 참고 이미지에는 존재하지 않는 가로형 은색 반사 띠가 임의로 추가되었습니다.",
    "fix_en": "Remove the horizontal silver reflective strips across the chest of the navy jacket, making the fabric solid navy blue. Preserve the jacket's overall shape, the patches, the man's face, and the foreground character.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "전택수의 왼쪽 팔에 부착된 방패 모양 패치의 글자가 'POURS'로 잘못 표기되어 있습니다.",
    "fix_en": "Smudge the text on the shield-shaped patch on the left arm so that it is entirely illegible. Preserve the jacket's design, the man's face, and the foreground character.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "서의용은 가장자리의 부드러운 어깨만 보여야 하는데 머리 뒷면 전체가 크게 들어와 있다",
    "fix_en": "Move the camera in to reduce the foreground character to just a soft, out-of-focus shoulder at the left edge. Preserve the characters' identities and the office background.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "전택수 얼굴 클로즈업이 아니라 상반신과 상대가 함께 보이는 오버숄더 미디엄 샷이다",
    "fix_en": "Push the camera in to a true close-up of the smiling man's face, excluding his lower torso and most of the foreground character. Preserve the lighting, the office background, and the smiling expression.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the foreground character's brown leather jacket with a generic dark grey wool suit jacket. Preserve the character's position, the background office, and the smiling man in the navy police jacket.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "이전 컷의 인물 의상(가죽 재킷)을 배제하라는 지시를 정확히 수행하여 전경 인물에 회색 재킷을 입혔으며, 전택수의 따뜻한 미소와 지정된 오버더숄더 구도를 훌륭하게 연출함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "전택수의 표정과 카메라 구도는 적절히 구현되었으나, 이전 컷 인물의 의상(가죽 재킷)을 전경에 그대로 노출하여 명시적인 배제 지시를 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 전경에 위치한 서의용의 어깨 너머 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사무실 내부. 가죽 소파와 배경의 나무 패널, 액자 등이 지시된 위치에 알맞게 배치됨.",
      "entities": "전택수는 참고 이미지와 동일한 인상 및 복장(수사 모자, 재킷)을 갖춤. 전경의 인물은 이전 컷의 갈색 가죽 재킷을 그대로 입고 있음.",
      "hard_violations": [],
      "physics": "전택수는 가죽 소파에 체중을 싣고 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 전경에 위치한 서의용의 어깨 너머 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사무실 내부. 가죽 소파와 배경의 나무 패널, 액자 등이 지시된 위치에 알맞게 배치됨.",
      "entities": "전택수는 참고 이미지와 동일한 인상 및 복장을 갖춤. 전경의 인물은 회색 재킷을 입고 있어 이전 컷 인물의 의상 배제 지시를 충족함.",
      "hard_violations": [],
      "physics": "전택수는 가죽 소파에 체중을 싣고 안정적으로 앉아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "이전 컷의 인물 의상(가죽 재킷)을 배제하라는 지시를 정확히 수행하여 전경 인물에 회색 재킷을 입혔으며, 전택수의 따뜻한 미소와 지정된 오버더숄더 구도를 훌륭하게 연출함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "전택수의 표정과 카메라 구도는 적절히 구현되었으나, 이전 컷 인물의 의상(가죽 재킷)을 전경에 그대로 노출하여 명시적인 배제 지시를 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 전경에 위치한 서의용의 어깨 너머 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사무실 내부. 가죽 소파와 배경의 나무 패널, 액자 등이 지시된 위치에 알맞게 배치됨.",
      "entities": "전택수는 참고 이미지와 동일한 인상 및 복장(수사 모자, 재킷)을 갖춤. 전경의 인물은 이전 컷의 갈색 가죽 재킷을 그대로 입고 있음.",
      "hard_violations": [],
      "physics": "전택수는 가죽 소파에 체중을 싣고 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 전경에 위치한 서의용의 어깨 너머 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사무실 내부. 가죽 소파와 배경의 나무 패널, 액자 등이 지시된 위치에 알맞게 배치됨.",
      "entities": "전택수는 참고 이미지와 동일한 인상 및 복장을 갖춤. 전경의 인물은 회색 재킷을 입고 있어 이전 컷 인물의 의상 배제 지시를 충족함.",
      "hard_violations": [],
      "physics": "전택수는 가죽 소파에 체중을 싣고 안정적으로 앉아 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "전택수의 외모와 배경은 훌륭하게 일치하나, 전경에 있는 서의용의 어깨가 이전 샷의 가죽 재킷이 아닌 회색 양복으로 렌더링되어 장면의 연속성을 훼손했습니다."
     },
     {
      "label": "B",
      "score": 1125,
      "verdict_ko": "전택수의 따뜻한 표정과 복장 등 핵심 요소를 잘 재현했으며, 서의용의 어깨 복장 또한 이전 샷과 동일한 갈색 가죽 재킷으로 유지하여 연속성을 훌륭하게 지켰습니다.  ★위반: [openrouter:x-ai/grok-4.6] 이전 스틸의 젊은 수사관 의상(가죽 재킷)을 서의용 어깨에 옮겨 ‘젊은 수사관 제외’와 ‘그 인물의 얼굴·몸·의상을 가져오지 말 것’을 위반"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 1.375
     },
     "adjusted": {
      "A": 1.714,
      "B": 1.125
     },
     "violations": {
      "B": [
       "[openrouter:x-ai/grok-4.6] 이전 스틸의 젊은 수사관 의상(가죽 재킷)을 서의용 어깨에 옮겨 ‘젊은 수사관 제외’와 ‘그 인물의 얼굴·몸·의상을 가져오지 말 것’을 위반"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.625,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "전택수의 외모와 배경은 훌륭하게 일치하나, 전경에 있는 서의용의 어깨가 이전 샷의 가죽 재킷이 아닌 회색 양복으로 렌더링되어 장면의 연속성을 훼손했습니다."
     },
     {
      "label": "A",
      "score": 1125,
      "verdict_ko": "전택수의 따뜻한 표정과 복장 등 핵심 요소를 잘 재현했으며, 서의용의 어깨 복장 또한 이전 샷과 동일한 갈색 가죽 재킷으로 유지하여 연속성을 훌륭하게 지켰습니다.  ★위반: [openrouter:x-ai/grok-4.6] 이전 스틸의 젊은 수사관 의상(가죽 재킷)을 서의용 어깨에 옮겨 ‘젊은 수사관 제외’와 ‘그 인물의 얼굴·몸·의상을 가져오지 말 것’을 위반"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1129,
     "B": 1721
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S46sh5"
  }
 },
 "S46sh8::cine": {
  "applied": true,
  "fingerprint": "1091314d58f36d7c12882586e0b1b928ac68518259e07ee191332ce965c3e74c",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S46sh8_sel.png",
  "source_sha256": "270fe195f1bb338ad40b378469bac722dfddcbd959796d805393a8db3c8e9932",
  "file": "S46sh8_cine.png",
  "latency_ms": 12487
 },
 "S47sh5::signage": {
  "fp": "0880a2926d30abd1",
  "inscriptions": []
 },
 "era_assess::e7cf509b5ce2797c": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 가정집 주방 싱크대와 가스레인지",
    "search_terms_native": [
     "2000년대 한국 싱크대",
     "일반 가정 가스레인지 싱크대",
     "낡은 주방 싱크대 가스렌지",
     "2000년대 아파트 주방 인테리어"
    ],
    "language_lock_native": "이 검색은 오직 한국어로만 진행되어야 하며, 다른 언어로 번역하거나 추가 검색어를 더해서는 안 됩니다.",
    "reason_ko": "2000년대에서 2010년대 초반 한국의 서민 가정집 주방 싱크대와 가스레인지 구조, 타일 스타일은 현대적이거나 서구적인 주방 이미지와 다른 독특한 생활감이 있어 철저한 고증이 필요합니다."
   }
  ]
 },
 "era_ref::fd7aef718365d78d": {
  "subject": "2000년대~2010년대 한국 가정집 주방 싱크대와 가스레인지",
  "terms": [
   "2000년대 한국 싱크대",
   "일반 가정 가스레인지 싱크대",
   "낡은 주방 싱크대 가스렌지",
   "2000년대 아파트 주방 인테리어"
  ],
  "queries": [
   [
    "2000년대 한국 싱크대 / 일반 가정 가스레인지 싱크대 / 낡은 주방 싱크대 가스렌지 / 2000년대 아파트 주방 인테리어"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://searchome-aws.hmgcdn.com/article/doc62379/imgs20201023161205450.jpg",
  "picked_reason_ko": "2번은 2000~2010년대 한국 아파트에서 흔히 볼 수 있는 일자형 주방으로, 싱크대와 2구 가스레인지, 상·하부장, 타일 벽의 구조와 비례가 가장 명확하게 보인다.",
  "sha256": "ce5d6b3b125bbf35b86d08dde4a37cac6229163a2344ec94521df614a9223d89",
  "file": "eraref_fd7aef718365d78d.png"
 },
 "groupbg::부녀 집 거실": {
  "input_fingerprint": "1bf4c17247e22b69",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "부녀 집 거실",
    "tags": [
     "S47sh5",
     "S47sh7"
    ]
   },
   "context_sig": "1a74612b853fac92",
   "era_research_sha": "ce5d6b3b125bbf35b86d08dde4a37cac6229163a2344ec94521df614a9223d89"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: In the sunlit kitchen-dining area of the detective's home, beside the breakfast table.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n서의용 집 내부: 아침 햇살이 비치는 소박하고 일상적인 홀아비 가정의 거실과 주방이다. (특징: 유리창 너머 베란다 건조대에 가지런히 널려 있는 빨래들; 가스 불 켜진 레인지와 국냄비가 있는 낡은 주방 싱크대; 플라스틱 반찬통 여러 개가 뚜껑이 열린 채 놓인 좁은 식탁)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 의용 집 - 아침\n- 거실을 환하게 비추는 아침햇살. 베란다엔 빨래들이 가지런히 널려있다.\n\nTIME OF DAY (lock): morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 가정집 주방 싱크대와 가스레인지: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: In the sunlit kitchen-dining area of the detective's home, beside the breakfast table.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n서의용 집 내부: 아침 햇살이 비치는 소박하고 일상적인 홀아비 가정의 거실과 주방이다. (특징: 유리창 너머 베란다 건조대에 가지런히 널려 있는 빨래들; 가스 불 켜진 레인지와 국냄비가 있는 낡은 주방 싱크대; 플라스틱 반찬통 여러 개가 뚜껑이 열린 채 놓인 좁은 식탁)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 의용 집 - 아침\n- 거실을 환하게 비추는 아침햇살. 베란다엔 빨래들이 가지런히 널려있다.\n\nTIME OF DAY (lock): morning.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 가정집 주방 싱크대와 가스레인지: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_부녀_집_거실_8d3b16.png",
  "asset_id": "88ce5d51-647c-40f3-a6b9-b47c38cea602",
  "input_asset_ids": [
   "c50fb120-1187-4396-97d7-dc97f3f17811"
  ],
  "origin_tag": "S47sh5",
  "place_text": "In the sunlit kitchen-dining area of the detective's home, beside the breakfast table.",
  "origin_inputs": {
   "place_text": "In the sunlit kitchen-dining area of the detective's home, beside the breakfast table.",
   "time_of_day_en": "morning",
   "conti_asset_id": "c50fb120-1187-4396-97d7-dc97f3f17811"
  },
  "era_research": {
   "subject": "2000년대~2010년대 한국 가정집 주방 싱크대와 가스레인지",
   "terms": [
    "2000년대 한국 싱크대",
    "일반 가정 가스레인지 싱크대",
    "낡은 주방 싱크대 가스렌지",
    "2000년대 아파트 주방 인테리어"
   ],
   "queries": [
    [
     "2000년대 한국 싱크대 / 일반 가정 가스레인지 싱크대 / 낡은 주방 싱크대 가스렌지 / 2000년대 아파트 주방 인테리어"
    ]
   ],
   "candidates": 4,
   "picked_index": 2,
   "picked_url": "https://searchome-aws.hmgcdn.com/article/doc62379/imgs20201023161205450.jpg",
   "picked_reason_ko": "2번은 2000~2010년대 한국 아파트에서 흔히 볼 수 있는 일자형 주방으로, 싱크대와 2구 가스레인지, 상·하부장, 타일 벽의 구조와 비례가 가장 명확하게 보인다.",
   "sha256": "ce5d6b3b125bbf35b86d08dde4a37cac6229163a2344ec94521df614a9223d89",
   "file": "eraref_fd7aef718365d78d.png",
   "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_fd7aef718365d78d.png"
  }
 },
 "era_assess::a8868cc98e989299": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 한국 아파트 주방 및 식당 구역",
    "search_terms_native": [
     "2000년대 아파트 주방",
     "2010년대 한국 가옥 식탁",
     "한국 가정집 싱크대 인테리어"
    ],
    "language_lock_native": "모든 검색어는 오직 한국어로만 작성해야 하며 다른 언어로 번역하거나 결합해서는 안 됩니다.",
    "reason_ko": "일반적인 이미지 모델은 2000년대~2010년대 한국 아파트 특유의 싱크대 형태, 주방 가구 스타일, 벽지 및 전형적인 레이아웃 대신 서구식 주방이나 아주 현대적인 북유럽풍 주방으로 잘못 그릴 가능성이 높습니다."
   }
  ]
 },
 "era_ref::3197bd2ded4299ba": {
  "subject": "2000년대-2010년대 한국 아파트 주방 및 식당 구역",
  "terms": [
   "2000년대 아파트 주방",
   "2010년대 한국 가옥 식탁",
   "한국 가정집 싱크대 인테리어"
  ],
  "queries": [
   [
    "2000년대 한국 아파트 주방 식당 식탁 싱크대 인테리어",
    "2010년대 한국 가정집 주방 식탁 싱크대 인테리어"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://octapi.lxzin.com/imageBlockProp/image/202412/20/720/0/e324e7f2-de2f-4365-9e3a-8b3b7551e168.jpg",
  "picked_reason_ko": "3번은 2000~2010년대 한국 아파트의 주방과 소규모 식사 공간이 한 화면에 담겨 있으며, 수납장·조리대·발코니 연결부·식탁의 실제 사용 모습과 비례를 가장 잘 읽을 수 있다.",
  "sha256": "a4135d924d9b7dd4c0c84b9f75c4cf7892b3036f5f22be8cbac3882a11dc0ed2",
  "file": "eraref_3197bd2ded4299ba.png"
 },
 "S47sh5::bgfirst_bg": {
  "input_fingerprint": "2a4b9e05e3c36b45",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 귀에 스마트폰을 바짝 댄 채 미간을 살짝 찌푸린 서의용의 얼굴\n\nLOCATION (lock): In the sunlit kitchen-dining area of the detective's home, beside the breakfast table.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly-in at 서의용's shoulder height in a close front three-quarter view, placing his furrowed face just off center and the smartphone tight against his ear along the frame edge. Camera distance is the only emphasized shift; his eyes lower in concentration while the dining area remains softly legible behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 식탁 (Set with soybean-paste stew and several side-dish containers) — Its near edge crosses the background at an oblique angle; used as Forms the lower background context for the interrupted breakfast; 스마트폰 (Held tightly to his ear during the call) — Its screen side is pressed toward his face while the outer side is visible to camera; used as Acts as the immediate narrative anchor against 서의용's face; 베란다의 빨래 (Hung neatly on the balcony); used as Provides distant domestic context beyond the dining area.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright morning sunlight illuminates the living room while the close view retains a subdued naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국 아파트 주방 및 식당 구역: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 귀에 스마트폰을 바짝 댄 채 미간을 살짝 찌푸린 서의용의 얼굴\n\nLOCATION (lock): In the sunlit kitchen-dining area of the detective's home, beside the breakfast table.\n\nTIME OF DAY (lock): morning.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: End the dolly-in at 서의용's shoulder height in a close front three-quarter view, placing his furrowed face just off center and the smartphone tight against his ear along the frame edge. Camera distance is the only emphasized shift; his eyes lower in concentration while the dining area remains softly legible behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 식탁 (Set with soybean-paste stew and several side-dish containers) — Its near edge crosses the background at an oblique angle; used as Forms the lower background context for the interrupted breakfast; 스마트폰 (Held tightly to his ear during the call) — Its screen side is pressed toward his face while the outer side is visible to camera; used as Acts as the immediate narrative anchor against 서의용's face; 베란다의 빨래 (Hung neatly on the balcony); used as Provides distant domestic context beyond the dining area.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright morning sunlight illuminates the living room while the close view retains a subdued naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국 아파트 주방 및 식당 구역: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S47sh5__bgfirst_bg.png",
  "asset_id": "8cf8653f-7c0c-4879-abaf-d75ad08cf0e7",
  "input_asset_ids": [
   "c50fb120-1187-4396-97d7-dc97f3f17811",
   "88ce5d51-647c-40f3-a6b9-b47c38cea602"
  ],
  "era_research": {
   "subject": "2000년대-2010년대 한국 아파트 주방 및 식당 구역",
   "queries": [
    [
     "2000년대 한국 아파트 주방 식당 식탁 싱크대 인테리어",
     "2010년대 한국 가정집 주방 식탁 싱크대 인테리어"
    ]
   ],
   "picked_url": "https://octapi.lxzin.com/imageBlockProp/image/202412/20/720/0/e324e7f2-de2f-4365-9e3a-8b3b7551e168.jpg",
   "sha256": "a4135d924d9b7dd4c0c84b9f75c4cf7892b3036f5f22be8cbac3882a11dc0ed2",
   "file": "eraref_3197bd2ded4299ba.png"
  }
 },
 "S47sh5": {
  "input_fingerprint": "2d753c82891318f4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 귀에 스마트폰을 바짝 댄 채 미간을 살짝 찌푸린 서의용의 얼굴\n\nLOCATION (lock): In the sunlit kitchen-dining area of the detective's home, beside the breakfast table. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the dolly-in at 서의용's shoulder height in a close front three-quarter view, placing his furrowed face just off center and the smartphone tight against his ear along the frame edge. Camera distance is the only emphasized shift; his eyes lower in concentration while the dining area remains softly legible behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 식탁 (Set with soybean-paste stew and several side-dish containers) — Its near edge crosses the background at an oblique angle; used as Forms the lower background context for the interrupted breakfast; 스마트폰 (Held tightly to his ear during the call) — Its screen side is pressed toward his face while the outer side is visible to camera; used as Acts as the immediate narrative anchor against 서의용's face; 베란다의 빨래 (Hung neatly on the balcony); used as Provides distant domestic context beyond the dining area.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright morning sunlight illuminates the living room while the close view retains a subdued naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong holds the smartphone to his ear throughout Sanghyeok's urgent call.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 귀에 스마트폰을 바짝 댄 채 미간을 살짝 찌푸린 서의용의 얼굴\n\nLOCATION (lock): In the sunlit kitchen-dining area of the detective's home, beside the breakfast table. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the dolly-in at 서의용's shoulder height in a close front three-quarter view, placing his furrowed face just off center and the smartphone tight against his ear along the frame edge. Camera distance is the only emphasized shift; his eyes lower in concentration while the dining area remains softly legible behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 식탁 (Set with soybean-paste stew and several side-dish containers) — Its near edge crosses the background at an oblique angle; used as Forms the lower background context for the interrupted breakfast; 스마트폰 (Held tightly to his ear during the call) — Its screen side is pressed toward his face while the outer side is visible to camera; used as Acts as the immediate narrative anchor against 서의용's face; 베란다의 빨래 (Hung neatly on the balcony); used as Provides distant domestic context beyond the dining area.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright morning sunlight illuminates the living room while the close view retains a subdued naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong holds the smartphone to his ear throughout Sanghyeok's urgent call.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 귀에 스마트폰을 바짝 댄 채 미간을 살짝 찌푸린 서의용의 얼굴\n\nLOCATION (lock): In the sunlit kitchen-dining area of the detective's home, beside the breakfast table. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the dolly-in at 서의용's shoulder height in a close front three-quarter view, placing his furrowed face just off center and the smartphone tight against his ear along the frame edge. Camera distance is the only emphasized shift; his eyes lower in concentration while the dining area remains softly legible behind him.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 식탁 (Set with soybean-paste stew and several side-dish containers) — Its near edge crosses the background at an oblique angle; used as Forms the lower background context for the interrupted breakfast; 스마트폰 (Held tightly to his ear during the call) — Its screen side is pressed toward his face while the outer side is visible to camera; used as Acts as the immediate narrative anchor against 서의용's face; 베란다의 빨래 (Hung neatly on the balcony); used as Provides distant domestic context beyond the dining area.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright morning sunlight illuminates the living room while the close view retains a subdued naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong holds the smartphone to his ear throughout Sanghyeok's urgent call.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S47sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S47sh5.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:867693>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_부녀_집_거실_8d3b16.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:867693>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "스마트폰의 방향(뒷면 노출), 인물의 표정, 식탁 및 발코니 빨래 등 배경 요소와 구도가 지시문과 잘 일치합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "스마트폰의 화면이 바깥을 향하도록 뒤집어 쥐고 있어 소품 방향 규칙을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선은 화면 아래쪽을 향하고 있음.",
      "built_space": "배경에 레퍼런스와 일치하는 발코니 창문과 주방 구조가 적절한 원근감으로 배치됨.",
      "entities": "인물의 외모와 의상이 일치함. 카메라 렌즈가 보이는 스마트폰 뒷면, 찌개와 반찬이 놓인 식탁, 발코니의 빨래가 모두 확인됨.",
      "hard_violations": [],
      "physics": "손이 스마트폰을 귀에 밀착하여 안정적으로 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선은 아래쪽을 향하고 있음.",
      "built_space": "배경 우측으로 식탁과 발코니가 배치되어 있음.",
      "entities": "인물의 외모/의상은 일치하나, 스마트폰의 화면(전면)이 카메라를 향해 있음. 식탁의 음식과 빨래는 확인됨.",
      "hard_violations": [
       "스마트폰을 거꾸로 들고 있음 (화면이 귀가 아닌 카메라 측을 향함)"
      ],
      "physics": "손이 스마트폰을 지탱하고 있으나 파지 방향이 잘못됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "스마트폰의 방향(뒷면 노출), 인물의 표정, 식탁 및 발코니 빨래 등 배경 요소와 구도가 지시문과 잘 일치합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "스마트폰의 화면이 바깥을 향하도록 뒤집어 쥐고 있어 소품 방향 규칙을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선은 화면 아래쪽을 향하고 있음.",
      "built_space": "배경에 레퍼런스와 일치하는 발코니 창문과 주방 구조가 적절한 원근감으로 배치됨.",
      "entities": "인물의 외모와 의상이 일치함. 카메라 렌즈가 보이는 스마트폰 뒷면, 찌개와 반찬이 놓인 식탁, 발코니의 빨래가 모두 확인됨.",
      "hard_violations": [],
      "physics": "손이 스마트폰을 귀에 밀착하여 안정적으로 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선은 아래쪽을 향하고 있음.",
      "built_space": "배경 우측으로 식탁과 발코니가 배치되어 있음.",
      "entities": "인물의 외모/의상은 일치하나, 스마트폰의 화면(전면)이 카메라를 향해 있음. 식탁의 음식과 빨래는 확인됨.",
      "hard_violations": [
       "스마트폰을 거꾸로 들고 있음 (화면이 귀가 아닌 카메라 측을 향함)"
      ],
      "physics": "손이 스마트폰을 지탱하고 있으나 파지 방향이 잘못됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "지정된 카메라 구도와 스마트폰의 방향을 정확히 구현하였으며, 인물의 표정과 배경의 배치 모두 프롬프트에 충실합니다."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "지정된 배경과 인물의 외형은 잘 반영했으나, 손가락 개수가 틀린 치명적인 해부학적 오류와 스마트폰의 방향 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (스마트폰을 잡고 있는 왼손의 바깥쪽 면에만 5개의 손가락이 묘사되어 보이지 않는 엄지까지 총 6개의 손가락을 가짐)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.429,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.179,
      "B": 1.75
     },
     "violations": {
      "A": [
       "[gemini-pro] 물리적으로 불가능한 해부학적 구조 (스마트폰을 잡고 있는 왼손의 바깥쪽 면에만 5개의 손가락이 묘사되어 보이지 않는 엄지까지 총 6개의 손가락을 가짐)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1750,
      "verdict_ko": "지정된 카메라 구도와 스마트폰의 방향을 정확히 구현하였으며, 인물의 표정과 배경의 배치 모두 프롬프트에 충실합니다."
     },
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "지정된 배경과 인물의 외형은 잘 반영했으나, 손가락 개수가 틀린 치명적인 해부학적 오류와 스마트폰의 방향 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학적 구조 (스마트폰을 잡고 있는 왼손의 바깥쪽 면에만 5개의 손가락이 묘사되어 보이지 않는 엄지까지 총 6개의 손가락을 가짐)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1757,
     "B": 1182
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선은 화면 아래쪽을 향하고 있음.",
    "built_space": "배경에 레퍼런스와 일치하는 발코니 창문과 주방 구조가 적절한 원근감으로 배치됨.",
    "entities": "인물의 외모와 의상이 일치함. 카메라 렌즈가 보이는 스마트폰 뒷면, 찌개와 반찬이 놓인 식탁, 발코니의 빨래가 모두 확인됨.",
    "hard_violations": [],
    "physics": "손이 스마트폰을 귀에 밀착하여 안정적으로 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "인물의 시선은 아래쪽을 향하고 있음.",
    "built_space": "배경 우측으로 식탁과 발코니가 배치되어 있음.",
    "entities": "인물의 외모/의상은 일치하나, 스마트폰의 화면(전면)이 카메라를 향해 있음. 식탁의 음식과 빨래는 확인됨.",
    "hard_violations": [
     "스마트폰을 거꾸로 들고 있음 (화면이 귀가 아닌 카메라 측을 향함)"
    ],
    "physics": "손이 스마트폰을 지탱하고 있으나 파지 방향이 잘못됨."
   }
  ],
  "totals": {
   "A": 1757,
   "B": 1182
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "스마트폰의 방향(뒷면 노출), 인물의 표정, 식탁 및 발코니 빨래 등 배경 요소와 구도가 지시문과 잘 일치합니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "스마트폰의 화면이 바깥을 향하도록 뒤집어 쥐고 있어 소품 방향 규칙을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_부녀_집_거실_8d3b16.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:867693>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "원본 배경을 그대로 유지하라는 지시와 달리, 주방 싱크대의 위치가 바뀌고 냉장고가 추가되는 등 배경 구조가 완전히 다르게 생성됨.",
     "fix_en": "Restore the background layout to match the reference, mitigating the error by blurring the incorrect kitchen elements on the left. Preserve the character, his pose, and the table setup.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "스마트폰을 쥐고 있는 왼손의 손가락 개수가 너무 많고 해부학적 구조가 기형적임.",
     "fix_en": "Redraw the left hand on the smartphone to have exactly four anatomically correct fingers resting on the back of the device, filling the extra finger space with the phone's metallic casing. Preserve the man's face, his clothing, the dining table, the food, and the background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "식탁 위에 지정된 '된장찌개' 대신 면발이 보이는 '라면'이 놓여 있음.",
     "fix_en": "Replace the ramen in the left bowl with a traditional Korean soybean-paste stew (doenjang jjigae), depicting a dark brown broth with chunks of tofu and vegetables instead of noodles. Preserve the character, his hand, the smartphone, the other dishes, the wooden table, and the background.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "왼쪽 식탁 상판이 배경 참조의 베이지 대리석·유리가 아니라 갈색 나무로 바뀌어 있다",
     "fix_en": "Change the dining table top from brown wood to a light beige marble surface. Preserve the character, the dishes on the table, and the background environment.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "중앙 베란다 빨래의 건조대 형태와 옷 배치가 배경 참조와 다르다",
     "fix_en": "Adjust the hanging clothes and the laundry rack to match the exact shapes and colors in the background reference. Preserve the character in the foreground and the surrounding room.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "미간이 살짝이 아니라 깊게 찌푸려져 있다",
     "fix_en": "Smooth out the deep wrinkles between the eyebrows to create a very subtle furrow, replacing the heavy creases with smooth skin. Preserve the man's overall expression, the rest of his face, his clothing, and the background.",
     "severity": "minor",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "원본 배경을 그대로 유지하라는 지시와 달리, 주방 싱크대의 위치가 바뀌고 냉장고가 추가되는 등 배경 구조가 완전히 다르게 생성됨.",
     "severity": "major"
    },
    {
     "issue_ko": "스마트폰을 쥐고 있는 왼손의 손가락 개수가 너무 많고 해부학적 구조가 기형적임.",
     "severity": "critical"
    },
    {
     "issue_ko": "식탁 위에 지정된 '된장찌개' 대신 면발이 보이는 '라면'이 놓여 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 식탁 상판이 배경 참조의 베이지 대리석·유리가 아니라 갈색 나무로 바뀌어 있다",
     "severity": "major"
    },
    {
     "issue_ko": "중앙 베란다 빨래의 건조대 형태와 옷 배치가 배경 참조와 다르다",
     "severity": "major"
    },
    {
     "issue_ko": "미간이 살짝이 아니라 깊게 찌푸려져 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "원본 배경을 그대로 유지하라는 지시와 달리, 주방 싱크대의 위치가 바뀌고 냉장고가 추가되는 등 배경 구조가 완전히 다르게 생성됨.",
    "fix_en": "Restore the background layout to match the reference, mitigating the error by blurring the incorrect kitchen elements on the left. Preserve the character, his pose, and the table setup.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "왼쪽 식탁 상판이 배경 참조의 베이지 대리석·유리가 아니라 갈색 나무로 바뀌어 있다",
    "fix_en": "Change the dining table top from brown wood to a light beige marble surface. Preserve the character, the dishes on the table, and the background environment.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "중앙 베란다 빨래의 건조대 형태와 옷 배치가 배경 참조와 다르다",
    "fix_en": "Adjust the hanging clothes and the laundry rack to match the exact shapes and colors in the background reference. Preserve the character in the foreground and the surrounding room.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "미간이 살짝이 아니라 깊게 찌푸려져 있다",
    "fix_en": "Smooth out the deep wrinkles between the eyebrows to create a very subtle furrow, replacing the heavy creases with smooth skin. Preserve the man's overall expression, the rest of his face, his clothing, and the background.",
    "severity": "minor",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the left hand on the smartphone to have exactly four anatomically correct fingers resting on the back of the device, filling the extra finger space with the phone's metallic casing. Preserve the man's face, his clothing, the dining table, the food, and the background.\n- Replace the ramen in the left bowl with a traditional Korean soybean-paste stew (doenjang jjigae), depicting a dark brown broth with chunks of tofu and vegetables instead of noodles. Preserve the character, his hand, the smartphone, the other dishes, the wooden table, and the background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 클로즈업 프레이밍과 레이아웃 스케치의 구도를 완벽하게 준수하였으며, 피사체의 찌푸린 표정과 밀착된 스마트폰, 배경 요소(식탁 위 음식, 베란다 빨래)의 얕은 심도 처리가 매우 사실적으로 구현되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "요구된 클로즈업 카메라 앵글과 레이아웃 스케치의 크기 비례를 완전히 무시하고 피사체를 전신에 가까운 와이드 샷으로 렌더링하여 프레이밍 조건(우선순위 2번)에서 크게 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 집중하여 약간 아래를 향하고 있으며, 스마트폰의 화면이 귀와 얼굴에 바짝 밀착되어 있음.",
      "built_space": "레이아웃 스케치가 지정한 대로 주방 공간 내 올바른 위치에 카메라가 자리잡고 있으며, 배경의 식탁과 베란다가 프레임에 알맞게 잘려 있음.",
      "entities": "레퍼런스와 일치하는 인물(서의용)의 얼굴, 비니, 남색 재킷이 확인되며, 스마트폰, 찌개와 반찬이 있는 식탁, 베란다의 빨래 등 모든 객체가 프롬프트대로 묘사됨.",
      "hard_violations": [],
      "physics": "손가락과 손바닥이 스마트폰을 실제 통화 시처럼 자연스럽고 안정적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "인물이 눈을 감고 있으며, 스마트폰을 귀에 대고 있음.",
      "built_space": "원본 배경 사진과 동일한 넓은 주방 공간이 그대로 노출됨.",
      "entities": "인물의 복장과 배경의 요소들은 존재하나, 프롬프트의 클로즈업 스케일에서 배제되어야 할 하반신과 배경 전체가 노출됨.",
      "hard_violations": [],
      "physics": "두 발로 바닥에 서서 스마트폰을 손으로 들고 있으며, 지탱 방식에 물리적 오류는 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지정된 클로즈업 프레이밍과 레이아웃 스케치의 구도를 완벽하게 준수하였으며, 피사체의 찌푸린 표정과 밀착된 스마트폰, 배경 요소(식탁 위 음식, 베란다 빨래)의 얕은 심도 처리가 매우 사실적으로 구현되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "요구된 클로즈업 카메라 앵글과 레이아웃 스케치의 크기 비례를 완전히 무시하고 피사체를 전신에 가까운 와이드 샷으로 렌더링하여 프레이밍 조건(우선순위 2번)에서 크게 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 집중하여 약간 아래를 향하고 있으며, 스마트폰의 화면이 귀와 얼굴에 바짝 밀착되어 있음.",
      "built_space": "레이아웃 스케치가 지정한 대로 주방 공간 내 올바른 위치에 카메라가 자리잡고 있으며, 배경의 식탁과 베란다가 프레임에 알맞게 잘려 있음.",
      "entities": "레퍼런스와 일치하는 인물(서의용)의 얼굴, 비니, 남색 재킷이 확인되며, 스마트폰, 찌개와 반찬이 있는 식탁, 베란다의 빨래 등 모든 객체가 프롬프트대로 묘사됨.",
      "hard_violations": [],
      "physics": "손가락과 손바닥이 스마트폰을 실제 통화 시처럼 자연스럽고 안정적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "인물이 눈을 감고 있으며, 스마트폰을 귀에 대고 있음.",
      "built_space": "원본 배경 사진과 동일한 넓은 주방 공간이 그대로 노출됨.",
      "entities": "인물의 복장과 배경의 요소들은 존재하나, 프롬프트의 클로즈업 스케일에서 배제되어야 할 하반신과 배경 전체가 노출됨.",
      "hard_violations": [],
      "physics": "두 발로 바닥에 서서 스마트폰을 손으로 들고 있으며, 지탱 방식에 물리적 오류는 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "프롬프트가 요구한 클로즈업 샷 사이즈와 레이아웃 스케치를 정확히 구현했으며, 미간을 찌푸리고 시선을 내린 표정과 스마트폰의 방향까지 완벽하게 표현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업(인물의 어깨 높이) 프레이밍과 레이아웃 스케치를 완전히 무시하고 전신에 가까운 와이드 샷을 렌더링하여 우선순위가 높은 카메라 지시를 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "서의용의 시선은 집중한 듯 아래를 향하고 있으며, 왼손으로 스마트폰을 왼쪽 귀에 바짝 대고 있음.",
      "built_space": "카메라가 어깨 높이에 위치하여 뒷배경의 식탁과 베란다가 아웃포커싱되어 깊이감 있게 배치됨. 기준 레퍼런스와 동일한 구조임.",
      "entities": "서의용(40대 중반 한국 남성, 비니와 남색 재킷 착용), 바깥면이 보이는 스마트폰, 된장찌개와 반찬이 놓인 식탁, 베란다의 빨래 등 프롬프트의 모든 요소가 정확히 일치함.",
      "hard_violations": [],
      "physics": "인물의 자세는 자연스러우며, 손이 스마트폰을 안정적으로 귀에 밀착시켜 쥐고 있음."
     },
     {
      "label": "A",
      "direction": "서의용의 시선은 눈을 감은 듯 아래를 향하고 있으며, 왼손으로 스마트폰을 왼쪽 귀에 대고 있음.",
      "built_space": "주방 및 식당의 구조는 기준 레퍼런스와 일치하나, 카메라 위치가 지정된 어깨 높이가 아닌 더 먼 거리에서 촬영되어 바닥까지 묘사됨.",
      "entities": "서의용, 스마트폰, 된장찌개와 반찬이 놓인 식탁, 베란다의 빨래가 존재하나, 프레이밍이 지시된 '클로즈업'이 아님.",
      "hard_violations": [],
      "physics": "인물이 바닥에 서 있으며, 왼손이 스마트폰을 쥐고 귀에 대고 있음. 물리적 오류는 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "프롬프트가 요구한 클로즈업 샷 사이즈와 레이아웃 스케치를 정확히 구현했으며, 미간을 찌푸리고 시선을 내린 표정과 스마트폰의 방향까지 완벽하게 표현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업(인물의 어깨 높이) 프레이밍과 레이아웃 스케치를 완전히 무시하고 전신에 가까운 와이드 샷을 렌더링하여 우선순위가 높은 카메라 지시를 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "서의용의 시선은 집중한 듯 아래를 향하고 있으며, 왼손으로 스마트폰을 왼쪽 귀에 바짝 대고 있음.",
      "built_space": "카메라가 어깨 높이에 위치하여 뒷배경의 식탁과 베란다가 아웃포커싱되어 깊이감 있게 배치됨. 기준 레퍼런스와 동일한 구조임.",
      "entities": "서의용(40대 중반 한국 남성, 비니와 남색 재킷 착용), 바깥면이 보이는 스마트폰, 된장찌개와 반찬이 놓인 식탁, 베란다의 빨래 등 프롬프트의 모든 요소가 정확히 일치함.",
      "hard_violations": [],
      "physics": "인물의 자세는 자연스러우며, 손이 스마트폰을 안정적으로 귀에 밀착시켜 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "서의용의 시선은 눈을 감은 듯 아래를 향하고 있으며, 왼손으로 스마트폰을 왼쪽 귀에 대고 있음.",
      "built_space": "주방 및 식당의 구조는 기준 레퍼런스와 일치하나, 카메라 위치가 지정된 어깨 높이가 아닌 더 먼 거리에서 촬영되어 바닥까지 묘사됨.",
      "entities": "서의용, 스마트폰, 된장찌개와 반찬이 놓인 식탁, 베란다의 빨래가 존재하나, 프레이밍이 지시된 '클로즈업'이 아님.",
      "hard_violations": [],
      "physics": "인물이 바닥에 서 있으며, 왼손이 스마트폰을 쥐고 귀에 대고 있음. 물리적 오류는 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S47sh5__bgfirst_bg.png",
   "bg_asset_id": "8cf8653f-7c0c-4879-abaf-d75ad08cf0e7",
   "bg_record_key": "S47sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "부녀 집 거실",
   "groupbg_asset_id": "88ce5d51-647c-40f3-a6b9-b47c38cea602"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S47sh5::cine": {
  "applied": true,
  "fingerprint": "b496f1aa100654d9b7896af8055636d5b2c2aaf1f4293b6692c36580af657448",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S47sh5_sel.png",
  "source_sha256": "c1eaa3b97022d8d66580586d78f5cb18771f947ab6b8818cd58cfe3023799328",
  "file": "S47sh5_cine.png",
  "latency_ms": 10617
 },
 "S47sh7::signage": {
  "fp": "8fcd1dd2811ea4a0",
  "inscriptions": []
 },
 "S47sh7": {
  "input_fingerprint": "a56a6a1b186173f6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 한 손에 낡은 노트를 쥔 채 허겁지겁 현관 쪽을 향해 한 발을 내디뎌 땅에서 갓 떨어진 mid-stride 자세의 서의용의 뒷모습\n\nLOCATION (lock): Across the home's living area toward the front entrance, with morning sunlight entering the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin the low rear tracking move at hip height, framing 서의용 in a medium rear three-quarter view as he drives one step toward the entryway with one foot newly lifted and the worn notebook gripped in one hand. His forward position is the sole emphasized change: the route opens ahead of him while the living room recedes behind, and his gaze stays fixed on where he is rushing.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-center of the frame, midground, moves toward entryway; entryway in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 현관 (Visible as his destination) — The entryway faces 서의용's route and sits ahead of him in the background; used as Defines the destination and forward axis of the tracking shot; 거실 가구 (Visible behind 서의용) — Furniture edges recede away from the entryway route; used as Provides a scale reference behind the moving figure; 낡은 취재 노트 (Held firmly in one hand) — Its closed cover is seen obliquely beside 서의용's moving leg; used as Serves as the handheld investigative object without overwhelming the figure.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Morning sunlight continues to brighten the living room with grounded, moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong has dressed to leave and carries the worn interview notebook as he rushes toward the door.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 한 손에 낡은 노트를 쥔 채 허겁지겁 현관 쪽을 향해 한 발을 내디뎌 땅에서 갓 떨어진 mid-stride 자세의 서의용의 뒷모습\n\nLOCATION (lock): Across the home's living area toward the front entrance, with morning sunlight entering the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin the low rear tracking move at hip height, framing 서의용 in a medium rear three-quarter view as he drives one step toward the entryway with one foot newly lifted and the worn notebook gripped in one hand. His forward position is the sole emphasized change: the route opens ahead of him while the living room recedes behind, and his gaze stays fixed on where he is rushing.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-center of the frame, midground, moves toward entryway; entryway in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 현관 (Visible as his destination) — The entryway faces 서의용's route and sits ahead of him in the background; used as Defines the destination and forward axis of the tracking shot; 거실 가구 (Visible behind 서의용) — Furniture edges recede away from the entryway route; used as Provides a scale reference behind the moving figure; 낡은 취재 노트 (Held firmly in one hand) — Its closed cover is seen obliquely beside 서의용's moving leg; used as Serves as the handheld investigative object without overwhelming the figure.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Morning sunlight continues to brighten the living room with grounded, moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong has dressed to leave and carries the worn interview notebook as he rushes toward the door.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning.\n\nSHOT TEXT (authoritative, Korean): 한 손에 낡은 노트를 쥔 채 허겁지겁 현관 쪽을 향해 한 발을 내디뎌 땅에서 갓 떨어진 mid-stride 자세의 서의용의 뒷모습\n\nLOCATION (lock): Across the home's living area toward the front entrance, with morning sunlight entering the room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin the low rear tracking move at hip height, framing 서의용 in a medium rear three-quarter view as he drives one step toward the entryway with one foot newly lifted and the worn notebook gripped in one hand. His forward position is the sole emphasized change: the route opens ahead of him while the living room recedes behind, and his gaze stays fixed on where he is rushing.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-center of the frame, midground, moves toward entryway; entryway in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 현관 (Visible as his destination) — The entryway faces 서의용's route and sits ahead of him in the background; used as Defines the destination and forward axis of the tracking shot; 거실 가구 (Visible behind 서의용) — Furniture edges recede away from the entryway route; used as Provides a scale reference behind the moving figure; 낡은 취재 노트 (Held firmly in one hand) — Its closed cover is seen obliquely beside 서의용's moving leg; used as Serves as the handheld investigative object without overwhelming the figure.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Morning sunlight continues to brighten the living room with grounded, moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong has dressed to leave and carries the worn interview notebook as he rushes toward the door.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "인물은 우측에 위치한 현관문을 향해 몸과 시선을 두고 이동 중임.",
    "built_space": "좌측에 이전 샷과 동일한 음식과 식탁이 있는 거실이며, 엉덩이 높이의 후면 로우 앵글 뷰를 취함.",
    "entities": "비니와 가죽 재킷을 입은 1명의 서의용이 오른손에 낡은 노트를 들고 있음.",
    "hard_violations": [],
    "physics": "왼발로 바닥을 딛고 오른발을 막 들어 올린 보행 자세이며, 노트는 손에 단단히 지지됨."
   },
   {
    "label": "B",
    "direction": "전경 인물은 앞을 향하고, 배경 인물은 카메라 방향으로 달려옴.",
    "built_space": "좌측에 식탁이 배치된 거실 공간임.",
    "entities": "서의용과 동일한 인물이 전경과 배경에 총 2명 등장하며, 전경 인물이 노트를 들고 있음.",
    "hard_violations": [
     "duplicated or extra bodies",
     "unsupported body"
    ],
    "physics": "배경 인물의 두 발이 모두 바닥에서 떨어져 아무런 지지 없이 허공에 떠 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "요청된 후면 3/4 앵글, 현관을 향해 걷는 자세, 낡은 노트 등 샷의 핵심 구성을 정확하고 사실적으로 구현함."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "인물이 두 명으로 복제되어 나타나고 배경 인물이 허공에 떠 있는 등 치명적인 규정 위반이 발생함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S47sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "목걸이형 경찰 배지가 인물의 앞가슴이 아닌 등 뒤에 비정상적으로 매달려 있습니다.",
     "fix_en": "Remove the police badge hanging from the character's back, filling the space with the worn black leather texture of the jacket. Preserve the person present and his position, his clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 샷 사진의 의상(회색 비니, 남색 집업)으로 고정(Lock)하라는 지시와 달리 다른 의상(남색 비니, 가죽 재킷)을 착용하고 있습니다.",
     "fix_en": "Replace the character's blue beanie, leather jacket, and jeans with the grey beanie and navy blue zip-up jacket from the previous shot. Preserve the person present and his position, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "미디엄 샷(Medium shot)으로 지시되었으나, 인물의 발끝까지 프레임에 포함되어 풀 샷에 가깝게 연출되었습니다.",
     "fix_en": "Adjust the camera closer to frame the character from the waist up in a medium shot. Preserve the person present and his position, his clothing, the set, and the light.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 중앙 상단에 배경으로 배치되어야 할 현관문이 화면 우측에 위치해 있습니다.",
     "fix_en": "Rebuild the scene geometry so the entryway is located in the upper-center background, directly ahead of the character's path. Preserve the person present and his position, his clothing, the light, and the framing.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "인물이 들고 있는 노트 표지에 해독할 수 없는 뭉개진 문자(Garbled text)가 그려져 있습니다.",
     "fix_en": "Erase the garbled text from the notebook cover, rendering it as a solid, plain, worn surface. Preserve the person present and his position, his clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "노트를 쥐고 있는 오른손의 손가락들이 하나로 뭉개져 해부학적으로 부자연스럽습니다.",
     "fix_en": "Redraw the right hand gripping the notebook to show anatomically distinct, normal fingers. Preserve the person present and his position, his clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "목걸이형 경찰 배지가 인물의 앞가슴이 아닌 등 뒤에 비정상적으로 매달려 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 샷 사진의 의상(회색 비니, 남색 집업)으로 고정(Lock)하라는 지시와 달리 다른 의상(남색 비니, 가죽 재킷)을 착용하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "미디엄 샷(Medium shot)으로 지시되었으나, 인물의 발끝까지 프레임에 포함되어 풀 샷에 가깝게 연출되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙 상단에 배경으로 배치되어야 할 현관문이 화면 우측에 위치해 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "인물이 들고 있는 노트 표지에 해독할 수 없는 뭉개진 문자(Garbled text)가 그려져 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "노트를 쥐고 있는 오른손의 손가락들이 하나로 뭉개져 해부학적으로 부자연스럽습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "서의용 옷차림이 이전 샷에 고정된 회색 비니·네이비 집업이 아니라 파란 비니와 가죽 재킷·청바지로 바뀌어 있다",
     "severity": "major"
    },
    {
     "issue_ko": "힙 높이 로우 리어 미디엄이 아니라 눈높이에 가까운 전신 프레이밍이다",
     "severity": "major"
    },
    {
     "issue_ko": "서의용이 화면 정중앙이 아니라 중우측에 있고, 현관도 상단 중앙이 아니라 우측 닫힌 문에 있다",
     "severity": "major"
    },
    {
     "issue_ko": "우측 문이 이전 샷의 밝은 문과 색·재질이 다르고 현관(단차·신발장)으로 읽히지 않는다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 6,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 샷 사진의 의상(회색 비니, 남색 집업)으로 고정(Lock)하라는 지시와 달리 다른 의상(남색 비니, 가죽 재킷)을 착용하고 있습니다.",
    "fix_en": "Replace the character's blue beanie, leather jacket, and jeans with the grey beanie and navy blue zip-up jacket from the previous shot. Preserve the person present and his position, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "미디엄 샷(Medium shot)으로 지시되었으나, 인물의 발끝까지 프레임에 포함되어 풀 샷에 가깝게 연출되었습니다.",
    "fix_en": "Adjust the camera closer to frame the character from the waist up in a medium shot. Preserve the person present and his position, his clothing, the set, and the light.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "화면 중앙 상단에 배경으로 배치되어야 할 현관문이 화면 우측에 위치해 있습니다.",
    "fix_en": "Rebuild the scene geometry so the entryway is located in the upper-center background, directly ahead of the character's path. Preserve the person present and his position, his clothing, the light, and the framing.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "인물이 들고 있는 노트 표지에 해독할 수 없는 뭉개진 문자(Garbled text)가 그려져 있습니다.",
    "fix_en": "Erase the garbled text from the notebook cover, rendering it as a solid, plain, worn surface. Preserve the person present and his position, his clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "노트를 쥐고 있는 오른손의 손가락들이 하나로 뭉개져 해부학적으로 부자연스럽습니다.",
    "fix_en": "Redraw the right hand gripping the notebook to show anatomically distinct, normal fingers. Preserve the person present and his position, his clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the police badge hanging from the character's back, filling the space with the worn black leather texture of the jacket. Preserve the person present and his position, his clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "뒷모습 앵글에 맞게 가려져야 할 배지를 자연스럽게 생략했으며, 지정된 의상과 걷는 자세, 노트 파지법을 물리적 오류 없이 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "캐릭터 레퍼런스의 배지를 무리하게 추가하려다 목이 아닌 재킷 소매에 끈이 이어지는 치명적인 구조적 오류를 범했습니다.  ★위반: [gemini-pro] 경찰 배지의 검은색 끈이 목을 감싸지 않고 가죽 재킷의 오른쪽 소매/등 부분에 물리적으로 불가능하게 융합되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.429,
      "B": 1.857
     },
     "adjusted": {
      "A": 1.179,
      "B": 1.857
     },
     "violations": {
      "A": [
       "[gemini-pro] 경찰 배지의 검은색 끈이 목을 감싸지 않고 가죽 재킷의 오른쪽 소매/등 부분에 물리적으로 불가능하게 융합되어 있음."
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.143,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "뒷모습 앵글에 맞게 가려져야 할 배지를 자연스럽게 생략했으며, 지정된 의상과 걷는 자세, 노트 파지법을 물리적 오류 없이 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "캐릭터 레퍼런스의 배지를 무리하게 추가하려다 목이 아닌 재킷 소매에 끈이 이어지는 치명적인 구조적 오류를 범했습니다.  ★위반: [gemini-pro] 경찰 배지의 검은색 끈이 목을 감싸지 않고 가죽 재킷의 오른쪽 소매/등 부분에 물리적으로 불가능하게 융합되어 있음."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "이전 샷 사진의 의상 잠금 지시를 무시하고 캐릭터 레퍼런스의 의상(가죽 재킷)을 입힌 감점 요소가 있으나, 후보 B에서 발생한 물리적 오류가 없어 최종 선정됨."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "의상 설정 오류와 더불어, 뒷모습 앵글임에도 앞가슴에 있어야 할 신분증이 등/옆구리 쪽에 부자연스럽게 매달려 있는 치명적인 물리적 오류가 발생함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 이동 방향과 시선 모두 배경 우측에 있는 현관문(닫힌 나무문)을 향하고 있음.",
      "built_space": "거실 공간으로, 왼쪽 전경에 음식이 놓인 식탁, 뒤쪽에 베란다와 빨래 건조대, 우측에 현관문이 위치함.",
      "entities": "서의용의 뒷모습. 낡은 노트를 오른손에 쥐고 있음. 두 후보 모두 이전 샷 스틸(회색 비니, 파란색 플리스)의 의상을 무시하고 캐릭터 레퍼런스의 의상(파란 비니, 가죽 재킷)을 따름.",
      "hard_violations": [],
      "physics": "오른발이 바닥을 지탱하고 왼발이 땅에서 갓 떨어진 걷는 자세이며, 오른손이 노트를 안정적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 이동 방향과 시선이 배경 우측의 현관문을 향함.",
      "built_space": "거실 공간(식탁, 베란다, 우측 현관문) 구성이 A와 동일함.",
      "entities": "서의용의 뒷모습 및 낡은 노트. A와 마찬가지로 이전 샷 스틸의 의상 잠금 지시를 어김.",
      "hard_violations": [
       "물리적으로 불가능한 구조 (앞가슴에 있어야 할 목걸이 신분증이 가죽 재킷 바깥쪽 등/옆구리에 뚫고 나온 듯 매달려 있음)"
      ],
      "physics": "걷는 자세와 노트를 쥔 손의 지지는 정상적이나, 우측 옆구리에 매달린 신분증은 물리적 연결이 불가능한 형태로 떠 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "이전 샷 사진의 의상 잠금 지시를 무시하고 캐릭터 레퍼런스의 의상(가죽 재킷)을 입힌 감점 요소가 있으나, 후보 B에서 발생한 물리적 오류가 없어 최종 선정됨."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "의상 설정 오류와 더불어, 뒷모습 앵글임에도 앞가슴에 있어야 할 신분증이 등/옆구리 쪽에 부자연스럽게 매달려 있는 치명적인 물리적 오류가 발생함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물의 이동 방향과 시선 모두 배경 우측에 있는 현관문(닫힌 나무문)을 향하고 있음.",
      "built_space": "거실 공간으로, 왼쪽 전경에 음식이 놓인 식탁, 뒤쪽에 베란다와 빨래 건조대, 우측에 현관문이 위치함.",
      "entities": "서의용의 뒷모습. 낡은 노트를 오른손에 쥐고 있음. 두 후보 모두 이전 샷 스틸(회색 비니, 파란색 플리스)의 의상을 무시하고 캐릭터 레퍼런스의 의상(파란 비니, 가죽 재킷)을 따름.",
      "hard_violations": [],
      "physics": "오른발이 바닥을 지탱하고 왼발이 땅에서 갓 떨어진 걷는 자세이며, 오른손이 노트를 안정적으로 쥐고 있음."
     },
     {
      "label": "A",
      "direction": "인물의 이동 방향과 시선이 배경 우측의 현관문을 향함.",
      "built_space": "거실 공간(식탁, 베란다, 우측 현관문) 구성이 A와 동일함.",
      "entities": "서의용의 뒷모습 및 낡은 노트. A와 마찬가지로 이전 샷 스틸의 의상 잠금 지시를 어김.",
      "hard_violations": [
       "물리적으로 불가능한 구조 (앞가슴에 있어야 할 목걸이 신분증이 가죽 재킷 바깥쪽 등/옆구리에 뚫고 나온 듯 매달려 있음)"
      ],
      "physics": "걷는 자세와 노트를 쥔 손의 지지는 정상적이나, 우측 옆구리에 매달린 신분증은 물리적 연결이 불가능한 형태로 떠 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1183,
     "B": 1863
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S47sh5"
  }
 },
 "S47sh7::cine": {
  "applied": true,
  "fingerprint": "0a479c063ebb051ddf709a11cfcf9e4bbaabf89b6eb098e5c3246c0adffe8daa",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S47sh7_sel.png",
  "source_sha256": "01b8335246501993dcb76846a410d03c055bdf16ba7c83d6abd8af6e8eb7f98b",
  "file": "S47sh7_cine.png",
  "latency_ms": 9679
 },
 "S48sh6::signage": {
  "fp": "896a9e7d76a81e52",
  "inscriptions": []
 },
 "era_assess::b112fefef90bc7a9": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 커피숍 인테리어",
    "search_terms_native": [
     "2000년대 카페 내부",
     "2010년대 커피숍 인테리어",
     "옛날 동네 카페"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성되어야 하며, 다른 언어로 번역하거나 추가적인 외국어 단어를 덧붙이지 마십시오.",
    "reason_ko": "2001~2017년 기간의 한국 카페는 목재 위주의 인테리어, 특유의 소파 좌석, 벽면 디자인 등 당대 한국 특유의 분위기가 있어 일반적인 현대식 또는 서구식 카페와 다릅니다."
   }
  ]
 },
 "era_ref::802b3a2829a2ac51": {
  "subject": "2000년대~2010년대 한국의 커피숍 인테리어",
  "terms": [
   "2000년대 카페 내부",
   "2010년대 커피숍 인테리어",
   "옛날 동네 카페"
  ],
  "queries": [
   [
    "2000년대 한국 카페 내부 옛날 동네 카페",
    "2010년대 한국 커피숍 인테리어"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://www.konest.com/varimg/photogallery/Article_Photo/16010/big_348528.jpg",
  "picked_reason_ko": "3번은 한국의 일상적인 다방형 커피숍 내부를 좌석, 테이블, 카운터, 메뉴판, 조명과 마감재까지 가장 온전히 보여 주어 형태 참고에 적합하다.",
  "sha256": "a099405089a18683095a01a7fe0668b813b835666cab31b4f41a2bea51893f69",
  "file": "eraref_802b3a2829a2ac51.png"
 },
 "S48sh6::bgfirst_bg": {
  "input_fingerprint": "2dacfb0380a016a5",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블 맞은편에 앉아 두 손으로 커피잔을 단단히 감싸 쥔 이미경의 긴장한 상체\n\nLOCATION (lock): At a café table inside the suburban coffee shop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the cut, frame 이미경 from slightly above seated chest height and from her three-quarter side, keeping her tense face and upper torso in the upper half while her hands and cup share the lower half without letting the cup exceed a natural proportion. The slightly tighter camera distance is the single emphasis; she hunches toward the cup and directs her guarded account across the table toward the unseen investigators.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 커피잔 (Held on the table with both hands) — Its side and upper rim are visible between 이미경's enclosing hands; used as Anchors both hands in the lower frame and makes her tension physically readable; 카페 테이블 (이미경 is seated on the far side) — The tabletop recedes from 이미경 toward the lower foreground; used as Separates 이미경 from the unseen interviewers across the camera axis.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime café, rendered in restrained neutral tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 커피숍 인테리어: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 테이블 맞은편에 앉아 두 손으로 커피잔을 단단히 감싸 쥔 이미경의 긴장한 상체\n\nLOCATION (lock): At a café table inside the suburban coffee shop.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: After the cut, frame 이미경 from slightly above seated chest height and from her three-quarter side, keeping her tense face and upper torso in the upper half while her hands and cup share the lower half without letting the cup exceed a natural proportion. The slightly tighter camera distance is the single emphasis; she hunches toward the cup and directs her guarded account across the table toward the unseen investigators.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 커피잔 (Held on the table with both hands) — Its side and upper rim are visible between 이미경's enclosing hands; used as Anchors both hands in the lower frame and makes her tension physically readable; 카페 테이블 (이미경 is seated on the far side) — The tabletop recedes from 이미경 toward the lower foreground; used as Separates 이미경 from the unseen interviewers across the camera axis.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime café, rendered in restrained neutral tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 커피숍 인테리어: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S48sh6__bgfirst_bg.png",
  "asset_id": "127992fa-7203-44c4-ba3c-d6626a1da291",
  "input_asset_ids": [
   "e6faba65-b5f4-47e1-b887-db3cbf61c6d5",
   "2b579b57-c12d-4b2d-bfb6-5f63aee38c93"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국의 커피숍 인테리어",
   "queries": [
    [
     "2000년대 한국 카페 내부 옛날 동네 카페",
     "2010년대 한국 커피숍 인테리어"
    ]
   ],
   "picked_url": "https://www.konest.com/varimg/photogallery/Article_Photo/16010/big_348528.jpg",
   "sha256": "a099405089a18683095a01a7fe0668b813b835666cab31b4f41a2bea51893f69",
   "file": "eraref_802b3a2829a2ac51.png"
  }
 },
 "S48sh6": {
  "input_fingerprint": "3a96d2f6ee069e84",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 맞은편에 앉아 두 손으로 커피잔을 단단히 감싸 쥔 이미경의 긴장한 상체\n\nLOCATION (lock): At a café table inside the suburban coffee shop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the cut, frame 이미경 from slightly above seated chest height and from her three-quarter side, keeping her tense face and upper torso in the upper half while her hands and cup share the lower half without letting the cup exceed a natural proportion. The slightly tighter camera distance is the single emphasis; she hunches toward the cup and directs her guarded account across the table toward the unseen investigators.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 커피잔 (Held on the table with both hands) — Its side and upper rim are visible between 이미경's enclosing hands; used as Anchors both hands in the lower frame and makes her tension physically readable; 카페 테이블 (이미경 is seated on the far side) — The tabletop recedes from 이미경 toward the lower foreground; used as Separates 이미경 from the unseen interviewers across the camera axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime café, rendered in restrained neutral tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong continues gripping the coffee cup tightly in both hands during the interview.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 이미경 right now — 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 맞은편에 앉아 두 손으로 커피잔을 단단히 감싸 쥔 이미경의 긴장한 상체\n\nLOCATION (lock): At a café table inside the suburban coffee shop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the cut, frame 이미경 from slightly above seated chest height and from her three-quarter side, keeping her tense face and upper torso in the upper half while her hands and cup share the lower half without letting the cup exceed a natural proportion. The slightly tighter camera distance is the single emphasis; she hunches toward the cup and directs her guarded account across the table toward the unseen investigators.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 커피잔 (Held on the table with both hands) — Its side and upper rim are visible between 이미경's enclosing hands; used as Anchors both hands in the lower frame and makes her tension physically readable; 카페 테이블 (이미경 is seated on the far side) — The tabletop recedes from 이미경 toward the lower foreground; used as Separates 이미경 from the unseen interviewers across the camera axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime café, rendered in restrained neutral tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong continues gripping the coffee cup tightly in both hands during the interview.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 이미경 right now — 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 맞은편에 앉아 두 손으로 커피잔을 단단히 감싸 쥔 이미경의 긴장한 상체\n\nLOCATION (lock): At a café table inside the suburban coffee shop. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the cut, frame 이미경 from slightly above seated chest height and from her three-quarter side, keeping her tense face and upper torso in the upper half while her hands and cup share the lower half without letting the cup exceed a natural proportion. The slightly tighter camera distance is the single emphasis; she hunches toward the cup and directs her guarded account across the table toward the unseen investigators.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 커피잔 (Held on the table with both hands) — Its side and upper rim are visible between 이미경's enclosing hands; used as Anchors both hands in the lower frame and makes her tension physically readable; 카페 테이블 (이미경 is seated on the far side) — The tabletop recedes from 이미경 toward the lower foreground; used as Separates 이미경 from the unseen interviewers across the camera axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime café, rendered in restrained neutral tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong continues gripping the coffee cup tightly in both hands during the interview.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 이미경 right now — 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S48sh6__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S48sh6.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L51B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1125,
      "verdict_ko": "지정된 4분의 3 측면(three-quarter side) 앵글과 인물의 긴장된 자세를 훌륭하게 구현했으며, 레퍼런스의 배경 요소를 정확하게 반영했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 스테이징이 정한 테이블 먼 쪽이 아니라 옆자리에 앉아 카메라 축 맞은편 좌석이 아님"
     },
     {
      "label": "B",
      "score": 1625,
      "verdict_ko": "캐릭터와 배경의 일치도는 높으나, 요구된 측면 앵글이 아닌 정면 위주의 구도로 촬영되어 카메라 지시사항을 어겼습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.375,
      "B": 1.625
     },
     "adjusted": {
      "A": 1.125,
      "B": 1.625
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 스테이징이 정한 테이블 먼 쪽이 아니라 옆자리에 앉아 카메라 축 맞은편 좌석이 아님"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.625,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1125,
      "verdict_ko": "지정된 4분의 3 측면(three-quarter side) 앵글과 인물의 긴장된 자세를 훌륭하게 구현했으며, 레퍼런스의 배경 요소를 정확하게 반영했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 스테이징이 정한 테이블 먼 쪽이 아니라 옆자리에 앉아 카메라 축 맞은편 좌석이 아님"
     },
     {
      "label": "B",
      "score": 1625,
      "verdict_ko": "캐릭터와 배경의 일치도는 높으나, 요구된 측면 앵글이 아닌 정면 위주의 구도로 촬영되어 카메라 지시사항을 어겼습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 미디엄 샷 프레이밍(상단에 상체, 하단에 손과 컵)과 캐릭터 및 배경 디테일을 정확히 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "카메라 앵글이 너무 넓어 치마가 화면에 포함되었으며, 요구된 타이트한 프레이밍 지시를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 테이블 맞은편(화면 좌측 밖)의 보이지 않는 대상을 향하고 있습니다.",
      "built_space": "레퍼런스의 카페 내부이며, 배경에 카운터가 보이고 캐릭터는 올바른 형태의 팔걸이 의자에 앉아 있습니다.",
      "entities": "이미경(얼굴, 머리스타일, 분홍색 가디건, 리본 블라우스 일치), 나무 테이블, 커피잔이 모두 존재합니다.",
      "hard_violations": [],
      "physics": "의자에 안정적으로 앉아 있으며, 양팔을 테이블에 지탱하고 두 손으로 컵을 자연스럽게 쥐고 있습니다."
     },
     {
      "label": "B",
      "direction": "시선은 테이블 맞은편(화면 좌측 밖)을 향하고 있습니다.",
      "built_space": "카페 내부이며 카운터와 창가 쪽 테이블들이 보입니다.",
      "entities": "이미경(얼굴, 의상, 민트색 치마 포함 일치), 나무 테이블, 커피잔이 존재합니다.",
      "hard_violations": [],
      "physics": "의자에 앉아 있으며, 테이블 위에서 두 손으로 컵을 쥐고 있는 자세가 유지되고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 미디엄 샷 프레이밍(상단에 상체, 하단에 손과 컵)과 캐릭터 및 배경 디테일을 정확히 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "카메라 앵글이 너무 넓어 치마가 화면에 포함되었으며, 요구된 타이트한 프레이밍 지시를 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 테이블 맞은편(화면 좌측 밖)의 보이지 않는 대상을 향하고 있습니다.",
      "built_space": "레퍼런스의 카페 내부이며, 배경에 카운터가 보이고 캐릭터는 올바른 형태의 팔걸이 의자에 앉아 있습니다.",
      "entities": "이미경(얼굴, 머리스타일, 분홍색 가디건, 리본 블라우스 일치), 나무 테이블, 커피잔이 모두 존재합니다.",
      "hard_violations": [],
      "physics": "의자에 안정적으로 앉아 있으며, 양팔을 테이블에 지탱하고 두 손으로 컵을 자연스럽게 쥐고 있습니다."
     },
     {
      "label": "A",
      "direction": "시선은 테이블 맞은편(화면 좌측 밖)을 향하고 있습니다.",
      "built_space": "카페 내부이며 카운터와 창가 쪽 테이블들이 보입니다.",
      "entities": "이미경(얼굴, 의상, 민트색 치마 포함 일치), 나무 테이블, 커피잔이 존재합니다.",
      "hard_violations": [],
      "physics": "의자에 앉아 있으며, 테이블 위에서 두 손으로 컵을 쥐고 있는 자세가 유지되고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1129,
     "B": 1632
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 1129,
   "B": 1632
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "지정된 4분의 3 측면(three-quarter side) 앵글과 인물의 긴장된 자세를 훌륭하게 구현했으며, 레퍼런스의 배경 요소를 정확하게 반영했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 스테이징이 정한 테이블 먼 쪽이 아니라 옆자리에 앉아 카메라 축 맞은편 좌석이 아님"
   },
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "캐릭터와 배경의 일치도는 높으나, 요구된 측면 앵글이 아닌 정면 위주의 구도로 촬영되어 카메라 지시사항을 어겼습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L51B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 우측, 커피잔을 쥔 인물의 왼쪽 손가락 개수가 정상보다 많아 보이고 관절 형태가 기형적으로 뭉개져 있음.",
     "fix_en": "Redraw the hand on the right side of the coffee cup to have exactly four anatomically normal fingers wrapped around the cup and one thumb resting near the top rim, removing the extra finger. Preserve Mi-gyeong's face, her clothing, her posture, the other hand, the coffee cup, and the cafe background exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이미경의 시선이 테이블 맞은편 조사관 쪽이 아니라 프레임 왼쪽 출입문 쪽을 향한다",
     "fix_en": "Adjust Mi-gyeong's eyes so her gaze is directed forward across the table rather than toward the far left.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "이미경이 컵 쪽으로 상체를 숙이지 않고 곧게 펴고 앉아 있다",
     "fix_en": "Redraw Mi-gyeong's posture so her shoulders and upper torso hunch forward toward the coffee cup.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "이미경 왼쪽 가슴(프레임 오른쪽)에 참조 의상의 이름표가 없다",
     "fix_en": "Add the name tag from the reference image to the left chest of Mi-gyeong's pink cardigan.",
     "severity": "minor",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 우측, 커피잔을 쥔 인물의 왼쪽 손가락 개수가 정상보다 많아 보이고 관절 형태가 기형적으로 뭉개져 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "이미경의 시선이 테이블 맞은편 조사관 쪽이 아니라 프레임 왼쪽 출입문 쪽을 향한다",
     "severity": "major"
    },
    {
     "issue_ko": "이미경이 컵 쪽으로 상체를 숙이지 않고 곧게 펴고 앉아 있다",
     "severity": "major"
    },
    {
     "issue_ko": "이미경 왼쪽 가슴(프레임 오른쪽)에 참조 의상의 이름표가 없다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "이미경의 시선이 테이블 맞은편 조사관 쪽이 아니라 프레임 왼쪽 출입문 쪽을 향한다",
    "fix_en": "Adjust Mi-gyeong's eyes so her gaze is directed forward across the table rather than toward the far left.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "이미경이 컵 쪽으로 상체를 숙이지 않고 곧게 펴고 앉아 있다",
    "fix_en": "Redraw Mi-gyeong's posture so her shoulders and upper torso hunch forward toward the coffee cup.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "이미경 왼쪽 가슴(프레임 오른쪽)에 참조 의상의 이름표가 없다",
    "fix_en": "Add the name tag from the reference image to the left chest of Mi-gyeong's pink cardigan.",
    "severity": "minor",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the hand on the right side of the coffee cup to have exactly four anatomically normal fingers wrapped around the cup and one thumb resting near the top rim, removing the extra finger. Preserve Mi-gyeong's face, her clothing, her posture, the other hand, the coffee cup, and the cafe background exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "어색했던 손가락의 구조가 자연스럽게 수정되었으며, 프롬프트가 요구한 인물의 긴장한 태도와 레퍼런스의 디테일을 안정적으로 구현함."
     },
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "지시된 구도와 배경, 인물의 외형을 잘 반영했으나, 컵을 쥔 양손의 손가락 형태와 관절이 해부학적으로 부자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.857,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.857,
      "B": 1.75
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "어색했던 손가락의 구조가 자연스럽게 수정되었으며, 프롬프트가 요구한 인물의 긴장한 태도와 레퍼런스의 디테일을 안정적으로 구현함."
     },
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "지시된 구도와 배경, 인물의 외형을 잘 반영했으나, 컵을 쥔 양손의 손가락 형태와 관절이 해부학적으로 부자연스러움."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A안의 기형적인 손가락 구조를 자연스럽게 수정하면서 캐릭터의 외형, 배경, 구도 등 모든 샷 텍스트 요구사항을 정확히 유지한 성공적인 수정본입니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "지시된 미디엄 샷 구도와 캐릭터의 외형, 배경 디테일은 잘 표현했으나, 컵을 쥔 화면 오른쪽 손의 손가락이 비정상적으로 길게 왜곡되는 치명적인 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선은 화면 왼쪽 밖의 보이지 않는 조사관을 향하고 있음.",
      "built_space": "카페 내부, 참고 이미지의 콘크리트 바닥, 카운터, 유리문 등이 올바르게 배치됨.",
      "entities": "이미경(참고 이미지와 얼굴, 분홍색 가디건, 리본 블라우스 일치), 두 손으로 쥔 커피잔, 나무 테이블 모두 존재함.",
      "hard_violations": [],
      "physics": "의자에 앉아 체중을 지탱하고 있으며, 두 팔은 테이블 위에 안정적으로 놓여 컵을 쥐고 있음."
     },
     {
      "label": "A",
      "direction": "시선은 화면 왼쪽 밖의 보이지 않는 조사관을 향하고 있음.",
      "built_space": "카페 내부, 참고 이미지의 카운터 및 공간 구조와 동일함.",
      "entities": "이미경(얼굴, 분홍 가디건, 블라우스 일치), 커피잔, 테이블 모두 지시대로 위치함.",
      "hard_violations": [
       "화면 오른쪽(캐릭터의 왼손) 검지손가락이 물리적으로 불가능할 정도로 길고 꺾여 있는 해부학적 오류"
      ],
      "physics": "의자에 앉아 테이블에 팔을 기대어 컵을 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "A안의 기형적인 손가락 구조를 자연스럽게 수정하면서 캐릭터의 외형, 배경, 구도 등 모든 샷 텍스트 요구사항을 정확히 유지한 성공적인 수정본입니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "지시된 미디엄 샷 구도와 캐릭터의 외형, 배경 디테일은 잘 표현했으나, 컵을 쥔 화면 오른쪽 손의 손가락이 비정상적으로 길게 왜곡되는 치명적인 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 화면 왼쪽 밖의 보이지 않는 조사관을 향하고 있음.",
      "built_space": "카페 내부, 참고 이미지의 콘크리트 바닥, 카운터, 유리문 등이 올바르게 배치됨.",
      "entities": "이미경(참고 이미지와 얼굴, 분홍색 가디건, 리본 블라우스 일치), 두 손으로 쥔 커피잔, 나무 테이블 모두 존재함.",
      "hard_violations": [],
      "physics": "의자에 앉아 체중을 지탱하고 있으며, 두 팔은 테이블 위에 안정적으로 놓여 컵을 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "시선은 화면 왼쪽 밖의 보이지 않는 조사관을 향하고 있음.",
      "built_space": "카페 내부, 참고 이미지의 카운터 및 공간 구조와 동일함.",
      "entities": "이미경(얼굴, 분홍 가디건, 블라우스 일치), 커피잔, 테이블 모두 지시대로 위치함.",
      "hard_violations": [
       "화면 오른쪽(캐릭터의 왼손) 검지손가락이 물리적으로 불가능할 정도로 길고 꺾여 있는 해부학적 오류"
      ],
      "physics": "의자에 앉아 테이블에 팔을 기대어 컵을 지지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1864,
     "B": 1755
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S48sh6__bgfirst_bg.png",
   "bg_asset_id": "127992fa-7203-44c4-ba3c-d6626a1da291",
   "bg_record_key": "S48sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S48sh6::cine": {
  "applied": true,
  "fingerprint": "91d93a57f8f04eb06f547e9092bec9b3a82a46d94058eea1e09eb7aefc4dedd8",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S48sh6_sel.png",
  "source_sha256": "6ea83f8c5168f922ea41fd14b0d35fbd4a6b01131efc55520fd20b87f067db52",
  "file": "S48sh6_cine.png",
  "latency_ms": 10707
 },
 "S48sh9::signage": {
  "fp": "9c03578200bb12fd",
  "inscriptions": []
 },
 "S48sh9": {
  "input_fingerprint": "ef5ba9b2f91f6cb4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이미경의 대답에 눈을 크게 뜨고 서로를 마주 보는 서의용과 나상혁의 굳은 모습\n\nLOCATION (lock): At the detectives' side of a table inside the suburban café. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the track just behind and beside 이미경's position, keeping her shoulder line outside the frame as a firm seated-eye-level two-shot finds 서의용 on the left and 나상혁 on the right across the table. Their gaze is the sole emphasized change: both hold widened eyes for an instant, then turn toward each other rather than toward the lens, their torsos remaining tense and contained.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 카페 테이블 (Both men are seated on the opposite side) — Its long axis recedes from 이미경's side toward the two men; used as Runs between the camera and the two men, preserving the interview geometry.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime café ambience remains subdued, with moderate contrast holding the men's arrested expressions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same quiet cafe table, cups, daylight, and the two investigators' clothing. Exclude the seated woman from the frame and show the investigators exchanging a startled look.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이미경의 대답에 눈을 크게 뜨고 서로를 마주 보는 서의용과 나상혁의 굳은 모습\n\nLOCATION (lock): At the detectives' side of a table inside the suburban café. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the track just behind and beside 이미경's position, keeping her shoulder line outside the frame as a firm seated-eye-level two-shot finds 서의용 on the left and 나상혁 on the right across the table. Their gaze is the sole emphasized change: both hold widened eyes for an instant, then turn toward each other rather than toward the lens, their torsos remaining tense and contained.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 카페 테이블 (Both men are seated on the opposite side) — Its long axis recedes from 이미경's side toward the two men; used as Runs between the camera and the two men, preserving the interview geometry.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime café ambience remains subdued, with moderate contrast holding the men's arrested expressions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same quiet cafe table, cups, daylight, and the two investigators' clothing. Exclude the seated woman from the frame and show the investigators exchanging a startled look.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이미경의 대답에 눈을 크게 뜨고 서로를 마주 보는 서의용과 나상혁의 굳은 모습\n\nLOCATION (lock): At the detectives' side of a table inside the suburban café. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the track just behind and beside 이미경's position, keeping her shoulder line outside the frame as a firm seated-eye-level two-shot finds 서의용 on the left and 나상혁 on the right across the table. Their gaze is the sole emphasized change: both hold widened eyes for an instant, then turn toward each other rather than toward the lens, their torsos remaining tense and contained.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 카페 테이블 (Both men are seated on the opposite side) — Its long axis recedes from 이미경's side toward the two men; used as Runs between the camera and the two men, preserving the interview geometry.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime café ambience remains subdued, with moderate contrast holding the men's arrested expressions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same quiet cafe table, cups, daylight, and the two investigators' clothing. Exclude the seated woman from the frame and show the investigators exchanging a startled look.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "서의용과 나상혁이 서로의 얼굴을 향해 시선을 정확히 맞추고 있으며, 눈을 다소 크게 뜬 상태임.",
    "built_space": "카페 테이블을 사이에 두고 두 남자가 나란히 앉아 있음. 카메라가 반대편을 향해야 하나, 레퍼런스와 동일한 카운터 및 출입문 배경이 뒤에 위치함.",
    "entities": "왼쪽의 서의용(40대, 비니, 가죽 재킷, 후드, 배지)과 오른쪽의 나상혁(30대, 짙은 정장, 연하늘색 셔츠) 모두 캐릭터 레퍼런스와 일치함. 백색 머그잔이 테이블 위에 있음.",
    "hard_violations": [],
    "physics": "두 인물 모두 의자에 안정적으로 앉아 체중을 지탱하고 있으며, 머그잔은 테이블 위에 정상적으로 놓여 있음."
   },
   {
    "label": "B",
    "direction": "두 인물이 서로를 마주보고 있으나, 눈을 크게 뜬 형태라기보다는 차분한 시선 교환에 가까움.",
    "built_space": "테이블 뒤에 두 남자가 앉아 있으며, A와 마찬가지로 이전 샷의 배경(카운터와 문)이 그대로 뒤에 배치됨.",
    "entities": "서의용과 나상혁의 인상, 연령대, 의상 모두 지정된 캐릭터 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "의자에 앉아있는 자세가 자연스럽고, 나상혁의 오른손 일부가 무릎 쪽에 놓인 것이 보이며, 테이블 위의 머그잔도 안정적임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "두 인물이 서로를 마주보며 눈을 크게 뜬 긴장된 표정을 프롬프트의 지시대로 잘 표현하였으며, 복장과 분위기 모두 훌륭하게 구현되었습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "인물들의 배치와 복장은 준수하나, 지시된 '눈을 크게 뜨고' 놀란 표정이 다소 평이하게 표현되어 프롬프트의 감정선 묘사가 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S48sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:771254>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "두 남자가 카운터를 등진 채 이미경과 같은 쪽에 앉아 테이블 맞은편이 아니다.",
     "fix_en": "Change the background behind the men to an out-of-focus cafe wall, removing the glass door and counter to imply the opposite side of the room, preserving the two men, their clothing, expressions, and the table in the foreground.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "테이블 앞 컵이 이전 스틸의 베이지색 머그가 아닌 흰색 손잡이 머그다.",
     "fix_en": "Change the white mug with a handle into a smooth, handleless beige ceramic cup holding black coffee, preserving the men, the table, and the background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "왼쪽 유리문 옆에 이전 스틸에 있던 검은 스툴이 없다.",
     "fix_en": "Add the thin black metal stand near the left glass door, preserving the men, the door, the counter, and the lighting.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라가 이미경의 뒤로 이동하여 맞은편을 바라보는 구도임에도 불구하고, 이전 컷에서 이미경 뒤에 있던 배경(왼쪽 유리문, 오른쪽 카운터)이 두 남자의 뒤에 동일하게 나타나 공간의 연속성이 어긋납니다.",
     "severity": "major"
    },
    {
     "issue_ko": "두 남자가 카운터를 등진 채 이미경과 같은 쪽에 앉아 테이블 맞은편이 아니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "테이블 앞 컵이 이전 스틸의 베이지색 머그가 아닌 흰색 손잡이 머그다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 유리문 옆에 이전 스틸에 있던 검은 스툴이 없다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "테이블 앞 컵이 이전 스틸의 베이지색 머그가 아닌 흰색 손잡이 머그다.",
    "fix_en": "Change the white mug with a handle into a smooth, handleless beige ceramic cup holding black coffee, preserving the men, the table, and the background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "왼쪽 유리문 옆에 이전 스틸에 있던 검은 스툴이 없다.",
    "fix_en": "Add the thin black metal stand near the left glass door, preserving the men, the door, the counter, and the lighting.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the background behind the men to an out-of-focus cafe wall, removing the glass door and counter to imply the opposite side of the room, preserving the two men, their clothing, expressions, and the table in the foreground.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 카메라 구도를 정확히 구현하여 이미경을 프레임에서 제외하고 두 남자가 서로 마주보는 모습을 성공적으로 연출했으며, 이전 샷의 배경을 충실히 유지했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레임에서 제외해야 할 여성이 중앙에 그대로 등장하며, 남성들이 마주보지 않고 서 있는 등 프롬프트의 지시사항을 다수 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 남자가 테이블을 사이에 두고 서로의 얼굴을 향해 시선을 맞추고 있습니다.",
      "built_space": "이전 샷과 동일한 카페 내부로, 전경에 테이블이 있고 배경에 커피 머신과 카운터, 좌측에 출입문이 올바르게 배치되어 있습니다.",
      "entities": "좌측에 서의용(비니, 가죽 재킷, 배지), 우측에 나상혁(네이비 정장)이 래퍼런스와 일치하게 구현되었으며, 지시대로 여성은 프레임에 없습니다.",
      "hard_violations": [],
      "physics": "두 남자 모두 의자에 앉아 무게 중심이 자연스럽게 실려 있으며 자세에 어색함이 없습니다."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 카메라 방향이나 앞쪽을 향해 시선을 두고 있어 마주보는 지시를 따르지 않았습니다.",
      "built_space": "벽돌 벽과 다른 형태의 테이블들이 있는 전혀 다른 카페 내부 구조를 보여줍니다.",
      "entities": "서의용과 나상혁의 외형은 래퍼런스와 유사하나, 배제해야 할 여성(이미경)이 프레임 중앙에 명확히 등장합니다.",
      "hard_violations": [
       "프레임에서 제외해야 할 인물(이미경)이 등장함",
       "지정된 좌식 투샷 카메라 앵글을 무시하고 인물들을 세움",
       "이전 샷과 위치(배경)가 일치하지 않음"
      ],
      "physics": "여성은 의자에 앉아 있고, 두 남자는 그녀 뒤에 서 있는 상태로 지면에 발을 딛고 있는 것으로 보입니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 카메라 구도를 정확히 구현하여 이미경을 프레임에서 제외하고 두 남자가 서로 마주보는 모습을 성공적으로 연출했으며, 이전 샷의 배경을 충실히 유지했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레임에서 제외해야 할 여성이 중앙에 그대로 등장하며, 남성들이 마주보지 않고 서 있는 등 프롬프트의 지시사항을 다수 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "두 남자가 테이블을 사이에 두고 서로의 얼굴을 향해 시선을 맞추고 있습니다.",
      "built_space": "이전 샷과 동일한 카페 내부로, 전경에 테이블이 있고 배경에 커피 머신과 카운터, 좌측에 출입문이 올바르게 배치되어 있습니다.",
      "entities": "좌측에 서의용(비니, 가죽 재킷, 배지), 우측에 나상혁(네이비 정장)이 래퍼런스와 일치하게 구현되었으며, 지시대로 여성은 프레임에 없습니다.",
      "hard_violations": [],
      "physics": "두 남자 모두 의자에 앉아 무게 중심이 자연스럽게 실려 있으며 자세에 어색함이 없습니다."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 카메라 방향이나 앞쪽을 향해 시선을 두고 있어 마주보는 지시를 따르지 않았습니다.",
      "built_space": "벽돌 벽과 다른 형태의 테이블들이 있는 전혀 다른 카페 내부 구조를 보여줍니다.",
      "entities": "서의용과 나상혁의 외형은 래퍼런스와 유사하나, 배제해야 할 여성(이미경)이 프레임 중앙에 명확히 등장합니다.",
      "hard_violations": [
       "프레임에서 제외해야 할 인물(이미경)이 등장함",
       "지정된 좌식 투샷 카메라 앵글을 무시하고 인물들을 세움",
       "이전 샷과 위치(배경)가 일치하지 않음"
      ],
      "physics": "여성은 의자에 앉아 있고, 두 남자는 그녀 뒤에 서 있는 상태로 지면에 발을 딛고 있는 것으로 보입니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레임에서 제외해야 할 여성을 중앙에 배치했고, 두 남성이 앉아서 서로를 보지 않고 서서 카메라를 응시하여 지시사항을 전면 위반했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "여성을 프레임에서 제외하고 두 남성이 테이블에 앉아 서로 마주보며 놀란 표정을 짓는 연출과 이전 샷의 배경을 정확히 구현했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "세 인물 모두 카메라 렌즈를 정면으로 응시하고 있으며, 서로 마주보지 않음.",
      "built_space": "테이블 앞에 여성이 앉아 있고 그 뒤로 두 남성이 서 있음. 배경 구조가 이전 샷과 다름.",
      "entities": "서의용과 나상혁 외에 명시적으로 제외해야 할 여성(이미경)이 프레임 중앙에 포함됨.",
      "hard_violations": [
       "프레임에서 제외해야 할 인물(이미경) 포함",
       "인물들이 앉아있지 않고 서 있음 (스테이징 위반)",
       "지정된 시선 방향 위반 (서로 마주보지 않고 카메라 응시)"
      ],
      "physics": "남성들은 바닥에 서 있고 여성은 의자에 지지해 앉아 있음."
     },
     {
      "label": "B",
      "direction": "서의용(왼쪽)과 나상혁(오른쪽)이 눈을 크게 뜨고 서로를 마주 보고 있음.",
      "built_space": "이전 샷과 동일한 카페 배경(문, 카운터, 커피 머신 등)에서 테이블 맞은편에 두 남성이 나란히 착석함.",
      "entities": "여성이 프레임에서 제외되었고, 서의용과 나상혁만 지시된 위치에 나타남. 복장과 외모가 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "두 남성 모두 의자에 안정적으로 앉아 체중을 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레임에서 제외해야 할 여성을 중앙에 배치했고, 두 남성이 앉아서 서로를 보지 않고 서서 카메라를 응시하여 지시사항을 전면 위반했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "여성을 프레임에서 제외하고 두 남성이 테이블에 앉아 서로 마주보며 놀란 표정을 짓는 연출과 이전 샷의 배경을 정확히 구현했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "세 인물 모두 카메라 렌즈를 정면으로 응시하고 있으며, 서로 마주보지 않음.",
      "built_space": "테이블 앞에 여성이 앉아 있고 그 뒤로 두 남성이 서 있음. 배경 구조가 이전 샷과 다름.",
      "entities": "서의용과 나상혁 외에 명시적으로 제외해야 할 여성(이미경)이 프레임 중앙에 포함됨.",
      "hard_violations": [
       "프레임에서 제외해야 할 인물(이미경) 포함",
       "인물들이 앉아있지 않고 서 있음 (스테이징 위반)",
       "지정된 시선 방향 위반 (서로 마주보지 않고 카메라 응시)"
      ],
      "physics": "남성들은 바닥에 서 있고 여성은 의자에 지지해 앉아 있음."
     },
     {
      "label": "A",
      "direction": "서의용(왼쪽)과 나상혁(오른쪽)이 눈을 크게 뜨고 서로를 마주 보고 있음.",
      "built_space": "이전 샷과 동일한 카페 배경(문, 카운터, 커피 머신 등)에서 테이블 맞은편에 두 남성이 나란히 착석함.",
      "entities": "여성이 프레임에서 제외되었고, 서의용과 나상혁만 지시된 위치에 나타남. 복장과 외모가 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "두 남성 모두 의자에 안정적으로 앉아 체중을 지지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S48sh6"
  }
 },
 "S48sh9::cine": {
  "applied": true,
  "fingerprint": "a114965797aa92427f6878f66f413eef3c622946fa74503e5a690b7b703e90fb",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S48sh9_sel.png",
  "source_sha256": "f603fa026fb985ba3f3464ff04c0dfcdc2750f7f28f08d418884c6895fe59258",
  "file": "S48sh9_cine.png",
  "latency_ms": 13224
 },
 "S48sh13::signage": {
  "fp": "15b49b1360d1dc52",
  "inscriptions": []
 },
 "S48sh13": {
  "input_fingerprint": "96d5ad149d2a9e53",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용과 나상혁 쪽으로 시선을 든 채 멈춰 있는 이미경의 결연한 얼굴\n\nLOCATION (lock): At the witness's seat across the café table from the detectives. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from the men's side at a slightly low angle, looking past 서의용's soft shoulder at lower left and 나상혁's soft shoulder at lower right toward 이미경's centered three-quarter face. Camera distance is the single emphasized shift; 이미경 lifts and holds her eyes toward the two men, her stillness reading as a settled decision rather than a lens-directed pose.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 서의용 shoulder line in the lower-left of the frame, foreground; 나상혁 shoulder line in the lower-right of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward the two investigators.\n- KEY BACKGROUND ELEMENTS: 카페 테이블 (The three participants remain positioned around it) — Its far edge runs beneath 이미경 and recedes toward the foreground men; used as Provides a narrow lower-edge spatial cue linking 이미경 to the two foreground shoulders.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime café ambience models 이미경's face gently while retaining restrained contrast and neutral color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same cafe interior, table setting, soft daylight, and the woman's clothing. Exclude the investigators from the close frame and retain only her resolved expression.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용과 나상혁 쪽으로 시선을 든 채 멈춰 있는 이미경의 결연한 얼굴\n\nLOCATION (lock): At the witness's seat across the café table from the detectives. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from the men's side at a slightly low angle, looking past 서의용's soft shoulder at lower left and 나상혁's soft shoulder at lower right toward 이미경's centered three-quarter face. Camera distance is the single emphasized shift; 이미경 lifts and holds her eyes toward the two men, her stillness reading as a settled decision rather than a lens-directed pose.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 서의용 shoulder line in the lower-left of the frame, foreground; 나상혁 shoulder line in the lower-right of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward the two investigators.\n- KEY BACKGROUND ELEMENTS: 카페 테이블 (The three participants remain positioned around it) — Its far edge runs beneath 이미경 and recedes toward the foreground men; used as Provides a narrow lower-edge spatial cue linking 이미경 to the two foreground shoulders.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime café ambience models 이미경's face gently while retaining restrained contrast and neutral color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same cafe interior, table setting, soft daylight, and the woman's clothing. Exclude the investigators from the close frame and retain only her resolved expression.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용과 나상혁 쪽으로 시선을 든 채 멈춰 있는 이미경의 결연한 얼굴\n\nLOCATION (lock): At the witness's seat across the café table from the detectives. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from the men's side at a slightly low angle, looking past 서의용's soft shoulder at lower left and 나상혁's soft shoulder at lower right toward 이미경's centered three-quarter face. Camera distance is the single emphasized shift; 이미경 lifts and holds her eyes toward the two men, her stillness reading as a settled decision rather than a lens-directed pose.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 서의용 shoulder line in the lower-left of the frame, foreground; 나상혁 shoulder line in the lower-right of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward the two investigators.\n- KEY BACKGROUND ELEMENTS: 카페 테이블 (The three participants remain positioned around it) — Its far edge runs beneath 이미경 and recedes toward the foreground men; used as Provides a narrow lower-edge spatial cue linking 이미경 to the two foreground shoulders.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime café ambience models 이미경's face gently while retaining restrained contrast and neutral color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same cafe interior, table setting, soft daylight, and the woman's clothing. Exclude the investigators from the close frame and retain only her resolved expression.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.857,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.857,
    "B": 1.667
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1667,
   "A": 1857
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "약간의 로우 앵글과 인물의 결연한 표정이 샷 텍스트의 지시를 잘 구현했으며, 전체적인 구도와 조명이 영화적인 느낌을 잘 살렸습니다."
   },
   {
    "label": "A",
    "score": 1857,
    "verdict_ko": "프레임과 위치 설정은 양호하나, 결연한 표정의 연출이 다소 부족하고 명찰에 생성된 영어 텍스트가 몰입을 방해합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S48sh9_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이미경의 옷에 달린 명찰에 한국 배경에 맞지 않는 'TRX.ALAN'이라는 임의의 영문 텍스트가 생성됨",
     "fix_en": "Replace the 'TRX.ALAN' text on the name tag with illegible smudges. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 이미경의 상반신과 카페 배경이 다 보이는 미디엄 샷이다",
     "fix_en": "Crop the image to a close-up of the woman's face. Preserve the people present, their clothing, the set, and the light.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "하단 좌우의 부드러운 어깨선이 아니라 두 남성의 뒷머리와 등이 프레임 좌우를 크게 채운다",
     "fix_en": "Lower the foreground men to show only their shoulders, removing their heads. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이전 스틸에서 이미경 쪽에 있던 커피잔이 전경 왼쪽 남성 앞 테이블 가장자리에 있다",
     "fix_en": "Move the white coffee cup to the center of the table. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이미경의 옷에 달린 명찰에 한국 배경에 맞지 않는 'TRX.ALAN'이라는 임의의 영문 텍스트가 생성됨",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 이미경의 상반신과 카페 배경이 다 보이는 미디엄 샷이다",
     "severity": "major"
    },
    {
     "issue_ko": "하단 좌우의 부드러운 어깨선이 아니라 두 남성의 뒷머리와 등이 프레임 좌우를 크게 채운다",
     "severity": "major"
    },
    {
     "issue_ko": "이미경 가슴 명찰에 장소·시대와 맞지 않는 영어 TRX ALAN이 적혀 있다",
     "severity": "major"
    },
    {
     "issue_ko": "이전 스틸에서 이미경 쪽에 있던 커피잔이 전경 왼쪽 남성 앞 테이블 가장자리에 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "이미경의 옷에 달린 명찰에 한국 배경에 맞지 않는 'TRX.ALAN'이라는 임의의 영문 텍스트가 생성됨",
    "fix_en": "Replace the 'TRX.ALAN' text on the name tag with illegible smudges. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "얼굴 클로즈업이 아니라 이미경의 상반신과 카페 배경이 다 보이는 미디엄 샷이다",
    "fix_en": "Crop the image to a close-up of the woman's face. Preserve the people present, their clothing, the set, and the light.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "하단 좌우의 부드러운 어깨선이 아니라 두 남성의 뒷머리와 등이 프레임 좌우를 크게 채운다",
    "fix_en": "Lower the foreground men to show only their shoulders, removing their heads. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "이전 스틸에서 이미경 쪽에 있던 커피잔이 전경 왼쪽 남성 앞 테이블 가장자리에 있다",
    "fix_en": "Move the white coffee cup to the center of the table. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S48sh9"
  }
 },
 "S48sh13::cine": {
  "applied": true,
  "fingerprint": "9f1e96899fdda3d5233e19a1c962ab78403b60e5715436ce3cf8e0b61d59b4d5",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S48sh13_sel.png",
  "source_sha256": "ffc66d1c3d564572bbc54b946b5555f85c30eecb8edbcbcd934df18cc1e0903e",
  "file": "S48sh13_cine.png",
  "latency_ms": 10993
 },
 "S49sh10::signage": {
  "fp": "9b7d4280c58a70c4",
  "inscriptions": [
   {
    "surface_native": "서류철 표지",
    "text_native": "사건기록",
    "reason_ko": "한국 강력반 형사의 어수선한 책상 위 사실감을 극대화하기 위해 배치된 수사 서류 봉투의 표기입니다."
   }
  ]
 },
 "S49sh10": {
  "input_fingerprint": "9d0fd49023a2240e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 꼬리뼈 부분에 피가 고인 부검 사진이 띄워진 노트북을 서의용 쪽을 향해 양팔을 뻗어 밀어내고 있는 나상혁의 손과 책상 위를 미끄러지는 노트북\n\nLOCATION (lock): At the cluttered detective's desk inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track laterally at desk height, perpendicular to the laptop's path, in a tight insert that keeps 나상혁's extended forearms and pushing hands beside the device while 서의용 remains a partial destination at the opposite edge. Camera distance is the sole emphasized change; the laptop slides naturally across the desk with its screen angled enough to show the clinical photograph without sensational enlargement, and both men concentrate downward on the evidence.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 노트북 화면 (On and sliding across the desk toward 서의용) — The active display faces obliquely toward the camera and shows an autopsy photograph with a small area of pooled blood near the tailbone; used as Carries the evidence across the center of the insert while remaining within realistic scale beside both men's hands; 책상 위 (Supports the moving laptop); used as Provides the movement plane and scale reference around the laptop and forearms.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime office keeps the evidence legible without graphic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same detective-office desk area, cluttered records, fluorescent light, and workstation layout. Exclude the earlier standing pose and show the laptop sliding toward the seated detective.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop remains open to the autopsy image showing pooled blood near Seonyeong's tailbone as Sanghyeok slides it toward Euiyong.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 나상혁 right now — 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"사건기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 꼬리뼈 부분에 피가 고인 부검 사진이 띄워진 노트북을 서의용 쪽을 향해 양팔을 뻗어 밀어내고 있는 나상혁의 손과 책상 위를 미끄러지는 노트북\n\nLOCATION (lock): At the cluttered detective's desk inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track laterally at desk height, perpendicular to the laptop's path, in a tight insert that keeps 나상혁's extended forearms and pushing hands beside the device while 서의용 remains a partial destination at the opposite edge. Camera distance is the sole emphasized change; the laptop slides naturally across the desk with its screen angled enough to show the clinical photograph without sensational enlargement, and both men concentrate downward on the evidence.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 노트북 화면 (On and sliding across the desk toward 서의용) — The active display faces obliquely toward the camera and shows an autopsy photograph with a small area of pooled blood near the tailbone; used as Carries the evidence across the center of the insert while remaining within realistic scale beside both men's hands; 책상 위 (Supports the moving laptop); used as Provides the movement plane and scale reference around the laptop and forearms.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime office keeps the evidence legible without graphic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same detective-office desk area, cluttered records, fluorescent light, and workstation layout. Exclude the earlier standing pose and show the laptop sliding toward the seated detective.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop remains open to the autopsy image showing pooled blood near Seonyeong's tailbone as Sanghyeok slides it toward Euiyong.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 나상혁 right now — 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"사건기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 꼬리뼈 부분에 피가 고인 부검 사진이 띄워진 노트북을 서의용 쪽을 향해 양팔을 뻗어 밀어내고 있는 나상혁의 손과 책상 위를 미끄러지는 노트북\n\nLOCATION (lock): At the cluttered detective's desk inside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track laterally at desk height, perpendicular to the laptop's path, in a tight insert that keeps 나상혁's extended forearms and pushing hands beside the device while 서의용 remains a partial destination at the opposite edge. Camera distance is the sole emphasized change; the laptop slides naturally across the desk with its screen angled enough to show the clinical photograph without sensational enlargement, and both men concentrate downward on the evidence.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 노트북 화면 (On and sliding across the desk toward 서의용) — The active display faces obliquely toward the camera and shows an autopsy photograph with a small area of pooled blood near the tailbone; used as Carries the evidence across the center of the insert while remaining within realistic scale beside both men's hands; 책상 위 (Supports the moving laptop); used as Provides the movement plane and scale reference around the laptop and forearms.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime office keeps the evidence legible without graphic emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same detective-office desk area, cluttered records, fluorescent light, and workstation layout. Exclude the earlier standing pose and show the laptop sliding toward the seated detective.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop remains open to the autopsy image showing pooled blood near Seonyeong's tailbone as Sanghyeok slides it toward Euiyong.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 나상혁 right now — 나상혁's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 나상혁 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"사건기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "양팔이 화면 우측의 서의용 쪽을 향해 뻗어 있으며 노트북을 밀고 있음.",
    "built_space": "어수선한 경찰서 책상 위, 서류 더미와 모니터 등 이전 샷의 배경 요소들이 잘 배치됨.",
    "entities": "나상혁의 팔은 레퍼런스와 일치하는 네이비 수트를 입고 있으며, 노트북 화면에는 지정된 부검 사진이 떠 있음. 우측 가장자리에 서의용이 회색 셔츠를 입고 부분적으로 보임. 서류철에 '사건기록' 글자가 정확히 적혀 있음.",
    "hard_violations": [],
    "physics": "손은 노트북 하단부에 접촉해 지지받고 있으며, 노트북은 서류 더미와 책상 위에 안정적으로 놓여 있음."
   },
   {
    "label": "A",
    "direction": "시선은 노트북 화면을 향하고 있으며, 양손은 화면 좌측을 향해 뻗어 있음.",
    "built_space": "책상 위에 서류와 노트북이 놓여 있으며 뒤쪽으로 사무실 배경이 보임.",
    "entities": "노트북 화면에 부검 사진이 떠 있고 '사건기록' 글자가 서류에 적혀 있으나, 나상혁이 지정된 수트가 아닌 파란색 셔츠를 입고 있어 복장 레퍼런스와 불일치함.",
    "hard_violations": [
     "미는 동작을 수행해야 할 양손이 노트북에 닿지 않고 허공에 떠 있음 (물리적 지지 없음)"
    ],
    "physics": "노트북은 책상 위에 놓여 있으나, 노트북을 밀어내는 양손이 어떤 표면이나 물체에도 닿지 않은 채 허공에 떠 있어 동작의 물리적 개연성이 없음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 타이트한 인서트 샷 구도를 완벽하게 따랐으며, 나상혁의 의상(네이비 수트)과 부분적으로 보이는 서의용의 배치가 지시사항과 정확히 일치합니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "프레임이 너무 넓어 인서트 샷 지시를 어겼고, 나상혁의 의상이 레퍼런스와 다르며, 노트북을 미는 양손이 허공에 떠 있는 물리적 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S24sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:771254>"
   },
   {
    "label": "PROP REFERENCE — 피해자 부검 사진 및 복사본: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1017104>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "노트북이 책상 면을 미끄러지지 않고 서류 더미 위에 올려진 채 손이 키보드에 얹혀 있다.",
     "fix_en": "Remove the stack of papers beneath the laptop so it rests flat on the wooden desk surface. Redraw the visible hands to rest against the laptop's left edge, actively pushing it to the right, instead of resting on the keyboard. Preserve the laptop's screen content, the dark suit sleeves, the person on the right, and all surrounding desk clutter.",
     "severity": "critical",
     "observation_index": 4
    },
    {
     "issue_ko": "화면 우측 하단의 흰색 문서 표면에 해독할 수 없는 뭉개진 가짜 글자들이 생성되어 있습니다.",
     "fix_en": "Remove the illegible gibberish text from the white document in the lower right corner, leaving the paper surface blank. Preserve the adjacent documents, laptop, and hands.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "오른쪽 가장자리의 인물이 레퍼런스 이미지 속 인물과 동일한 디자인의 셔츠를 입고 있어 의상 복사 금지 지시를 위반했습니다.",
     "fix_en": "Change the light blue shirt and tie on the person at the right edge to a different outfit, such as a dark grey jacket, avoiding replication of the reference clothing. Preserve the person's placement, the laptop, and the desk set.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "노트북이 책상 표면을 미끄러지는 대신 두꺼운 서류 더미 위에 올려져 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "나상혁의 손이 노트북을 밀어내지 않고, 한 손은 키보드 위에 얹혀 있으며 다른 한 손은 노트북 뒤편 책상에 놓여 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 가장자리의 인물이 레퍼런스 이미지 속 인물과 동일한 디자인의 셔츠를 입고 있어 의상 복사 금지 지시를 위반했습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 우측 하단의 흰색 문서 표면에 해독할 수 없는 뭉개진 가짜 글자들이 생성되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "노트북이 책상 면을 미끄러지지 않고 서류 더미 위에 올려진 채 손이 키보드에 얹혀 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "샷 텍스트가 허용하지 않은 서의용의 상반신·넥타이가 프레임 오른쪽 가장자리에 보인다.",
     "severity": "major"
    },
    {
     "issue_ko": "두 사람이 아래로 증거를 보지 않고 노트북 화면이 카메라를 향해 열려 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 우측 하단의 흰색 문서 표면에 해독할 수 없는 뭉개진 가짜 글자들이 생성되어 있습니다.",
    "fix_en": "Remove the illegible gibberish text from the white document in the lower right corner, leaving the paper surface blank. Preserve the adjacent documents, laptop, and hands.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "오른쪽 가장자리의 인물이 레퍼런스 이미지 속 인물과 동일한 디자인의 셔츠를 입고 있어 의상 복사 금지 지시를 위반했습니다.",
    "fix_en": "Change the light blue shirt and tie on the person at the right edge to a different outfit, such as a dark grey jacket, avoiding replication of the reference clothing. Preserve the person's placement, the laptop, and the desk set.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the stack of papers beneath the laptop so it rests flat on the wooden desk surface. Redraw the visible hands to rest against the laptop's left edge, actively pushing it to the right, instead of resting on the keyboard. Preserve the laptop's screen content, the dark suit sleeves, the person on the right, and all surrounding desk clutter.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "노트북이 책상 표면 위를 미끄러지는 물리적 상황을 정확히 구현하였으며, 인물의 복장과 프레이밍, 소품의 디테일이 지시사항과 잘 일치함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "노트북이 두꺼운 서류 더미 위에 올려져 있어 '책상 위를 미끄러지는' 프롬프트의 핵심 물리적 행동을 불가능하게 만듦."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "양팔이 노트북을 우측에 앉은 인물(서의용)을 향해 밀어내고 있음.",
      "built_space": "경찰서 사무실 책상 위, 노트북이 책상이 아닌 두꺼운 서류 더미 위에 올려져 있음.",
      "entities": "나상혁의 양팔(남색 정장, 하늘색 셔츠), 부검 사진이 띄워진 노트북, 서의용의 신체 일부(우측), '사건기록' 표지가 적힌 서류철.",
      "hard_violations": [
       "물리적 불가능성/스테이징 위반: 노트북이 울퉁불퉁하고 두꺼운 서류 더미 위에 위치해 있어 프롬프트가 지시한 '책상 위를 미끄러지는' 액션이 불가능함."
      ],
      "physics": "노트북이 서류 더미 위에 얹혀 있어 불안정하며, 매끄럽게 미끄러지는 동작을 뒷받침할 평면이 제공되지 않음."
     },
     {
      "label": "B",
      "direction": "양팔이 노트북을 우측에 앉은 인물(서의용)을 향해 밀어내고 있음.",
      "built_space": "경찰서 사무실 책상 위, 노트북이 나무 책상 표면 위에 바르게 놓여 있음.",
      "entities": "나상혁의 양팔(남색 정장, 하늘색 셔츠), 부검 사진이 띄워진 노트북, 서의용의 신체 일부(우측), '사건기록' 표지가 적힌 서류철.",
      "hard_violations": [],
      "physics": "노트북이 책상 평면 위에 닿아 있어 밀어내는 힘에 의해 미끄러지는 물리적 상태가 자연스럽게 구현됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "노트북이 책상 표면 위를 미끄러지는 물리적 상황을 정확히 구현하였으며, 인물의 복장과 프레이밍, 소품의 디테일이 지시사항과 잘 일치함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "노트북이 두꺼운 서류 더미 위에 올려져 있어 '책상 위를 미끄러지는' 프롬프트의 핵심 물리적 행동을 불가능하게 만듦."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "양팔이 노트북을 우측에 앉은 인물(서의용)을 향해 밀어내고 있음.",
      "built_space": "경찰서 사무실 책상 위, 노트북이 책상이 아닌 두꺼운 서류 더미 위에 올려져 있음.",
      "entities": "나상혁의 양팔(남색 정장, 하늘색 셔츠), 부검 사진이 띄워진 노트북, 서의용의 신체 일부(우측), '사건기록' 표지가 적힌 서류철.",
      "hard_violations": [
       "물리적 불가능성/스테이징 위반: 노트북이 울퉁불퉁하고 두꺼운 서류 더미 위에 위치해 있어 프롬프트가 지시한 '책상 위를 미끄러지는' 액션이 불가능함."
      ],
      "physics": "노트북이 서류 더미 위에 얹혀 있어 불안정하며, 매끄럽게 미끄러지는 동작을 뒷받침할 평면이 제공되지 않음."
     },
     {
      "label": "B",
      "direction": "양팔이 노트북을 우측에 앉은 인물(서의용)을 향해 밀어내고 있음.",
      "built_space": "경찰서 사무실 책상 위, 노트북이 나무 책상 표면 위에 바르게 놓여 있음.",
      "entities": "나상혁의 양팔(남색 정장, 하늘색 셔츠), 부검 사진이 띄워진 노트북, 서의용의 신체 일부(우측), '사건기록' 표지가 적힌 서류철.",
      "hard_violations": [],
      "physics": "노트북이 책상 평면 위에 닿아 있어 밀어내는 힘에 의해 미끄러지는 물리적 상태가 자연스럽게 구현됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "책상 표면 위에서 노트북을 밀어내는 물리적 동작과 구도를 텍스트의 지시대로 정확하게 구현하였으며, 인물의 의상과 프롭도 훌륭하게 일치합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "노트북이 높게 쌓인 서류 더미 위에 올려져 있어 '책상 위를 미끄러지는 노트북'이라는 핵심 동작 지시를 심각하게 훼손하였습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "양팔과 노트북이 화면 우측 가장자리에 있는 인물(서의용)을 향해 정확히 방향을 잡고 있습니다.",
      "built_space": "경찰서 책상 위를 배경으로 하며, 뒤쪽의 모니터와 사무용품 배치가 이전 샷의 구도와 자연스럽게 이어집니다.",
      "entities": "네이비 정장을 입은 나상혁의 팔, 회색 셔츠를 입은 서의용의 우측 가장자리 일부, 꼬리뼈 부검 사진이 띄워진 노트북, '사건기록'이라 적힌 서류철이 모두 명확히 확인됩니다.",
      "hard_violations": [],
      "physics": "노트북이 책상 표면에 안정적으로 맞닿아 있으며, 두 손이 노트북 양옆을 쥐고 밀어내는 동작을 물리적으로 올바르게 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "양팔과 노트북이 화면 우측의 인물(서의용)을 향해 뻗어 있습니다.",
      "built_space": "책상과 배경의 모니터, 집기 배치가 이전 샷의 공간적 특징을 잘 반영하고 있습니다.",
      "entities": "나상혁의 정장 팔, 서의용의 셔츠 일부, 부검 사진이 띄워진 노트북, '사건기록' 서류철이 존재합니다.",
      "hard_violations": [],
      "physics": "노트북이 책상이 아닌 높게 쌓인 흩어진 서류 더미 위에 얹혀 있어, 지시된 '미끄러지는' 동작을 수행하기에 물리적으로 매우 불안정하고 앞뒤가 맞지 않는 지탱 상태입니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "책상 표면 위에서 노트북을 밀어내는 물리적 동작과 구도를 텍스트의 지시대로 정확하게 구현하였으며, 인물의 의상과 프롭도 훌륭하게 일치합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "노트북이 높게 쌓인 서류 더미 위에 올려져 있어 '책상 위를 미끄러지는 노트북'이라는 핵심 동작 지시를 심각하게 훼손하였습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "양팔과 노트북이 화면 우측 가장자리에 있는 인물(서의용)을 향해 정확히 방향을 잡고 있습니다.",
      "built_space": "경찰서 책상 위를 배경으로 하며, 뒤쪽의 모니터와 사무용품 배치가 이전 샷의 구도와 자연스럽게 이어집니다.",
      "entities": "네이비 정장을 입은 나상혁의 팔, 회색 셔츠를 입은 서의용의 우측 가장자리 일부, 꼬리뼈 부검 사진이 띄워진 노트북, '사건기록'이라 적힌 서류철이 모두 명확히 확인됩니다.",
      "hard_violations": [],
      "physics": "노트북이 책상 표면에 안정적으로 맞닿아 있으며, 두 손이 노트북 양옆을 쥐고 밀어내는 동작을 물리적으로 올바르게 지탱하고 있습니다."
     },
     {
      "label": "A",
      "direction": "양팔과 노트북이 화면 우측의 인물(서의용)을 향해 뻗어 있습니다.",
      "built_space": "책상과 배경의 모니터, 집기 배치가 이전 샷의 공간적 특징을 잘 반영하고 있습니다.",
      "entities": "나상혁의 정장 팔, 서의용의 셔츠 일부, 부검 사진이 띄워진 노트북, '사건기록' 서류철이 존재합니다.",
      "hard_violations": [],
      "physics": "노트북이 책상이 아닌 높게 쌓인 흩어진 서류 더미 위에 얹혀 있어, 지시된 '미끄러지는' 동작을 수행하기에 물리적으로 매우 불안정하고 앞뒤가 맞지 않는 지탱 상태입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S24sh5"
  }
 },
 "S49sh10::cine": {
  "applied": false,
  "fingerprint": "e25bb301ec901551b28061722008c809fe9f4151b739d751756d0996ef29dda8",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S49sh10_sel.png",
  "source_sha256": "6c5083c198a7a0a7f8a166e8e8db4939b1de02811ec00aac61fd40891e9b4f28",
  "error": "RuntimeError: moderation blocked: {\"message\": \"Generated image rejected by content moderation.\", \"code\": 400}"
 },
 "S49sh14::signage": {
  "fp": "08d145b85a1abdad",
  "inscriptions": [
   {
    "surface_native": "다이어리 페이지",
    "text_native": "10/24 (화) 23:30 삼거리",
    "reason_ko": "서의용이 응시하는 다이어리 페이지에 사건의 단서가 되는 일시와 장소가 손글씨로 적혀 있어 장면에 개연성을 부여합니다."
   }
  ]
 },
 "S49sh14": {
  "input_fingerprint": "c15e97339568d2bb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 쥔 다이어리 페이지를 향해 시선을 꽂은 채 눈을 번쩍 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the keepsake box and desks in a corner of the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly-in slightly below 서의용's eye line from the three-quarter side between his face and the diary, allowing his widened eyes to occupy most of the close frame while the open page remains along the lower edge of his sightline. His gaze is the sole emphasized change: it locks downward onto the discovered entry as his body pitches toward the page, never turning toward the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 다이어리 페이지 (Held open in 서의용's hand) — The open written face angles upward toward 서의용 and partly toward the camera, with the discovered date and '매직~' entry visible; used as Remains at the lower frame edge as the visible source of 서의용's recognition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime office isolates the recognition while preserving natural facial tone and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same cluttered office desk, records, fluorescent lighting, and the detective's clothing. Exclude the laptop-moving action and isolate him studying the diary.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's keepsake box remains in the office corner, and Euiyong has taken her diary from it and is examining the pages.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다이어리 페이지: \"10/24 (화) 23:30 삼거리\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 쥔 다이어리 페이지를 향해 시선을 꽂은 채 눈을 번쩍 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the keepsake box and desks in a corner of the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly-in slightly below 서의용's eye line from the three-quarter side between his face and the diary, allowing his widened eyes to occupy most of the close frame while the open page remains along the lower edge of his sightline. His gaze is the sole emphasized change: it locks downward onto the discovered entry as his body pitches toward the page, never turning toward the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 다이어리 페이지 (Held open in 서의용's hand) — The open written face angles upward toward 서의용 and partly toward the camera, with the discovered date and '매직~' entry visible; used as Remains at the lower frame edge as the visible source of 서의용's recognition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime office isolates the recognition while preserving natural facial tone and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same cluttered office desk, records, fluorescent lighting, and the detective's clothing. Exclude the laptop-moving action and isolate him studying the diary.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's keepsake box remains in the office corner, and Euiyong has taken her diary from it and is examining the pages.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다이어리 페이지: \"10/24 (화) 23:30 삼거리\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 손에 쥔 다이어리 페이지를 향해 시선을 꽂은 채 눈을 번쩍 뜬 서의용의 얼굴\n\nLOCATION (lock): Beside the keepsake box and desks in a corner of the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly-in slightly below 서의용's eye line from the three-quarter side between his face and the diary, allowing his widened eyes to occupy most of the close frame while the open page remains along the lower edge of his sightline. His gaze is the sole emphasized change: it locks downward onto the discovered entry as his body pitches toward the page, never turning toward the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 다이어리 페이지 (Held open in 서의용's hand) — The open written face angles upward toward 서의용 and partly toward the camera, with the discovered date and '매직~' entry visible; used as Remains at the lower frame edge as the visible source of 서의용's recognition.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to the nighttime office isolates the recognition while preserving natural facial tone and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same cluttered office desk, records, fluorescent lighting, and the detective's clothing. Exclude the laptop-moving action and isolate him studying the diary.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's keepsake box remains in the office corner, and Euiyong has taken her diary from it and is examining the pages.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다이어리 페이지: \"10/24 (화) 23:30 삼거리\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "서의용의 시선이 다이어리 페이지를 향하고 있음.",
    "built_space": "사무실 배경으로 책상과 블라인드가 보임.",
    "entities": "서의용과 다이어리가 존재하며 텍스트가 일치함. 그러나 프롬프트에 없는 정장 차림 인물의 손과 팔이 나타남.",
    "hard_violations": [
     "명시되지 않은 인물(정장 차림의 팔과 손) 등장"
    ],
    "physics": "서의용과 다른 인물의 손이 다이어리를 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "서의용의 시선이 다이어리에 꽂혀 있음.",
    "built_space": "사무실 내 책상, 모니터, 창문이 보임.",
    "entities": "서의용과 다이어리가 존재하나, 프롬프트에 금지된 정장 차림의 인물 뒷모습이 좌측에 나타남.",
    "hard_violations": [
     "명시되지 않은 인물(정장 입은 사람의 어깨와 뒷모습) 등장"
    ],
    "physics": "서의용이 다이어리를 손으로 받쳐 들고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 3,
   "B": 2
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "지정되지 않은 인물(정장 차림의 손과 팔)이 등장하여 프롬프트의 인물 제한 조건을 심각하게 위반함."
   },
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "명시되지 않은 인물(정장 입은 뒷모습)이 화면을 차지하여 인물 묘사 지시를 완전히 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S49sh10_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "PROP REFERENCE — 김선영의 낡은 노트와 다이어리: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1240939>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 하단 전경에 서의용이 아닌, 이전 샷 인물의 복장(남색 정장과 하늘색 셔츠 소매)을 입은 다른 사람의 두 손이 다이어리를 잡고 있습니다.",
     "fix_en": "Remove the hands wearing the blue suit and light blue shirt at the bottom foreground, replacing them with the lower portion of the diary's brown leather cover and the out-of-focus desk surface. Preserve Seo Eui-yong's face, his own hand, his black leather jacket, the open diary page, and the background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "서의용이 들고 있는 다이어리가 레퍼런스 이미지의 작고 얇은 낡은 가죽 수첩이 아닌, 매우 두껍고 큰 책으로 잘못 묘사되었습니다.",
     "fix_en": "Slim down the open diary from a thick, large book to a small, thin, worn pocket-sized leather notebook, reducing the bulk of the page blocks. Preserve Seo Eui-yong's pose, facial expression, his leather jacket, the text on the page, and the office background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "다이어리 페이지에 적힌 텍스트가 종이의 질감이나 표면 굴곡을 자연스럽게 따르지 않고, 화면 위에 평면적인 텍스트를 오버레이하여 합성한 것처럼 보입니다.",
     "fix_en": "Blend the text into the paper's surface so it follows the page curvature and texture like actual handwritten ink, removing the flat digital sticker effect. Preserve the exact text content, the page shape, Seo Eui-yong's face, and the lighting.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "서의용 얼굴이 렌즈 쪽으로 거의 정면을 향하고 시선이 페이지가 아니라 카메라 근처를 본다.",
     "fix_en": "Redirect Seo Eui-yong's eyes to look sharply downward, locking onto the diary page at the bottom frame instead of toward the camera lens. Preserve his head position, facial features, black leather jacket, the diary, and the office background.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "카메라가 눈높이 약간 아래 얼굴과 다이어리 사이 3/4 측면이 아니라 거의 정면 클로즈업이다.",
     "fix_en": "Deepen the shadows on the side of the face to create an illusion of a slightly more angled perspective, as the camera cannot be relocated in place. Preserve Seo Eui-yong's identity, the diary, and the background elements.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 하단 전경에 서의용이 아닌, 이전 샷 인물의 복장(남색 정장과 하늘색 셔츠 소매)을 입은 다른 사람의 두 손이 다이어리를 잡고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용이 들고 있는 다이어리가 레퍼런스 이미지의 작고 얇은 낡은 가죽 수첩이 아닌, 매우 두껍고 큰 책으로 잘못 묘사되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "다이어리 페이지에 적힌 텍스트가 종이의 질감이나 표면 굴곡을 자연스럽게 따르지 않고, 화면 위에 평면적인 텍스트를 오버레이하여 합성한 것처럼 보입니다.",
     "severity": "major"
    },
    {
     "issue_ko": "하단 왼쪽에 서의용 옷이 아닌 남색 정장·흰 셔츠 소매의 다른 손이 다이어리를 잡고 있어 샷에 없는 인물의 신체 일부가 보인다.",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용 얼굴이 렌즈 쪽으로 거의 정면을 향하고 시선이 페이지가 아니라 카메라 근처를 본다.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 눈높이 약간 아래 얼굴과 다이어리 사이 3/4 측면이 아니라 거의 정면 클로즈업이다.",
     "severity": "major"
    },
    {
     "issue_ko": "다이어리의 ‘10/24 (화) 23:30 삼거리’ 글자가 종이에 앉은 손글씨가 아니라 평면 그래픽처럼 붙어 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "다이어리 페이지에 적힌 텍스트가 종이의 질감이나 표면 굴곡을 자연스럽게 따르지 않고, 화면 위에 평면적인 텍스트를 오버레이하여 합성한 것처럼 보입니다.",
    "fix_en": "Blend the text into the paper's surface so it follows the page curvature and texture like actual handwritten ink, removing the flat digital sticker effect. Preserve the exact text content, the page shape, Seo Eui-yong's face, and the lighting.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "서의용 얼굴이 렌즈 쪽으로 거의 정면을 향하고 시선이 페이지가 아니라 카메라 근처를 본다.",
    "fix_en": "Redirect Seo Eui-yong's eyes to look sharply downward, locking onto the diary page at the bottom frame instead of toward the camera lens. Preserve his head position, facial features, black leather jacket, the diary, and the office background.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "카메라가 눈높이 약간 아래 얼굴과 다이어리 사이 3/4 측면이 아니라 거의 정면 클로즈업이다.",
    "fix_en": "Deepen the shadows on the side of the face to create an illusion of a slightly more angled perspective, as the camera cannot be relocated in place. Preserve Seo Eui-yong's identity, the diary, and the background elements.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the hands wearing the blue suit and light blue shirt at the bottom foreground, replacing them with the lower portion of the diary's brown leather cover and the out-of-focus desk surface. Preserve Seo Eui-yong's face, his own hand, his black leather jacket, the open diary page, and the background.\n- Slim down the open diary from a thick, large book to a small, thin, worn pocket-sized leather notebook, reducing the bulk of the page blocks. Preserve Seo Eui-yong's pose, facial expression, his leather jacket, the text on the page, and the office background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 583,
      "verdict_ko": "신원은 일치하나, 요구된 클로즈업 샷을 무시하고 다이어리를 들고 있지 않으며 이전 샷의 인물이 화면 우측에 그대로 등장하여 재촬영이 필요함.  ★위반: [gemini-pro] 프롬프트에 없는 인물 등장 (화면 우측의 흰 셔츠 입은 인물) / [gemini-pro] 프레이밍 및 연출 위반 (클로즈업이 아니며 다이어리를 손에 쥐고 있지 않음) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 보이지 말라고 한 이전 스틸의 우측 인물이 남아 발명된 인물이다"
     },
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "클로즈업과 텍스트는 반영했으나, 레퍼런스와 전혀 다른 얼굴로 신원이 심각하게 훼손되었고 화면에 없는 인물의 손이 다이어리를 쥐고 있어 실격됨.  ★위반: [gemini-pro] 캐릭터 신원 불일치 (레퍼런스와 완전히 다른 얼굴) / [gemini-pro] 프롬프트에 없는 인물 등장 (다이어리를 쥐고 있는 정장 소매의 손) / [openrouter:x-ai/grok-4.6] 샷에 없는 정장 소매의 타인 손이 프레임에 들어와 추가 신체/발명된 인물이 된다"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.0,
      "B": 0.583
     },
     "violations": {
      "B": [
       "[gemini-pro] 프롬프트에 없는 인물 등장 (화면 우측의 흰 셔츠 입은 인물)",
       "[gemini-pro] 프레이밍 및 연출 위반 (클로즈업이 아니며 다이어리를 손에 쥐고 있지 않음)",
       "[openrouter:x-ai/grok-4.6] 샷 텍스트가 보이지 말라고 한 이전 스틸의 우측 인물이 남아 발명된 인물이다"
      ],
      "A": [
       "[gemini-pro] 캐릭터 신원 불일치 (레퍼런스와 완전히 다른 얼굴)",
       "[gemini-pro] 프롬프트에 없는 인물 등장 (다이어리를 쥐고 있는 정장 소매의 손)",
       "[openrouter:x-ai/grok-4.6] 샷에 없는 정장 소매의 타인 손이 프레임에 들어와 추가 신체/발명된 인물이 된다"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 583,
      "verdict_ko": "신원은 일치하나, 요구된 클로즈업 샷을 무시하고 다이어리를 들고 있지 않으며 이전 샷의 인물이 화면 우측에 그대로 등장하여 재촬영이 필요함.  ★위반: [gemini-pro] 프롬프트에 없는 인물 등장 (화면 우측의 흰 셔츠 입은 인물) / [gemini-pro] 프레이밍 및 연출 위반 (클로즈업이 아니며 다이어리를 손에 쥐고 있지 않음) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 보이지 말라고 한 이전 스틸의 우측 인물이 남아 발명된 인물이다"
     },
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "클로즈업과 텍스트는 반영했으나, 레퍼런스와 전혀 다른 얼굴로 신원이 심각하게 훼손되었고 화면에 없는 인물의 손이 다이어리를 쥐고 있어 실격됨.  ★위반: [gemini-pro] 캐릭터 신원 불일치 (레퍼런스와 완전히 다른 얼굴) / [gemini-pro] 프롬프트에 없는 인물 등장 (다이어리를 쥐고 있는 정장 소매의 손) / [openrouter:x-ai/grok-4.6] 샷에 없는 정장 소매의 타인 손이 프레임에 들어와 추가 신체/발명된 인물이 된다"
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "요구된 클로즈업 샷, 눈을 번쩍 뜬 표정, 다이어리의 지정 텍스트를 정확히 구현했으나, 좌측 하단에 금지된 제3자의 팔(정장 소매)이 등장해 감점됨."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트가 지시한 구도와 표정을 무시했으며, 다이어리의 영문 텍스트, 우측의 추가 인물, 물리적으로 불가능한 사물 배치 등 다수의 치명적 오류가 있음."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 아래를 향하지만 눈을 크게 뜨지 않음.",
      "built_space": "사무실 책상과 뒤쪽의 모니터가 보임.",
      "entities": "서의용의 의상과 얼굴은 일치함. 다이어리에 지시되지 않은 영문이 적혀 있으며, 우측에 금지된 타인이 노출됨.",
      "hard_violations": [
       "허가되지 않은 추가 인물 등장(우측)",
       "지지하는 손 없이 수직으로 서 있는 다이어리"
      ],
      "physics": "펼쳐진 다이어리가 손에 들려있지 않고 아무런 지지대 없이 책상 위에 비현실적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "눈을 번쩍 뜨고 다이어리 페이지를 강렬하게 응시함.",
      "built_space": "아웃포커싱된 사무실 배경.",
      "entities": "서의용의 인상과 의상이 일치하며, 다이어리에 요구된 한글 텍스트가 정확히 적혀 있음. 좌측 하단에 정장 소매를 입은 타인의 팔이 보임.",
      "hard_violations": [
       "프롬프트에 없는 제3자의 팔(정장 소매) 등장"
      ],
      "physics": "다이어리가 손에 들려 있으나 서의용의 손과 함께 제3자의 손이 잡고 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "요구된 클로즈업 샷, 눈을 번쩍 뜬 표정, 다이어리의 지정 텍스트를 정확히 구현했으나, 좌측 하단에 금지된 제3자의 팔(정장 소매)이 등장해 감점됨."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 지시한 구도와 표정을 무시했으며, 다이어리의 영문 텍스트, 우측의 추가 인물, 물리적으로 불가능한 사물 배치 등 다수의 치명적 오류가 있음."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 아래를 향하지만 눈을 크게 뜨지 않음.",
      "built_space": "사무실 책상과 뒤쪽의 모니터가 보임.",
      "entities": "서의용의 의상과 얼굴은 일치함. 다이어리에 지시되지 않은 영문이 적혀 있으며, 우측에 금지된 타인이 노출됨.",
      "hard_violations": [
       "허가되지 않은 추가 인물 등장(우측)",
       "지지하는 손 없이 수직으로 서 있는 다이어리"
      ],
      "physics": "펼쳐진 다이어리가 손에 들려있지 않고 아무런 지지대 없이 책상 위에 비현실적으로 서 있음."
     },
     {
      "label": "A",
      "direction": "눈을 번쩍 뜨고 다이어리 페이지를 강렬하게 응시함.",
      "built_space": "아웃포커싱된 사무실 배경.",
      "entities": "서의용의 인상과 의상이 일치하며, 다이어리에 요구된 한글 텍스트가 정확히 적혀 있음. 좌측 하단에 정장 소매를 입은 타인의 팔이 보임.",
      "hard_violations": [
       "프롬프트에 없는 제3자의 팔(정장 소매) 등장"
      ],
      "physics": "다이어리가 손에 들려 있으나 서의용의 손과 함께 제3자의 손이 잡고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1005,
     "B": 586
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S49sh10"
  }
 },
 "S49sh14::cine": {
  "applied": true,
  "fingerprint": "b74fbb798f764c10fb5daa773230770fdfad32a4e6cd839a73b0e94f6f1ff736",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S49sh14_sel.png",
  "source_sha256": "d4f82326acdab9c3221cb24683c61dbfea68f64c67fedcd5974c6c57cbb8c255",
  "file": "S49sh14_cine.png",
  "latency_ms": 10438
 },
 "S49sh16::signage": {
  "fp": "c2ba1b976ac28398",
  "inscriptions": [
   {
    "surface_native": "다이어리 내지",
    "text_native": "약속 장소: 영등포 물류창고 22:00",
    "reason_ko": "형사들이 수사 중 발견한 핵심 단서나 메모를 보여주는 다이어리 페이지의 글귀가 필요합니다."
   }
  ]
 },
 "S49sh16": {
  "input_fingerprint": "dc5a0559fdee9d56",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 특정 글귀가 적힌 다이어리 페이지를 나상혁 쪽으로 뻗은 채 쥐고 있는 서의용의 팔\n\nLOCATION (lock): At the shared workstation inside the violent-crimes office, over the opened diary. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From upper-torso height beside the handoff line, track close along 서의용's extended forearm as the open diary runs diagonally from the lower left toward 나상혁 at the upper right. Use an insert-like composition in which the marked page and gripping hand remain sharp while both men's faces are cropped beyond the frame edges; 나상혁's near hand and torso edge provide scale without interrupting the handoff direction.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: open diary page (Open to the discovered entry) — The written face is angled toward the camera, with the marked entry legible on the exposed page; used as The marked page forms the sharp diagonal focal plane between the two investigators; desk edge (Supporting the handoff area); used as A narrow edge beneath the hands preserves the office context and physical support for the diary.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime office illumination with a neutral-to-cool procedural tone and enough restrained contrast to separate the writing from the hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office workstation, stacked case files, fluorescent light, and both investigators' clothing. Exclude the autopsy photographs from emphasis and focus on the diary page being extended.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong continues holding out the open diary at the page marked “Magic,” with the keepsake box remaining nearby.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다이어리 내지: \"약속 장소: 영등포 물류창고 22:00\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 특정 글귀가 적힌 다이어리 페이지를 나상혁 쪽으로 뻗은 채 쥐고 있는 서의용의 팔\n\nLOCATION (lock): At the shared workstation inside the violent-crimes office, over the opened diary. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From upper-torso height beside the handoff line, track close along 서의용's extended forearm as the open diary runs diagonally from the lower left toward 나상혁 at the upper right. Use an insert-like composition in which the marked page and gripping hand remain sharp while both men's faces are cropped beyond the frame edges; 나상혁's near hand and torso edge provide scale without interrupting the handoff direction.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: open diary page (Open to the discovered entry) — The written face is angled toward the camera, with the marked entry legible on the exposed page; used as The marked page forms the sharp diagonal focal plane between the two investigators; desk edge (Supporting the handoff area); used as A narrow edge beneath the hands preserves the office context and physical support for the diary.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime office illumination with a neutral-to-cool procedural tone and enough restrained contrast to separate the writing from the hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office workstation, stacked case files, fluorescent light, and both investigators' clothing. Exclude the autopsy photographs from emphasis and focus on the diary page being extended.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong continues holding out the open diary at the page marked “Magic,” with the keepsake box remaining nearby.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다이어리 내지: \"약속 장소: 영등포 물류창고 22:00\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 특정 글귀가 적힌 다이어리 페이지를 나상혁 쪽으로 뻗은 채 쥐고 있는 서의용의 팔\n\nLOCATION (lock): At the shared workstation inside the violent-crimes office, over the opened diary. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From upper-torso height beside the handoff line, track close along 서의용's extended forearm as the open diary runs diagonally from the lower left toward 나상혁 at the upper right. Use an insert-like composition in which the marked page and gripping hand remain sharp while both men's faces are cropped beyond the frame edges; 나상혁's near hand and torso edge provide scale without interrupting the handoff direction.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: open diary page (Open to the discovered entry) — The written face is angled toward the camera, with the marked entry legible on the exposed page; used as The marked page forms the sharp diagonal focal plane between the two investigators; desk edge (Supporting the handoff area); used as A narrow edge beneath the hands preserves the office context and physical support for the diary.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime office illumination with a neutral-to-cool procedural tone and enough restrained contrast to separate the writing from the hands.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office workstation, stacked case files, fluorescent light, and both investigators' clothing. Exclude the autopsy photographs from emphasis and focus on the diary page being extended.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong continues holding out the open diary at the page marked “Magic,” with the keepsake box remaining nearby.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다이어리 내지: \"약속 장소: 영등포 물류창고 22:00\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "지정된 인서트 구도와 인물들의 의상(가죽 재킷과 후드, 수트)을 잘 구현했으며, 명시된 보관함 소품과 텍스트가 정확히 반영되었습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "구도와 다이어리 질감은 양호하나, 프롬프트에서 요구한 보관함(keepsake box)이 누락되었고 숫자 텍스트가 다소 일그러졌습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S49sh14_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "PROP REFERENCE — 김선영의 낡은 노트와 다이어리: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1240939>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "다이어리 페이지에 적힌 텍스트가 종이의 기울기와 곡면을 따르지 않고, 프레임에 수평으로 평평하게 붙여넣은 2D 그래픽 오버레이처럼 보입니다.",
     "fix_en": "Transform the flat text overlay on the diary page to follow the paper's tilt and curvature with realistic ink shading. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "인서트 클로즈업이 아니라 두 사람 상반신과 사무실·책상이 넓게 보이는 미디엄에 가깝다.",
     "fix_en": "Would require reframing as a tighter insert close-up on the diary and hands.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "펼친 다이어리가 좌하단에서 우상단 나상혁 쪽으로 대각선을 이루지 않고 화면 중앙에 정면으로 들려 있다.",
     "fix_en": "Would require reorienting the diary and arms diagonally from lower left to upper right.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "서의용이 나상혁 쪽으로 뻗어 쥐는 것이 아니라 오른쪽 정장 손도 페이지를 붙잡아 건네는 주체가 불명확하다.",
     "fix_en": "Would require removing the right hand on the top page edge to clarify who is handing the diary.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "다이어리 페이지에 적힌 텍스트가 종이의 기울기와 곡면을 따르지 않고, 프레임에 수평으로 평평하게 붙여넣은 2D 그래픽 오버레이처럼 보입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "인서트 클로즈업이 아니라 두 사람 상반신과 사무실·책상이 넓게 보이는 미디엄에 가깝다.",
     "severity": "major"
    },
    {
     "issue_ko": "펼친 다이어리가 좌하단에서 우상단 나상혁 쪽으로 대각선을 이루지 않고 화면 중앙에 정면으로 들려 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "서의용이 나상혁 쪽으로 뻗어 쥐는 것이 아니라 오른쪽 정장 손도 페이지를 붙잡아 건네는 주체가 불명확하다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "인서트 클로즈업이 아니라 두 사람 상반신과 사무실·책상이 넓게 보이는 미디엄에 가깝다.",
    "fix_en": "Would require reframing as a tighter insert close-up on the diary and hands.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "펼친 다이어리가 좌하단에서 우상단 나상혁 쪽으로 대각선을 이루지 않고 화면 중앙에 정면으로 들려 있다.",
    "fix_en": "Would require reorienting the diary and arms diagonally from lower left to upper right.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "서의용이 나상혁 쪽으로 뻗어 쥐는 것이 아니라 오른쪽 정장 손도 페이지를 붙잡아 건네는 주체가 불명확하다.",
    "fix_en": "Would require removing the right hand on the top page edge to clarify who is handing the diary.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Transform the flat text overlay on the diary page to follow the paper's tilt and curvature with realistic ink shading. Preserve the people present and their positions, their clothing, the set, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프레이밍(인서트 클로즈업, 얼굴 크롭), 지정된 텍스트, 의상 및 소품(다이어리, 보관함) 지시를 모두 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "카메라 프레이밍 지시(얼굴 크롭)를 위반하여 인물의 얼굴을 그대로 노출했고, 다이어리의 필수 텍스트도 누락되어 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽의 가죽 재킷을 입은 팔이 다이어리를 오른쪽 정장 입은 사람 쪽으로 뻗고 있음.",
      "built_space": "사무실 책상 위, 배경에 서류 더미와 천장 형광등이 보임.",
      "entities": "가죽 재킷 소매(서의용), 정장 소매(나상혁), 지정된 문구가 정확히 적힌 다이어리, 나무 보관함.",
      "hard_violations": [],
      "physics": "양측의 손이 다이어리의 양 끝을 안정적으로 쥐어 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 펼쳐진 다이어리를 내려다보고 있음.",
      "built_space": "사무실 책상, 배경에 흐릿한 사무 공간이 배치됨.",
      "entities": "서의용의 전체 얼굴과 상반신, 글씨가 없는 빈 다이어리, 앞쪽 인물의 소매 일부.",
      "hard_violations": [
       "카메라 프레이밍 위반 (얼굴이 프레임 밖으로 잘려야 하나 전체가 노출됨)",
       "필수 텍스트 누락 (다이어리 페이지가 비어 있음)"
      ],
      "physics": "인물의 손이 다이어리를 쥐고 책상 위에 기대어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프레이밍(인서트 클로즈업, 얼굴 크롭), 지정된 텍스트, 의상 및 소품(다이어리, 보관함) 지시를 모두 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "카메라 프레이밍 지시(얼굴 크롭)를 위반하여 인물의 얼굴을 그대로 노출했고, 다이어리의 필수 텍스트도 누락되어 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽의 가죽 재킷을 입은 팔이 다이어리를 오른쪽 정장 입은 사람 쪽으로 뻗고 있음.",
      "built_space": "사무실 책상 위, 배경에 서류 더미와 천장 형광등이 보임.",
      "entities": "가죽 재킷 소매(서의용), 정장 소매(나상혁), 지정된 문구가 정확히 적힌 다이어리, 나무 보관함.",
      "hard_violations": [],
      "physics": "양측의 손이 다이어리의 양 끝을 안정적으로 쥐어 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 펼쳐진 다이어리를 내려다보고 있음.",
      "built_space": "사무실 책상, 배경에 흐릿한 사무 공간이 배치됨.",
      "entities": "서의용의 전체 얼굴과 상반신, 글씨가 없는 빈 다이어리, 앞쪽 인물의 소매 일부.",
      "hard_violations": [
       "카메라 프레이밍 위반 (얼굴이 프레임 밖으로 잘려야 하나 전체가 노출됨)",
       "필수 텍스트 누락 (다이어리 페이지가 비어 있음)"
      ],
      "physics": "인물의 손이 다이어리를 쥐고 책상 위에 기대어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 인서트 샷 프레이밍을 정확히 준수하여 얼굴을 자르고 손과 다이어리에 집중했으며, 지정된 텍스트와 팔의 방향까지 프롬프트를 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "얼굴이 잘려야 한다는 카메라 프레이밍 지시를 위반하고 이전 샷의 구도를 그대로 사용했으며, 다이어리의 텍스트도 누락되었습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "다이어리가 좌측 하단에서 우측 상단의 상대방을 향해 대각선으로 내밀어짐.",
      "built_space": "사무실 책상 위, 서류 더미와 보관함 상자가 배치됨.",
      "entities": "가죽 재킷을 입은 서의용의 팔, 정장을 입은 나상혁의 손, 정확한 문구('약속 장소: 영등포 물류창고 22:00')가 적힌 다이어리 일치함.",
      "hard_violations": [],
      "physics": "양측 인물의 손이 다이어리를 물리적으로 자연스럽게 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "인물의 시선이 아래쪽 다이어리를 향함.",
      "built_space": "사무실 공간, 배경에 책상과 형광등이 보임.",
      "entities": "서의용의 얼굴과 상체가 보이나, 다이어리 내지는 텍스트 없이 비어 있음.",
      "hard_violations": [
       "두 인물의 얼굴을 프레임 밖으로 자르라는 카메라 지시(insert close-up)를 완전히 무시하고 서의용의 얼굴을 노출함."
      ],
      "physics": "손이 책상 위에서 다이어리를 받치고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 인서트 샷 프레이밍을 정확히 준수하여 얼굴을 자르고 손과 다이어리에 집중했으며, 지정된 텍스트와 팔의 방향까지 프롬프트를 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "얼굴이 잘려야 한다는 카메라 프레이밍 지시를 위반하고 이전 샷의 구도를 그대로 사용했으며, 다이어리의 텍스트도 누락되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "다이어리가 좌측 하단에서 우측 상단의 상대방을 향해 대각선으로 내밀어짐.",
      "built_space": "사무실 책상 위, 서류 더미와 보관함 상자가 배치됨.",
      "entities": "가죽 재킷을 입은 서의용의 팔, 정장을 입은 나상혁의 손, 정확한 문구('약속 장소: 영등포 물류창고 22:00')가 적힌 다이어리 일치함.",
      "hard_violations": [],
      "physics": "양측 인물의 손이 다이어리를 물리적으로 자연스럽게 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 아래쪽 다이어리를 향함.",
      "built_space": "사무실 공간, 배경에 책상과 형광등이 보임.",
      "entities": "서의용의 얼굴과 상체가 보이나, 다이어리 내지는 텍스트 없이 비어 있음.",
      "hard_violations": [
       "두 인물의 얼굴을 프레임 밖으로 자르라는 카메라 지시(insert close-up)를 완전히 무시하고 서의용의 얼굴을 노출함."
      ],
      "physics": "손이 책상 위에서 다이어리를 받치고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S49sh14"
  }
 },
 "S49sh16::cine": {
  "applied": true,
  "fingerprint": "505565334b67464840b9345d503c03cc50915bc77b371e300535a1180e95e818",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S49sh16_sel.png",
  "source_sha256": "f47387ce87da97723199229f48c8c24130964d12970c94159368d3b617eb1084",
  "file": "S49sh16_cine.png",
  "latency_ms": 11926
 },
 "S50sh1::signage": {
  "fp": "b5fecf9c615b060a",
  "inscriptions": [
   {
    "surface_native": "노트북 화면의 문서 헤더",
    "text_native": "부검감정서",
    "reason_ko": "형사들이 노트북으로 부검 사진과 문서를 함께 검토하는 수사 상황을 현실적으로 묘사하기 위해 화면에 명확한 한글 표기가 필요합니다."
   }
  ]
 },
 "S50sh1": {
  "input_fingerprint": "18953be5a4917399",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 꼬리뼈 부검 사진이 띄워진 노트북 화면을 중심으로 둥글게 모여앉은 전택수, 강주철, 서의용, 나상혁의 전경\n\nLOCATION (lock): Around a laptop in the meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin at a high oblique angle just outside the seated circle, holding the laptop at the center while 전택수, 강주철, 서의용, and 나상혁 occupy distinct arcs around it. The crane is caught at its initial static beat: each investigator leans toward the displayed autopsy image at a slightly different angle, avoiding a mechanically symmetrical arrangement.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: laptop screen (On and displaying the autopsy photograph) — The display face is visible to the camera and shows the autopsy photograph focused on the coccyx area with a small amount of pooled blood; used as Central evidence anchor around which the four seated figures form an uneven circle; chairs around the laptop (Occupied by the four investigators) — The chair fronts turn inward toward the laptop from different directions; used as Defines the seated circular arrangement and keeps each investigator at an individualized angle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination rendered in a subdued neutral-to-cool palette with moderate, non-sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa arrangement, daytime lighting, and institutional decor. Exclude the earlier empty floor focus and add the four investigators gathered around the laptop.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop remains open at the center of the group with the tailbone autopsy photograph displayed on its screen.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 노트북 화면의 문서 헤더: \"부검감정서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 꼬리뼈 부검 사진이 띄워진 노트북 화면을 중심으로 둥글게 모여앉은 전택수, 강주철, 서의용, 나상혁의 전경\n\nLOCATION (lock): Around a laptop in the meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin at a high oblique angle just outside the seated circle, holding the laptop at the center while 전택수, 강주철, 서의용, and 나상혁 occupy distinct arcs around it. The crane is caught at its initial static beat: each investigator leans toward the displayed autopsy image at a slightly different angle, avoiding a mechanically symmetrical arrangement.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: laptop screen (On and displaying the autopsy photograph) — The display face is visible to the camera and shows the autopsy photograph focused on the coccyx area with a small amount of pooled blood; used as Central evidence anchor around which the four seated figures form an uneven circle; chairs around the laptop (Occupied by the four investigators) — The chair fronts turn inward toward the laptop from different directions; used as Defines the seated circular arrangement and keeps each investigator at an individualized angle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination rendered in a subdued neutral-to-cool palette with moderate, non-sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa arrangement, daytime lighting, and institutional decor. Exclude the earlier empty floor focus and add the four investigators gathered around the laptop.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop remains open at the center of the group with the tailbone autopsy photograph displayed on its screen.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 노트북 화면의 문서 헤더: \"부검감정서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 꼬리뼈 부검 사진이 띄워진 노트북 화면을 중심으로 둥글게 모여앉은 전택수, 강주철, 서의용, 나상혁의 전경\n\nLOCATION (lock): Around a laptop in the meeting area of the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin at a high oblique angle just outside the seated circle, holding the laptop at the center while 전택수, 강주철, 서의용, and 나상혁 occupy distinct arcs around it. The crane is caught at its initial static beat: each investigator leans toward the displayed autopsy image at a slightly different angle, avoiding a mechanically symmetrical arrangement.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: laptop screen (On and displaying the autopsy photograph) — The display face is visible to the camera and shows the autopsy photograph focused on the coccyx area with a small amount of pooled blood; used as Central evidence anchor around which the four seated figures form an uneven circle; chairs around the laptop (Occupied by the four investigators) — The chair fronts turn inward toward the laptop from different directions; used as Defines the seated circular arrangement and keeps each investigator at an individualized angle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination rendered in a subdued neutral-to-cool palette with moderate, non-sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa arrangement, daytime lighting, and institutional decor. Exclude the earlier empty floor focus and add the four investigators gathered around the laptop.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop remains open at the center of the group with the tailbone autopsy photograph displayed on its screen.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 노트북 화면의 문서 헤더: \"부검감정서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.667,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.25,
    "B": 0.583
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 인물(전택수) 누락",
     "[gemini-pro] 창조된 인물(베이지색 니트 남성) 등장",
     "[openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 베이지 스웨터 남성을 추가하고 전택수를 빠뜨림"
    ],
    "A": [
     "[gemini-pro] 지정된 인물(전택수) 누락",
     "[gemini-pro] 금지된 인물(이전 샷의 정장 남성) 재사용 및 창조"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 583,
   "A": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 583,
    "verdict_ko": "화면의 텍스트와 꼬리뼈 형태는 지침에 가깝게 구현했으나, 필수 인물인 전택수를 누락하고 임의의 인물을 생성하여 치명적인 오류가 발생함.  ★위반: [gemini-pro] 지정된 인물(전택수) 누락 / [gemini-pro] 창조된 인물(베이지색 니트 남성) 등장 / [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 베이지 스웨터 남성을 추가하고 전택수를 빠뜨림"
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "전택수를 배제하고 이전 샷에서 금지된 정장 인물을 잘못 포함시켰으며, 꼬리뼈가 아닌 팔 부검 사진을 띄워 주요 지침을 모두 위반함.  ★위반: [gemini-pro] 지정된 인물(전택수) 누락 / [gemini-pro] 금지된 인물(이전 샷의 정장 남성) 재사용 및 창조"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S46sh8_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프레임 전경에 등장해야 할 전택수(경찰 점퍼와 모자 착용)가 누락되고, 이전 컷 스틸에서 제외하도록 지시된 인물(회색 정장)이 잘못 포함됨.",
     "fix_en": "Replace the man in the dark grey suit in the foreground with the character 전택수. He must wear a navy blue SAURS police jacket over a plaid shirt and a navy blue cap with gold Korean lettering, matching his reference exactly. Preserve the other three seated men, the table, the laptop, and the background room.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "노트북 화면의 부검 사진이 지시된 꼬리뼈 부위가 아니라 피가 묻은 팔 부위를 보여줌.",
     "fix_en": "Change the image displayed on the laptop screen to an autopsy photograph of a human tailbone (coccyx) area with a small amount of pooled blood, replacing the bloody arm. Preserve the '부검감정서' text above the image, the laptop itself, the table, and the four investigators.",
     "severity": "critical",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프레임 전경에 등장해야 할 전택수(경찰 점퍼와 모자 착용)가 누락되고, 이전 컷 스틸에서 제외하도록 지시된 인물(회색 정장)이 잘못 포함됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "노트북 화면의 부검 사진이 지시된 꼬리뼈 부위가 아니라 피가 묻은 팔 부위를 보여줌.",
     "severity": "critical"
    },
    {
     "issue_ko": "노트북 화면의 부검 사진이 꼬리뼈가 아니라 피가 고인 팔·다리로 보인다",
     "severity": "critical"
    },
    {
     "issue_ko": "등이 보이는 인물이 전택수의 잠긴 수사 모자와 남색 경찰 점퍼가 아닌 어두운 정장 차림이다",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 6,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the man in the dark grey suit in the foreground with the character 전택수. He must wear a navy blue SAURS police jacket over a plaid shirt and a navy blue cap with gold Korean lettering, matching his reference exactly. Preserve the other three seated men, the table, the laptop, and the background room.\n- Change the image displayed on the laptop screen to an autopsy photograph of a human tailbone (coccyx) area with a small amount of pooled blood, replacing the bloody arm. Preserve the '부검감정서' text above the image, the laptop itself, the table, and the four investigators.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "3명의 캐릭터 외형과 배경, 원형 착석 구도를 잘 구현했으나, 핵심 인물인 전택수를 누락하고 제외 대상인 이전 컷의 인물(회색 정장)을 포함하여 실격입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "배경이 불일치하며 3명의 캐릭터 디자인이 레퍼런스와 전혀 다르고 전택수가 서 있는 등 인물과 구도에서 심각한 오류가 다수 존재합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "4명이 중앙의 노트북 화면(피 묻은 팔 사진과 '부검감정서' 텍스트)을 향해 시선을 고정함.",
      "built_space": "이전 컷과 일치하는 가죽 소파와 우드톤 사무실 배경이며, 둥근 테이블 주변에 4개의 의자가 적절히 배치됨.",
      "entities": "서의용, 강주철, 나상혁은 레퍼런스와 정확히 일치함. 노트북 화면 텍스트 일치. 그러나 전택수 대신 이전 컷에서 제외해야 할 회색 정장 인물이 뒷모습으로 등장함.",
      "hard_violations": [
       "필수 인물(전택수) 누락",
       "금지된 인물(이전 컷의 회색 정장 남성) 포함"
      ],
      "physics": "4명 모두 의자에 안정적으로 앉아 있으며, 사물들도 테이블 위에 잘 놓여 있음."
     },
     {
      "label": "B",
      "direction": "1명이 서고 3명이 앉아 노트북 화면(관절뼈 사진과 '부검감정서' 텍스트)을 응시함.",
      "built_space": "화이트보드와 철제 캐비닛이 있어 지시된 이전 컷의 장소(가죽 소파 등)와 불일치함.",
      "entities": "서있는 인물은 전택수의 의상을 흉내냈으나 모자 문구가 틀림('수사'가 아닌 '부검감정'). 나머지 3명은 지정된 캐릭터 레퍼런스와 의상 및 외모가 완전히 다름.",
      "hard_violations": [
       "구도 위반 (전원 착석이 아닌 1명 기립 상태)",
       "3명의 캐릭터(서의용, 강주철, 나상혁) 레퍼런스 완전 불일치 및 발명된 인물 사용"
      ],
      "physics": "서 있거나 앉아 있는 인물 모두 바닥과 의자에 의해 지지되며, 사물들도 안정적임."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "3명의 캐릭터 외형과 배경, 원형 착석 구도를 잘 구현했으나, 핵심 인물인 전택수를 누락하고 제외 대상인 이전 컷의 인물(회색 정장)을 포함하여 실격입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "배경이 불일치하며 3명의 캐릭터 디자인이 레퍼런스와 전혀 다르고 전택수가 서 있는 등 인물과 구도에서 심각한 오류가 다수 존재합니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "4명이 중앙의 노트북 화면(피 묻은 팔 사진과 '부검감정서' 텍스트)을 향해 시선을 고정함.",
      "built_space": "이전 컷과 일치하는 가죽 소파와 우드톤 사무실 배경이며, 둥근 테이블 주변에 4개의 의자가 적절히 배치됨.",
      "entities": "서의용, 강주철, 나상혁은 레퍼런스와 정확히 일치함. 노트북 화면 텍스트 일치. 그러나 전택수 대신 이전 컷에서 제외해야 할 회색 정장 인물이 뒷모습으로 등장함.",
      "hard_violations": [
       "필수 인물(전택수) 누락",
       "금지된 인물(이전 컷의 회색 정장 남성) 포함"
      ],
      "physics": "4명 모두 의자에 안정적으로 앉아 있으며, 사물들도 테이블 위에 잘 놓여 있음."
     },
     {
      "label": "B",
      "direction": "1명이 서고 3명이 앉아 노트북 화면(관절뼈 사진과 '부검감정서' 텍스트)을 응시함.",
      "built_space": "화이트보드와 철제 캐비닛이 있어 지시된 이전 컷의 장소(가죽 소파 등)와 불일치함.",
      "entities": "서있는 인물은 전택수의 의상을 흉내냈으나 모자 문구가 틀림('수사'가 아닌 '부검감정'). 나머지 3명은 지정된 캐릭터 레퍼런스와 의상 및 외모가 완전히 다름.",
      "hard_violations": [
       "구도 위반 (전원 착석이 아닌 1명 기립 상태)",
       "3명의 캐릭터(서의용, 강주철, 나상혁) 레퍼런스 완전 불일치 및 발명된 인물 사용"
      ],
      "physics": "서 있거나 앉아 있는 인물 모두 바닥과 의자에 의해 지지되며, 사물들도 안정적임."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "카메라 앵글, 지정된 장소, 그리고 3명의 캐릭터 신원 및 복장(서의용, 강주철, 나상혁)을 정확히 구현했으나, 전택수를 배제해야 할 정장 입은 인물로 대체한 점과 부검 사진이 꼬리뼈가 아닌 팔로 묘사된 점이 아쉽습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "장소 일관성이 없고 전택수가 서 있는 등 화면 구성이 잘못되었으며, 나머지 3명의 신원과 복장이 레퍼런스와 전혀 다른 인물로 창조되어 지시를 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "네 명의 인물 모두 중앙 테이블에 놓인 노트북 화면을 향해 시선을 고정하고 있음.",
      "built_space": "이전 숏과 일치하는 가죽 소파와 인테리어가 있는 사무실. 둥근 테이블을 중심으로 4개의 의자에 인물들이 착석함.",
      "entities": "서의용, 나상혁, 강주철은 레퍼런스의 얼굴과 복장을 정확히 반영함. 전택수 자리에 있는 인물은 이전 숏에서 배제 지시된 정장 입은 남성의 뒷모습임. 노트북 화면에는 '부검감정서' 텍스트와 피 묻은 팔 부위가 띄워져 있음.",
      "hard_violations": [
       "전택수의 고정된 외형을 무시하고 배제해야 할 정장 차림의 인물로 대체함(아이덴티티 훼손)",
       "부검 사진의 부위가 꼬리뼈가 아닌 팔로 묘사됨"
      ],
      "physics": "네 명 모두 의자에 안정적으로 앉아 체중을 지지하고 있으며, 노트북은 테이블 위에 정상적으로 놓여 있음. 물리적 오류 없음."
     },
     {
      "label": "A",
      "direction": "인물들의 시선이 중앙의 노트북 화면과 서 있는 전택수를 향해 교차하고 있음.",
      "built_space": "화이트보드와 비품이 있는 일반적인 직사각형 형태의 회의 공간. 3명은 착석하고 1명은 서 있음.",
      "entities": "전택수는 얼굴이 레퍼런스와 일치하나 모자에 '수사' 대신 '부검감정'이라는 잘못된 텍스트가 합성됨. 나머지 세 명은 제공된 캐릭터 레퍼런스와 전혀 다른 외모와 복장임. 화면에는 '부검감정서'와 뼈 관절 이미지가 있음.",
      "hard_violations": [
       "모두 앉아 있어야 한다는 지시를 어기고 전택수가 서 있음",
       "이전 숏에 고정된 장소(조사반장실 소파 구역)와 전혀 다른 공간을 배경으로 함",
       "서의용, 강주철, 나상혁의 신원과 복장이 레퍼런스와 전혀 일치하지 않음",
       "전택수의 모자에 화면 헤더 텍스트가 유출되어 변형됨"
      ],
      "physics": "앉은 3명은 의자에, 서 있는 1명은 바닥에 지지되어 있음. 물리적으로 불가능한 포즈나 부유하는 물체는 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "카메라 앵글, 지정된 장소, 그리고 3명의 캐릭터 신원 및 복장(서의용, 강주철, 나상혁)을 정확히 구현했으나, 전택수를 배제해야 할 정장 입은 인물로 대체한 점과 부검 사진이 꼬리뼈가 아닌 팔로 묘사된 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "장소 일관성이 없고 전택수가 서 있는 등 화면 구성이 잘못되었으며, 나머지 3명의 신원과 복장이 레퍼런스와 전혀 다른 인물로 창조되어 지시를 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "네 명의 인물 모두 중앙 테이블에 놓인 노트북 화면을 향해 시선을 고정하고 있음.",
      "built_space": "이전 숏과 일치하는 가죽 소파와 인테리어가 있는 사무실. 둥근 테이블을 중심으로 4개의 의자에 인물들이 착석함.",
      "entities": "서의용, 나상혁, 강주철은 레퍼런스의 얼굴과 복장을 정확히 반영함. 전택수 자리에 있는 인물은 이전 숏에서 배제 지시된 정장 입은 남성의 뒷모습임. 노트북 화면에는 '부검감정서' 텍스트와 피 묻은 팔 부위가 띄워져 있음.",
      "hard_violations": [
       "전택수의 고정된 외형을 무시하고 배제해야 할 정장 차림의 인물로 대체함(아이덴티티 훼손)",
       "부검 사진의 부위가 꼬리뼈가 아닌 팔로 묘사됨"
      ],
      "physics": "네 명 모두 의자에 안정적으로 앉아 체중을 지지하고 있으며, 노트북은 테이블 위에 정상적으로 놓여 있음. 물리적 오류 없음."
     },
     {
      "label": "B",
      "direction": "인물들의 시선이 중앙의 노트북 화면과 서 있는 전택수를 향해 교차하고 있음.",
      "built_space": "화이트보드와 비품이 있는 일반적인 직사각형 형태의 회의 공간. 3명은 착석하고 1명은 서 있음.",
      "entities": "전택수는 얼굴이 레퍼런스와 일치하나 모자에 '수사' 대신 '부검감정'이라는 잘못된 텍스트가 합성됨. 나머지 세 명은 제공된 캐릭터 레퍼런스와 전혀 다른 외모와 복장임. 화면에는 '부검감정서'와 뼈 관절 이미지가 있음.",
      "hard_violations": [
       "모두 앉아 있어야 한다는 지시를 어기고 전택수가 서 있음",
       "이전 숏에 고정된 장소(조사반장실 소파 구역)와 전혀 다른 공간을 배경으로 함",
       "서의용, 강주철, 나상혁의 신원과 복장이 레퍼런스와 전혀 일치하지 않음",
       "전택수의 모자에 화면 헤더 텍스트가 유출되어 변형됨"
      ],
      "physics": "앉은 3명은 의자에, 서 있는 1명은 바닥에 지지되어 있음. 물리적으로 불가능한 포즈나 부유하는 물체는 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S46sh8"
  }
 },
 "S50sh1::cine": {
  "applied": true,
  "fingerprint": "50dd95fe0a5f9a6ae5bf50fb1cdd410bf9530c642484645581b6fc3d41d289cd",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S50sh1_sel.png",
  "source_sha256": "2dc5fb90de57b13ca047200d545c04cddb43c00fc1e068d5765a071dd1a29f0f",
  "file": "S50sh1_cine.png",
  "latency_ms": 12015
 },
 "S50sh5::signage": {
  "fp": "83dcf998cd8c68ba",
  "inscriptions": []
 },
 "S50sh5": {
  "input_fingerprint": "5d004c9394f41052",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 본능적인 단서를 깨달은 듯 눈이 동그랗게 커진 채 굳어버린 강주철과 서의용의 모습\n\nLOCATION (lock): Among the seats surrounding the laptop inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height on the circle's outer edge, hold a tight oblique two-shot with 강주철 on the nearer left and 서의용 set slightly deeper on the right. Their eyes remain fixed toward the laptop below the frame line as their forward motions halt at different instants—강주철 caught mid-lean and 서의용 with one hand suspended—making the shared realization readable without exaggeration.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: edge of laptop display (On with the autopsy photograph displayed) — Only the camera-facing upper edge of the active display is visible; the detailed image remains mostly below frame; used as A restrained lower-frame reference tying both reactions to the same clue; adjacent chairs (Occupied) — The chair sides angle inward toward the laptop outside the lower frame; used as Compressed surrounding context that preserves their placement within the seated circle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same natural daytime interior illumination remains neutral-to-cool and restrained, preserving detail in both arrested expressions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 강주철, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office, laptop-centered seating, daylight, and the detectives' clothing. Exclude the senior officer and younger investigator from the tighter frame, focusing on the two startled men.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tailbone autopsy photograph remains displayed on the laptop as the team recognizes the significance of the clue.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 본능적인 단서를 깨달은 듯 눈이 동그랗게 커진 채 굳어버린 강주철과 서의용의 모습\n\nLOCATION (lock): Among the seats surrounding the laptop inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height on the circle's outer edge, hold a tight oblique two-shot with 강주철 on the nearer left and 서의용 set slightly deeper on the right. Their eyes remain fixed toward the laptop below the frame line as their forward motions halt at different instants—강주철 caught mid-lean and 서의용 with one hand suspended—making the shared realization readable without exaggeration.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: edge of laptop display (On with the autopsy photograph displayed) — Only the camera-facing upper edge of the active display is visible; the detailed image remains mostly below frame; used as A restrained lower-frame reference tying both reactions to the same clue; adjacent chairs (Occupied) — The chair sides angle inward toward the laptop outside the lower frame; used as Compressed surrounding context that preserves their placement within the seated circle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same natural daytime interior illumination remains neutral-to-cool and restrained, preserving detail in both arrested expressions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 강주철, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office, laptop-centered seating, daylight, and the detectives' clothing. Exclude the senior officer and younger investigator from the tighter frame, focusing on the two startled men.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tailbone autopsy photograph remains displayed on the laptop as the team recognizes the significance of the clue.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 본능적인 단서를 깨달은 듯 눈이 동그랗게 커진 채 굳어버린 강주철과 서의용의 모습\n\nLOCATION (lock): Among the seats surrounding the laptop inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height on the circle's outer edge, hold a tight oblique two-shot with 강주철 on the nearer left and 서의용 set slightly deeper on the right. Their eyes remain fixed toward the laptop below the frame line as their forward motions halt at different instants—강주철 caught mid-lean and 서의용 with one hand suspended—making the shared realization readable without exaggeration.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: edge of laptop display (On with the autopsy photograph displayed) — Only the camera-facing upper edge of the active display is visible; the detailed image remains mostly below frame; used as A restrained lower-frame reference tying both reactions to the same clue; adjacent chairs (Occupied) — The chair sides angle inward toward the laptop outside the lower frame; used as Compressed surrounding context that preserves their placement within the seated circle.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The same natural daytime interior illumination remains neutral-to-cool and restrained, preserving detail in both arrested expressions.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 강주철, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office, laptop-centered seating, daylight, and the detectives' clothing. Exclude the senior officer and younger investigator from the tighter frame, focusing on the two startled men.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The tailbone autopsy photograph remains displayed on the laptop as the team recognizes the significance of the clue.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "두 사람 모두 테이블 위의 노트북으로 시선을 내리고 있음.",
    "built_space": "사무실 원형 테이블 앞. 카메라가 강주철을 왼쪽, 서의용을 오른쪽에 두라는 지시를 어기고 기존 참조 사진의 좌우 배치(서의용 좌, 강주철 우)를 그대로 유지함.",
    "entities": "서의용(왼쪽, 파란 비니)과 강주철(오른쪽, 검은 비니)의 외모와 의상은 참조와 일치함. 화면에 부검 사진이 나타남.",
    "hard_violations": [
     "물리적으로 불가능한 사물: 노트북 키보드가 인물들을 향해 있으나, 활성화된 화면과 텍스트가 덮개 외부(카메라 방향)에 렌더링됨."
    ],
    "physics": "서의용의 오른손이 프롬프트 지시대로 허공에 멈춰 있음. 노트북 덮개 뒷면에 디스플레이가 나타나는 물리적 불가능성."
   },
   {
    "label": "A",
    "direction": "두 캐릭터 모두 눈 앞의 노트북을 응시함.",
    "built_space": "사무실 내 테이블 배치. 프롬프트가 요구한 좌우 인물 위치(강주철 좌측 전경, 서의용 우측)를 무시하고 기존 좌우 배치를 유지함.",
    "entities": "서의용과 강주철의 인상착의는 일치함. 노트북에 부검감정서가 선명하게 보임.",
    "hard_violations": [
     "배치/물리적 모순: 노트북 키보드와 화면이 완전히 카메라를 향하고 있어, 정작 노트북을 응시하는 인물들은 화면의 뒷면을 보고 있는 불가능한 상황."
    ],
    "physics": "노트북이 카메라 쪽으로 열려 있어 인물들과 사물 간의 상호작용 방향이 물리적으로 성립하지 않음. 두 사람의 손은 테이블과 다리에 지지되어 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 3,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "지시된 표정(커진 눈)과 정지된 손 동작을 잘 포착했으나, 인물 좌우 배치가 지시와 반대이며 노트북 화면이 덮개 외부에 표시되는 치명적인 물리적 오류가 있습니다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "인물 배치가 지시와 반대일 뿐 아니라, 노트북 전체가 카메라를 향해 있어 인물들이 화면 뒷면을 쳐다보는 모순된 샷이 연출되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 강주철, 서의용 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S50sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 강주철이 왼쪽, 서의용이 오른쪽에 위치하도록 지시했으나, 이미지에서는 서의용(가죽 재킷)이 왼쪽, 강주철(오렌지색 재킷)이 오른쪽에 있어 두 인물의 위치가 뒤바뀌었습니다.",
     "fix_en": "Swap the characters: redraw the left person as 강주철 wearing the orange jacket and grey beanie, and the right person as 서의용 wearing the black leather jacket and blue beanie. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "두 인물이 노트북 화면을 주시하고 있는 상황이나, 노트북의 실제 화면(이미지와 텍스트가 있는 면)이 인물들의 반대편인 카메라를 향하고 있어 공간적 방향이 잘못되었습니다.",
     "fix_en": "Redraw the laptop in the lower foreground to show its plain grey metallic back cover facing the camera, so the screen faces away toward the characters. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "노트북 화면 상단의 텍스트가 참조 이미지의 '부검감정서'가 아닌 '부검검정서'로 잘못 표기되었습니다.",
     "fix_en": "Redraw the text on the laptop screen into illegible blurred marks. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "강주철(오른쪽 주황 점퍼)이 앞으로 기울이다 굳은 게 아니라 입을 벌리고 오른손을 든 제스처이고, 서의용(왼쪽)은 한 손이 허공에 멈춘 상태가 아니다.",
     "fix_en": "Redraw the right character leaning forward with hands out of view, and the left character with one hand suspended in mid-air. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "노트북은 상단 테두리만 보이게 하라 했으나 하단 프레임에 부검 사진과 제목이 크게 드러나 있다.",
     "fix_en": "Redraw the lower frame to conceal the laptop display, leaving only its top edge visible at the very bottom. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 강주철이 왼쪽, 서의용이 오른쪽에 위치하도록 지시했으나, 이미지에서는 서의용(가죽 재킷)이 왼쪽, 강주철(오렌지색 재킷)이 오른쪽에 있어 두 인물의 위치가 뒤바뀌었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "두 인물이 노트북 화면을 주시하고 있는 상황이나, 노트북의 실제 화면(이미지와 텍스트가 있는 면)이 인물들의 반대편인 카메라를 향하고 있어 공간적 방향이 잘못되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "노트북 화면 상단의 텍스트가 참조 이미지의 '부검감정서'가 아닌 '부검검정서'로 잘못 표기되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "강주철(오른쪽 주황 점퍼)이 앞으로 기울이다 굳은 게 아니라 입을 벌리고 오른손을 든 제스처이고, 서의용(왼쪽)은 한 손이 허공에 멈춘 상태가 아니다.",
     "severity": "major"
    },
    {
     "issue_ko": "노트북은 상단 테두리만 보이게 하라 했으나 하단 프레임에 부검 사진과 제목이 크게 드러나 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "노트북 화면 제목이 이전 컷의 '부검감정서'가 아니라 '부검검정서'로 잘못 적혀 있다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "노트북 화면 상단의 텍스트가 참조 이미지의 '부검감정서'가 아닌 '부검검정서'로 잘못 표기되었습니다.",
    "fix_en": "Redraw the text on the laptop screen into illegible blurred marks. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "강주철(오른쪽 주황 점퍼)이 앞으로 기울이다 굳은 게 아니라 입을 벌리고 오른손을 든 제스처이고, 서의용(왼쪽)은 한 손이 허공에 멈춘 상태가 아니다.",
    "fix_en": "Redraw the right character leaning forward with hands out of view, and the left character with one hand suspended in mid-air. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "노트북은 상단 테두리만 보이게 하라 했으나 하단 프레임에 부검 사진과 제목이 크게 드러나 있다.",
    "fix_en": "Redraw the lower frame to conceal the laptop display, leaving only its top edge visible at the very bottom. Preserve the people present and their positions, their clothing, the set, the light, the framing.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Swap the characters: redraw the left person as 강주철 wearing the orange jacket and grey beanie, and the right person as 서의용 wearing the black leather jacket and blue beanie. Preserve the people present and their positions, their clothing, the set, the light, the framing.\n- Redraw the laptop in the lower foreground to show its plain grey metallic back cover facing the camera, so the screen faces away toward the characters. Preserve the people present and their positions, their clothing, the set, the light, the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍과 인물들의 놀란 표정 지시는 훌륭하게 따랐으나, 노트북 화면 방향이 물리적으로 불가능하게 연출되었고 화면 우측 하단에 다른 사람의 잘린 손이 남아있어 큰 감점이 발생했습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "지정된 두 인물의 좌우 위치는 텍스트에 맞췄으나, 프롬프트에서 명시적으로 제외하라고 한 두 인물을 그대로 포함시켰으며 클로즈업 샷과 표정 연기 요구사항을 완전히 무시했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 인물 모두 아래쪽의 노트북을 향해 시선을 고정하고 있으나, 노트북 화면이 카메라를 향해 열려 있어 실제로는 인물들이 화면 내용을 볼 수 없는 상태입니다.",
      "built_space": "이전 샷과 동일한 사무실 배경에서 원형 테이블을 사이에 두고 두 인물이 앉아 있습니다. 노트북은 테이블 가장자리 카메라 쪽에 위치해 있습니다.",
      "entities": "서의용(가죽 재킷, 파란색 비니)과 강주철(주황색 재킷, 회색 비니)의 외형은 잘 유지되었으나, 프롬프트의 지시(강주철 왼쪽, 서의용 오른쪽)와 반대로 서의용이 왼쪽, 강주철이 오른쪽에 배치되었습니다.",
      "hard_violations": [
       "물리적으로 불가능한 무대 연출 (노트북 화면이 카메라를 향하고 있어 인물들이 화면을 보고 놀라는 행동이 성립되지 않음)",
       "신체 부위 오류 (우측 하단 프레임 경계에 몸통이 없는 정체불명의 녹색 소매 손이 테이블 위에 놓여 있음)"
      ],
      "physics": "두 인물 모두 의자에 앉아 상체를 약간 기울인 채로 몸을 지지하고 있으나, 우측 하단의 손은 주인이 없는 상태로 허공/테이블에 얹혀 있습니다."
     },
     {
      "label": "B",
      "direction": "인물들의 시선이 테이블 중앙에 놓인 노트북을 향하고 있습니다.",
      "built_space": "사무실 원형 테이블을 둘러싸고 4개의 의자가 배치되어 있으며, 노트북은 정상적으로 사용자들을 향해 열려 있습니다.",
      "entities": "강주철이 왼쪽, 서의용이 오른쪽에 배치되어 텍스트의 위치 지시는 맞췄으나, 화면에 등장해서는 안 되는 이전 샷의 두 인물(정장 남성과 남색 재킷 남성)이 그대로 등장했습니다.",
      "hard_violations": [
       "지시되지 않은 인물 추가 및 제외 지시 위반 (프롬프트에서 제외를 명시한 정장 입은 남성과 남색 재킷을 입은 남성이 프레임에 포함됨)"
      ],
      "physics": "네 명의 인물 모두 의자에 앉아 바닥과 테이블을 통해 몸을 정상적으로 지지하고 있습니다."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍과 인물들의 놀란 표정 지시는 훌륭하게 따랐으나, 노트북 화면 방향이 물리적으로 불가능하게 연출되었고 화면 우측 하단에 다른 사람의 잘린 손이 남아있어 큰 감점이 발생했습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "지정된 두 인물의 좌우 위치는 텍스트에 맞췄으나, 프롬프트에서 명시적으로 제외하라고 한 두 인물을 그대로 포함시켰으며 클로즈업 샷과 표정 연기 요구사항을 완전히 무시했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "두 인물 모두 아래쪽의 노트북을 향해 시선을 고정하고 있으나, 노트북 화면이 카메라를 향해 열려 있어 실제로는 인물들이 화면 내용을 볼 수 없는 상태입니다.",
      "built_space": "이전 샷과 동일한 사무실 배경에서 원형 테이블을 사이에 두고 두 인물이 앉아 있습니다. 노트북은 테이블 가장자리 카메라 쪽에 위치해 있습니다.",
      "entities": "서의용(가죽 재킷, 파란색 비니)과 강주철(주황색 재킷, 회색 비니)의 외형은 잘 유지되었으나, 프롬프트의 지시(강주철 왼쪽, 서의용 오른쪽)와 반대로 서의용이 왼쪽, 강주철이 오른쪽에 배치되었습니다.",
      "hard_violations": [
       "물리적으로 불가능한 무대 연출 (노트북 화면이 카메라를 향하고 있어 인물들이 화면을 보고 놀라는 행동이 성립되지 않음)",
       "신체 부위 오류 (우측 하단 프레임 경계에 몸통이 없는 정체불명의 녹색 소매 손이 테이블 위에 놓여 있음)"
      ],
      "physics": "두 인물 모두 의자에 앉아 상체를 약간 기울인 채로 몸을 지지하고 있으나, 우측 하단의 손은 주인이 없는 상태로 허공/테이블에 얹혀 있습니다."
     },
     {
      "label": "B",
      "direction": "인물들의 시선이 테이블 중앙에 놓인 노트북을 향하고 있습니다.",
      "built_space": "사무실 원형 테이블을 둘러싸고 4개의 의자가 배치되어 있으며, 노트북은 정상적으로 사용자들을 향해 열려 있습니다.",
      "entities": "강주철이 왼쪽, 서의용이 오른쪽에 배치되어 텍스트의 위치 지시는 맞췄으나, 화면에 등장해서는 안 되는 이전 샷의 두 인물(정장 남성과 남색 재킷 남성)이 그대로 등장했습니다.",
      "hard_violations": [
       "지시되지 않은 인물 추가 및 제외 지시 위반 (프롬프트에서 제외를 명시한 정장 입은 남성과 남색 재킷을 입은 남성이 프레임에 포함됨)"
      ],
      "physics": "네 명의 인물 모두 의자에 앉아 바닥과 테이블을 통해 몸을 정상적으로 지지하고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "지시된 클로즈업 프레이밍과 두 인물의 놀란 표정, 행동(기울인 몸, 들린 손)은 잘 구현했으나, 인물의 좌우 배치가 뒤바뀌었고 노트북 화면이 카메라를 향해 있어 인물들이 모니터 뒷면을 보고 놀라는 논리적 오류가 발생했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "두 인물에 집중하는 타이트한 클로즈업 지시를 완전히 무시하고 배제해야 할 두 인물을 그대로 포함한 와이드 샷을 생성하여 프롬프트의 핵심 요구사항을 전면 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "두 인물의 시선이 하단의 노트북을 향해 있으나, 화면이 카메라 쪽을 향하고 있어 실제로는 노트북 덮개 뒷면을 응시하고 있음.",
      "built_space": "두 인물이 나란히 앉아 있으며, 카메라 앞에 놓인 노트북은 화면이 인물이 아닌 카메라를 정면으로 향하고 있어 공간적 모순이 있음.",
      "entities": "프롬프트 지시(강주철 왼쪽, 서의용 오른쪽)와 반대로, 왼쪽이 서의용(가죽 재킷), 오른쪽이 강주철(주황색 재킷)로 배치됨.",
      "hard_violations": [
       "인물의 좌우 배치 오류 (강주철과 서의용의 위치 뒤바뀜)",
       "인물들이 모니터 뒷면을 보고 화면 내용에 반응하는 물리적 시선 및 공간 설정 오류"
      ],
      "physics": "인물들은 의자에 착석해 있으며, 몸을 앞으로 기울이거나 손을 허공에 둔 지정된 자세는 안정적으로 지지되어 있음."
     },
     {
      "label": "A",
      "direction": "네 명의 인물 모두 테이블 중앙의 노트북 화면 쪽으로 시선을 향하고 있음.",
      "built_space": "원형 테이블 주변에 네 개의 의자가 배치되어 있고 각 인물이 착석해 있으며, 노트북은 좌측의 두 인물을 향해 열려 있음.",
      "entities": "서의용과 강주철의 인상착의는 맞으나, 프롬프트에서 명시적으로 배제할 것을 지시한 두 명의 인물(정장 차림 등)이 그대로 등장함.",
      "hard_violations": [
       "명시적으로 배제 지시된 두 명의 추가 인물 포함",
       "클로즈업(close-up) 프레이밍 지시를 완전히 위반한 와이드 샷"
      ],
      "physics": "모든 인물이 의자에 안정적으로 앉아 테이블이나 다리에 손을 올리고 있으며 물리적인 오류는 없음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 클로즈업 프레이밍과 두 인물의 놀란 표정, 행동(기울인 몸, 들린 손)은 잘 구현했으나, 인물의 좌우 배치가 뒤바뀌었고 노트북 화면이 카메라를 향해 있어 인물들이 모니터 뒷면을 보고 놀라는 논리적 오류가 발생했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "두 인물에 집중하는 타이트한 클로즈업 지시를 완전히 무시하고 배제해야 할 두 인물을 그대로 포함한 와이드 샷을 생성하여 프롬프트의 핵심 요구사항을 전면 위반했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "두 인물의 시선이 하단의 노트북을 향해 있으나, 화면이 카메라 쪽을 향하고 있어 실제로는 노트북 덮개 뒷면을 응시하고 있음.",
      "built_space": "두 인물이 나란히 앉아 있으며, 카메라 앞에 놓인 노트북은 화면이 인물이 아닌 카메라를 정면으로 향하고 있어 공간적 모순이 있음.",
      "entities": "프롬프트 지시(강주철 왼쪽, 서의용 오른쪽)와 반대로, 왼쪽이 서의용(가죽 재킷), 오른쪽이 강주철(주황색 재킷)로 배치됨.",
      "hard_violations": [
       "인물의 좌우 배치 오류 (강주철과 서의용의 위치 뒤바뀜)",
       "인물들이 모니터 뒷면을 보고 화면 내용에 반응하는 물리적 시선 및 공간 설정 오류"
      ],
      "physics": "인물들은 의자에 착석해 있으며, 몸을 앞으로 기울이거나 손을 허공에 둔 지정된 자세는 안정적으로 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "네 명의 인물 모두 테이블 중앙의 노트북 화면 쪽으로 시선을 향하고 있음.",
      "built_space": "원형 테이블 주변에 네 개의 의자가 배치되어 있고 각 인물이 착석해 있으며, 노트북은 좌측의 두 인물을 향해 열려 있음.",
      "entities": "서의용과 강주철의 인상착의는 맞으나, 프롬프트에서 명시적으로 배제할 것을 지시한 두 명의 인물(정장 차림 등)이 그대로 등장함.",
      "hard_violations": [
       "명시적으로 배제 지시된 두 명의 추가 인물 포함",
       "클로즈업(close-up) 프레이밍 지시를 완전히 위반한 와이드 샷"
      ],
      "physics": "모든 인물이 의자에 안정적으로 앉아 테이블이나 다리에 손을 올리고 있으며 물리적인 오류는 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 3
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh1"
  }
 },
 "S50sh5::cine": {
  "applied": true,
  "fingerprint": "90fd59a132449b5ee13ca23135a3fb6dd46c4aa116cac83274dd0fd5215f8da3",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S50sh5_sel.png",
  "source_sha256": "dc277adfc01c86fa6ce76e60f5fd0bac6d316b84aebaad6a3d427f4ed1f11977",
  "file": "S50sh5_cine.png",
  "latency_ms": 10825
 },
 "S50sh6::signage": {
  "fp": "b0efa1f20991a0c5",
  "inscriptions": [
   {
    "surface_native": "책상 위 명패",
    "text_native": "수사과장",
    "reason_ko": "수사과장실 내부의 회의 테이블이나 책상 위에 놓인 명패를 통해 사건이 진행되는 공간의 직책과 설정을 명확히 보여줍니다."
   }
  ]
 },
 "S50sh6": {
  "input_fingerprint": "34285632f214d8e0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흥분한 기색으로 두 주먹을 불끈 쥔 채 환하게 웃으며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the head of the laptop-centered meeting inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the short dolly-in from slightly below 전택수's seated eye line, holding him in an off-axis three-quarter medium close frame with his upper torso and both fists visible. He shifts forward in the chair, clenches both fists near his torso, and opens into an excited smile while looking toward the other investigators outside the frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: chair back (Occupied) — Its inner side sits directly behind 전택수's shoulders; used as A limited contextual edge behind his torso confirms that he remains seated within the discussion circle; investigation room background (Out of focus); used as Soft peripheral context suggesting the continuing group discussion without adding another visible subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent natural daytime interior illumination keeps the palette subdued and neutral-to-cool despite 전택수's sudden enthusiasm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sofa area, laptop setup, daylight, and the senior officer's clothing. Exclude the other three people from the close composition and show him raising clenched fists.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop continues displaying the tailbone autopsy photograph while Taeksu reacts to the breakthrough.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흥분한 기색으로 두 주먹을 불끈 쥔 채 환하게 웃으며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the head of the laptop-centered meeting inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the short dolly-in from slightly below 전택수's seated eye line, holding him in an off-axis three-quarter medium close frame with his upper torso and both fists visible. He shifts forward in the chair, clenches both fists near his torso, and opens into an excited smile while looking toward the other investigators outside the frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: chair back (Occupied) — Its inner side sits directly behind 전택수's shoulders; used as A limited contextual edge behind his torso confirms that he remains seated within the discussion circle; investigation room background (Out of focus); used as Soft peripheral context suggesting the continuing group discussion without adding another visible subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent natural daytime interior illumination keeps the palette subdued and neutral-to-cool despite 전택수's sudden enthusiasm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sofa area, laptop setup, daylight, and the senior officer's clothing. Exclude the other three people from the close composition and show him raising clenched fists.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop continues displaying the tailbone autopsy photograph while Taeksu reacts to the breakthrough.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 흥분한 기색으로 두 주먹을 불끈 쥔 채 환하게 웃으며 입을 벌린 전택수의 상체\n\nLOCATION (lock): At the head of the laptop-centered meeting inside the investigation chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the short dolly-in from slightly below 전택수's seated eye line, holding him in an off-axis three-quarter medium close frame with his upper torso and both fists visible. He shifts forward in the chair, clenches both fists near his torso, and opens into an excited smile while looking toward the other investigators outside the frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: chair back (Occupied) — Its inner side sits directly behind 전택수's shoulders; used as A limited contextual edge behind his torso confirms that he remains seated within the discussion circle; investigation room background (Out of focus); used as Soft peripheral context suggesting the continuing group discussion without adding another visible subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent natural daytime interior illumination keeps the palette subdued and neutral-to-cool despite 전택수's sudden enthusiasm.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sofa area, laptop setup, daylight, and the senior officer's clothing. Exclude the other three people from the close composition and show him raising clenched fists.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The laptop continues displaying the tailbone autopsy photograph while Taeksu reacts to the breakthrough.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"수사과장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "화면 밖(다른 수사관들 방향)을 향해 시선을 두고 환하게 웃고 있음.",
    "built_space": "사무실 책상 중앙에 노트북 1대와 명패 1개가 있으며, 인물 1명이 의자에 바르게 앉아 있음.",
    "entities": "캐릭터 레퍼런스와 일치하는 전택수(수사 모자, 경찰 점퍼 착용). 노트북 화면의 부검 사진과 명패의 '수사과장' 텍스트가 정확함.",
    "hard_violations": [],
    "physics": "의자 좌석에 엉덩이를 대고 앉아 상체를 안정적으로 지지하며 두 주먹을 허공에 쥐고 있음."
   },
   {
    "label": "B",
    "direction": "화면 밖을 향해 시선을 두고 흥분한 표정으로 입을 벌리고 있음.",
    "built_space": "사무실 책상 위에 노트북 1대와 명패 1개가 있으며, 인물 1명이 사무실 의자에 앉아 있음.",
    "entities": "전택수 레퍼런스와 불일치하며, 배제해야 할 이전 샷의 인물(비니, 가죽 재킷, 경찰 배지 목걸이)이 등장함.",
    "hard_violations": [
     "이전 샷 스틸의 인물 및 복장을 배제하라는 지시를 위반하고 잘못된 인물을 묘사함"
    ],
    "physics": "의자에 앉은 상태로 상체의 움직임을 지지하며 두 주먹을 공중에 띄우고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 8,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "지정된 캐릭터 레퍼런스의 외양과 복장을 정확히 반영했으며, 환하게 웃으며 두 주먹을 쥔 상체 묘사가 프롬프트와 완벽히 일치합니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "이전 샷의 인물과 복장을 화면에 포함하지 말라는 지시를 무시하고 잘못된 인물을 등장시켜 핵심 조건을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S50sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "책상 위 노트북의 키보드와 화면이 카메라 쪽을 향해 열려 있어, 책상 뒤에 앉은 인물에게는 노트북 뒷면이 향하게 되는 물리적으로 잘못된 방향입니다.",
     "fix_en": "Rotate the laptop 180 degrees so its solid back lid faces the camera and the screen faces the man. Preserve the desk, the man, and the background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "전택수 얼굴이 캐릭터 레퍼런스와 다른 인물로 보임(프레임 중앙)",
     "fix_en": "Modify the man's face to match the provided character reference. Preserve his expression, clothing, pose, and the background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "상체가 거의 정면이라 오프액시스 3/4 구도가 아님",
     "fix_en": "Angle the man's upper torso to a three-quarter pose. Preserve his expression, clenched fists, clothing, and background.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "재킷 왼쪽 가슴에 레퍼런스에 없는 '수사과장' 글자가 있음",
     "fix_en": "Remove the '수사과장' text from the jacket's chest, leaving only plain fabric. Preserve the jacket, the man's pose, and the background.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "책상 명패 글자가 '수사과장'이 아니라 띄어 쓴 '수 사 과 장'임",
     "fix_en": "Change the text on the nameplate to '수사과장' without spaces. Preserve the nameplate, desk, man, and background.",
     "severity": "minor",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "책상 위 노트북의 키보드와 화면이 카메라 쪽을 향해 열려 있어, 책상 뒤에 앉은 인물에게는 노트북 뒷면이 향하게 되는 물리적으로 잘못된 방향입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수 얼굴이 캐릭터 레퍼런스와 다른 인물로 보임(프레임 중앙)",
     "severity": "major"
    },
    {
     "issue_ko": "시선이 프레임 밖 수사관이 아니라 카메라를 향함",
     "severity": "major"
    },
    {
     "issue_ko": "상체가 거의 정면이라 오프액시스 3/4 구도가 아님",
     "severity": "major"
    },
    {
     "issue_ko": "노트북 화면이 전택수가 아니라 카메라를 향해 있어 사용자가 볼 수 없음",
     "severity": "major"
    },
    {
     "issue_ko": "재킷 왼쪽 가슴에 레퍼런스에 없는 '수사과장' 글자가 있음",
     "severity": "major"
    },
    {
     "issue_ko": "책상 명패 글자가 '수사과장'이 아니라 띄어 쓴 '수 사 과 장'임",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 6
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "전택수 얼굴이 캐릭터 레퍼런스와 다른 인물로 보임(프레임 중앙)",
    "fix_en": "Modify the man's face to match the provided character reference. Preserve his expression, clothing, pose, and the background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "상체가 거의 정면이라 오프액시스 3/4 구도가 아님",
    "fix_en": "Angle the man's upper torso to a three-quarter pose. Preserve his expression, clenched fists, clothing, and background.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "재킷 왼쪽 가슴에 레퍼런스에 없는 '수사과장' 글자가 있음",
    "fix_en": "Remove the '수사과장' text from the jacket's chest, leaving only plain fabric. Preserve the jacket, the man's pose, and the background.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "책상 명패 글자가 '수사과장'이 아니라 띄어 쓴 '수 사 과 장'임",
    "fix_en": "Change the text on the nameplate to '수사과장' without spaces. Preserve the nameplate, desk, man, and background.",
    "severity": "minor",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Rotate the laptop 180 degrees so its solid back lid faces the camera and the screen faces the man. Preserve the desk, the man, and the background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "레퍼런스 인물의 외모와 복장을 완벽하게 일치시켰으며, 명패의 텍스트와 노트북 화면의 부검 사진까지 프롬프트의 모든 주요 지시사항을 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물의 표정과 명패 텍스트는 훌륭하나, 노트북 화면에 사진이 띄워지지 않고 기기의 형태가 물리적으로 모순되게 렌더링되어 심각한 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 우측 프레임 밖의 다른 수사관들을 향하고 있으며, 두 주먹은 자신의 상체 앞으로 불끈 쥐고 있음.",
      "built_space": "수사과장 사무실 내부. 전택수는 책상 뒤 의자에 착석해 있으며, 어깨 뒤로 의자 등받이가 위치함. 책상 위에는 카메라를 향해 열린 노트북과 명패가 놓여 있고 배경에 소파와 수납장이 보임.",
      "entities": "전택수의 얼굴과 50대 남성의 특징, '수사' 텍스트가 적힌 모자와 경찰 재킷 등 레퍼런스와 정확히 일치함. 명패에 '수사과장'이 올바르게 적혀 있으며, 노트북 화면에 부검 사진과 텍스트가 명확히 표시됨.",
      "hard_violations": [],
      "physics": "인물의 하중은 의자가 지탱하고 있으며, 공중에 들어 올린 두 주먹은 팔 근육이 지탱함. 노트북과 명패는 책상 표면 위에 안정적으로 놓여 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 화면 우측 프레임 밖을 향하고 있으며, 양손의 주먹을 들어 올리고 있음.",
      "built_space": "사무실 내부 책상 뒤에 인물이 착석해 있으며 어깨 뒤로 의자 등받이가 보임. 책상 위에 명패와 노트북이 배치됨.",
      "entities": "전택수의 인물 특징과 복장, 명패의 '수사과장' 텍스트는 프롬프트와 일치함. 그러나 노트북 화면에 부검 사진이 묘사되지 않고 회색의 텅 빈 면으로 나타남.",
      "hard_violations": [
       "키보드와 트랙패드가 카메라 쪽을 향하고 있으나 화면부가 표시되어야 할 안쪽 면이 밋밋한 금속판처럼 묘사된 물리적으로 불가능한 노트북 기기 구조"
      ],
      "physics": "인물은 의자에 정상적으로 앉아 있고 두 팔은 들어 올려진 상태를 유지함. 노트북이 책상 위에 놓여 있으나 기기 내부의 물리적 구조가 성립하지 않음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "레퍼런스 인물의 외모와 복장을 완벽하게 일치시켰으며, 명패의 텍스트와 노트북 화면의 부검 사진까지 프롬프트의 모든 주요 지시사항을 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "인물의 표정과 명패 텍스트는 훌륭하나, 노트북 화면에 사진이 띄워지지 않고 기기의 형태가 물리적으로 모순되게 렌더링되어 심각한 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 우측 프레임 밖의 다른 수사관들을 향하고 있으며, 두 주먹은 자신의 상체 앞으로 불끈 쥐고 있음.",
      "built_space": "수사과장 사무실 내부. 전택수는 책상 뒤 의자에 착석해 있으며, 어깨 뒤로 의자 등받이가 위치함. 책상 위에는 카메라를 향해 열린 노트북과 명패가 놓여 있고 배경에 소파와 수납장이 보임.",
      "entities": "전택수의 얼굴과 50대 남성의 특징, '수사' 텍스트가 적힌 모자와 경찰 재킷 등 레퍼런스와 정확히 일치함. 명패에 '수사과장'이 올바르게 적혀 있으며, 노트북 화면에 부검 사진과 텍스트가 명확히 표시됨.",
      "hard_violations": [],
      "physics": "인물의 하중은 의자가 지탱하고 있으며, 공중에 들어 올린 두 주먹은 팔 근육이 지탱함. 노트북과 명패는 책상 표면 위에 안정적으로 놓여 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 화면 우측 프레임 밖을 향하고 있으며, 양손의 주먹을 들어 올리고 있음.",
      "built_space": "사무실 내부 책상 뒤에 인물이 착석해 있으며 어깨 뒤로 의자 등받이가 보임. 책상 위에 명패와 노트북이 배치됨.",
      "entities": "전택수의 인물 특징과 복장, 명패의 '수사과장' 텍스트는 프롬프트와 일치함. 그러나 노트북 화면에 부검 사진이 묘사되지 않고 회색의 텅 빈 면으로 나타남.",
      "hard_violations": [
       "키보드와 트랙패드가 카메라 쪽을 향하고 있으나 화면부가 표시되어야 할 안쪽 면이 밋밋한 금속판처럼 묘사된 물리적으로 불가능한 노트북 기기 구조"
      ],
      "physics": "인물은 의자에 정상적으로 앉아 있고 두 팔은 들어 올려진 상태를 유지함. 노트북이 책상 위에 놓여 있으나 기기 내부의 물리적 구조가 성립하지 않음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1750,
      "verdict_ko": "인물의 표정과 제스처, 명패 텍스트를 정확히 구현했으며, 프롬프트의 지시대로 기하학적 구조에 맞게 노트북이 사용자를 향하도록 올바르게 배치했습니다."
     },
     {
      "label": "B",
      "score": 1083,
      "verdict_ko": "화면 내용을 보여주기 위해 노트북 전체를 카메라 방향으로 뒤집어 배치함으로써, 사용자가 노트북 뒷면을 마주하게 되는 치명적인 무대 구성 오류(Hard Violation)를 범했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 무대 구성: 노트북의 키보드와 화면이 사용자가 아닌 카메라를 향해 거꾸로 배치됨"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.75,
      "B": 1.083
     },
     "violations": {
      "B": [
       "[gemini-pro] 물리적으로 불가능한 무대 구성: 노트북의 키보드와 화면이 사용자가 아닌 카메라를 향해 거꾸로 배치됨"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "인물의 표정과 제스처, 명패 텍스트를 정확히 구현했으며, 프롬프트의 지시대로 기하학적 구조에 맞게 노트북이 사용자를 향하도록 올바르게 배치했습니다."
     },
     {
      "label": "A",
      "score": 1083,
      "verdict_ko": "화면 내용을 보여주기 위해 노트북 전체를 카메라 방향으로 뒤집어 배치함으로써, 사용자가 노트북 뒷면을 마주하게 되는 치명적인 무대 구성 오류(Hard Violation)를 범했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 무대 구성: 노트북의 키보드와 화면이 사용자가 아닌 카메라를 향해 거꾸로 배치됨"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1092,
     "B": 1754
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S50sh5"
  }
 },
 "S50sh6::cine": {
  "applied": true,
  "fingerprint": "4a229b660c6e7cce0a10e7813a1a84feae4f623b0bf201cd553aaba06a687f76",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S50sh6_sel.png",
  "source_sha256": "5510e36816b7227dca15732cfa53fad602ca621ef3cbb9daffc4eab94e2b8ff5",
  "file": "S50sh6_cine.png",
  "latency_ms": 12344
 },
 "S51sh1::signage": {
  "fp": "4ca616e787a94792",
  "inscriptions": [
   {
    "surface_native": "계단 벽면 안내판",
    "text_native": "2F 형사과",
    "reason_ko": "경찰서 내부 계단 벽면에 층별 안내 표지판을 배치하여 사실적인 경찰서 분위기를 연출하기 위함입니다."
   }
  ]
 },
 "S51sh1": {
  "input_fingerprint": "37ca7692c0f82bca",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 계단 중간에 선 채 아래쪽을 향해 눈을 동그랗게 뜨고 멈칫한 전택수의 전경\n\nLOCATION (lock): On the police-station stairway where it meets the main corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the side of the staircase at slightly above head height, frame 전택수 in a medium-long diagonal view as he arrests his descent on a step. His weight remains over the lower foot while the trailing foot stays higher, shoulders tightening and eyes directed down toward the police chief and reporter outside the frame.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: staircase (전택수 is stopped midway down) — The steps descend diagonally from the upper side of frame toward the corridor below; used as Creates the descending diagonal and visibly interrupts 전택수's movement; corridor below the stairs (Visible beyond the staircase); used as Provides the lower destination of his eyeline while keeping the people he noticed outside this opening frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool palette and moderate contrast supports the awkward pause.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station stairwell and adjoining corridor, institutional walls, and daytime light. Exclude the station chief previously on the stairs and show the senior officer stopping mid-descent.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 계단 벽면 안내판: \"2F 형사과\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 계단 중간에 선 채 아래쪽을 향해 눈을 동그랗게 뜨고 멈칫한 전택수의 전경\n\nLOCATION (lock): On the police-station stairway where it meets the main corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the side of the staircase at slightly above head height, frame 전택수 in a medium-long diagonal view as he arrests his descent on a step. His weight remains over the lower foot while the trailing foot stays higher, shoulders tightening and eyes directed down toward the police chief and reporter outside the frame.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: staircase (전택수 is stopped midway down) — The steps descend diagonally from the upper side of frame toward the corridor below; used as Creates the descending diagonal and visibly interrupts 전택수's movement; corridor below the stairs (Visible beyond the staircase); used as Provides the lower destination of his eyeline while keeping the people he noticed outside this opening frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool palette and moderate contrast supports the awkward pause.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station stairwell and adjoining corridor, institutional walls, and daytime light. Exclude the station chief previously on the stairs and show the senior officer stopping mid-descent.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 계단 벽면 안내판: \"2F 형사과\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 계단 중간에 선 채 아래쪽을 향해 눈을 동그랗게 뜨고 멈칫한 전택수의 전경\n\nLOCATION (lock): On the police-station stairway where it meets the main corridor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the side of the staircase at slightly above head height, frame 전택수 in a medium-long diagonal view as he arrests his descent on a step. His weight remains over the lower foot while the trailing foot stays higher, shoulders tightening and eyes directed down toward the police chief and reporter outside the frame.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: staircase (전택수 is stopped midway down) — The steps descend diagonally from the upper side of frame toward the corridor below; used as Creates the descending diagonal and visibly interrupts 전택수's movement; corridor below the stairs (Visible beyond the staircase); used as Provides the lower destination of his eyeline while keeping the people he noticed outside this opening frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool palette and moderate contrast supports the awkward pause.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station stairwell and adjoining corridor, institutional walls, and daytime light. Exclude the station chief previously on the stairs and show the senior officer stopping mid-descent.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 계단 벽면 안내판: \"2F 형사과\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "시선이 전경에 있는 인물의 뒷모습을 향해 아래로 향함.",
    "built_space": "이전 샷과 동일한 계단 구조와 안내판 위치가 정확히 묘사됨.",
    "entities": "전택수의 얼굴과 복장(모자, 점퍼)이 레퍼런스와 일치함. 안내판의 '2F 형사과' 텍스트가 정확함. 프롬프트에 없는 전경 인물이 존재함.",
    "hard_violations": [
     "프롬프트에 명시되지 않은 추가 인물(전경의 뒷모습) 포함",
     "이전 샷의 앵글과 전경 인물 배치를 금지 지시에도 불구하고 그대로 복사함"
    ],
    "physics": "계단 중간에서 아래쪽 발에 체중을 싣고 안정적으로 정지한 자세를 유지함."
   },
   {
    "label": "B",
    "direction": "시선이 전경에 있는 인물의 뒷모습을 향해 아래로 향함.",
    "built_space": "이전 샷과 동일한 계단 구조와 안내판 위치가 정확히 묘사됨.",
    "entities": "전택수의 의상(가죽 재킷, 모자 부재)이 레퍼런스와 완전히 다름. 안내판의 '2F 형사과' 텍스트는 정확함. 프롬프트에 없는 전경 인물이 존재함.",
    "hard_violations": [
     "프롬프트에 명시되지 않은 추가 인물(전경의 뒷모습) 포함",
     "이전 샷의 앵글과 전경 인물 배치를 금지 지시에도 불구하고 그대로 복사함"
    ],
    "physics": "계단 중간에서 아래쪽 발에 체중을 싣고 안정적으로 정지한 자세를 유지함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 3,
   "B": 2
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "전택수의 외형과 표지판 텍스트는 정확히 구현했으나, 프레임 밖에 있어야 할 이전 샷의 인물(전경의 뒷모습)을 포함시켜 핵심 지시를 위반했습니다."
   },
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "금지된 추가 인물을 화면에 포함시켰을 뿐만 아니라, 캐릭터의 의상(모자 제외, 가죽 재킷)마저 레퍼런스를 전혀 따르지 않았습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S10sh13_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트에서 프레임 밖에 있어야 한다고 명시했으며 추가하지 말라고 한 인물(이전 샷의 전경 인물)이 화면 왼쪽 앞부분에 포함되어 있습니다.",
     "fix_en": "Remove the person in the left foreground entirely, filling that area with the continuation of the plain white wall and the empty corridor space. Preserve Jeon Taek-soo exactly as he is, including his face, clothing, and pose, as well as the staircase, the framed sign, and the lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "전택수의 시선이 프레임 밖 아래가 아니라 전경 인물을 향한다.",
     "fix_en": "Redirect the subject's eyes to look further downward toward the bottom edge of the frame. Preserve the head angle, the rest of his expression, and the background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "계단 측면 약간 위 대각 구도가 아니라 이전 스틸과 같은 복도 쪽 오버숄더 구도이다.",
     "fix_en": "Remove the foreground obstruction to slightly mitigate the over-the-shoulder perspective, as the camera position itself cannot be moved in-place. Preserve the subject and the staircase.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "눈을 동그랗게 뜨고 멈칫한 것이 아니라 무표정에 가깝게 걸어 내려온다.",
     "fix_en": "Widen the subject's eyes to convey an abrupt, startled pause. Preserve the rest of the face, the clothing, and the background.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "아랫발에 무게를 두고 뒷발이 높은 정지 자세가 아니라 전경 쪽으로 걷는 보행이다.",
     "fix_en": "Adjust the subject's posture to shift his weight onto a firmly planted lower foot with the trailing foot visibly raised. Preserve the clothing and the surrounding steps.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트에서 프레임 밖에 있어야 한다고 명시했으며 추가하지 말라고 한 인물(이전 샷의 전경 인물)이 화면 왼쪽 앞부분에 포함되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "샷 텍스트에 없는 인물이 화면 왼쪽 전경에 등 뒤로 들어와 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수의 시선이 프레임 밖 아래가 아니라 전경 인물을 향한다.",
     "severity": "major"
    },
    {
     "issue_ko": "계단 측면 약간 위 대각 구도가 아니라 이전 스틸과 같은 복도 쪽 오버숄더 구도이다.",
     "severity": "major"
    },
    {
     "issue_ko": "눈을 동그랗게 뜨고 멈칫한 것이 아니라 무표정에 가깝게 걸어 내려온다.",
     "severity": "major"
    },
    {
     "issue_ko": "아랫발에 무게를 두고 뒷발이 높은 정지 자세가 아니라 전경 쪽으로 걷는 보행이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "전택수의 시선이 프레임 밖 아래가 아니라 전경 인물을 향한다.",
    "fix_en": "Redirect the subject's eyes to look further downward toward the bottom edge of the frame. Preserve the head angle, the rest of his expression, and the background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "계단 측면 약간 위 대각 구도가 아니라 이전 스틸과 같은 복도 쪽 오버숄더 구도이다.",
    "fix_en": "Remove the foreground obstruction to slightly mitigate the over-the-shoulder perspective, as the camera position itself cannot be moved in-place. Preserve the subject and the staircase.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "눈을 동그랗게 뜨고 멈칫한 것이 아니라 무표정에 가깝게 걸어 내려온다.",
    "fix_en": "Widen the subject's eyes to convey an abrupt, startled pause. Preserve the rest of the face, the clothing, and the background.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "아랫발에 무게를 두고 뒷발이 높은 정지 자세가 아니라 전경 쪽으로 걷는 보행이다.",
    "fix_en": "Adjust the subject's posture to shift his weight onto a firmly planted lower foot with the trailing foot visibly raised. Preserve the clothing and the surrounding steps.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the person in the left foreground entirely, filling that area with the continuation of the plain white wall and the empty corridor space. Preserve Jeon Taek-soo exactly as he is, including his face, clothing, and pose, as well as the staircase, the framed sign, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프롬프트의 지시대로 이전 샷의 전경 인물을 완벽히 제거하고 전택수 단독 샷을 구성했으며, 안내판 텍스트와 하행 정지 포즈를 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시사항에 어긋나게 프롬프트에 없는 이전 샷의 전경 인물이 화면 좌측에 그대로 남아 있어 감점 요소가 큽니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 좌측 하단 프레임 밖을 향함.",
      "built_space": "경찰서 내부 계단과 벽면이 보이며, 우측 벽에 안내판이 위치함.",
      "entities": "전택수의 복장 및 외모가 레퍼런스와 일치하고, 안내판에 '2F 형사과'가 올바르게 적혀 있으나, 좌측 전경에 프롬프트에 없는 인물의 뒷모습이 등장함.",
      "hard_violations": [
       "invented people (프롬프트가 허용하지 않은 좌측 전경의 인물 등장)"
      ],
      "physics": "전택수의 두 발이 계단 표면을 딛고 체중을 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 화면 좌측 하단 프레임 밖을 향함.",
      "built_space": "경찰서 내부 계단과 복도가 보이며, 벽면 구조와 안내판 위치가 정확함.",
      "entities": "화면에 전택수만 등장하며(추가 인물 없음), 캐릭터 레퍼런스와 복장이 일치하고 안내판의 '2F 형사과' 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "아래쪽 발에 체중을 싣고 계단에 멈춰 선 자세가 바닥에 의해 안정적으로 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프롬프트의 지시대로 이전 샷의 전경 인물을 완벽히 제거하고 전택수 단독 샷을 구성했으며, 안내판 텍스트와 하행 정지 포즈를 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시사항에 어긋나게 프롬프트에 없는 이전 샷의 전경 인물이 화면 좌측에 그대로 남아 있어 감점 요소가 큽니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 화면 좌측 하단 프레임 밖을 향함.",
      "built_space": "경찰서 내부 계단과 벽면이 보이며, 우측 벽에 안내판이 위치함.",
      "entities": "전택수의 복장 및 외모가 레퍼런스와 일치하고, 안내판에 '2F 형사과'가 올바르게 적혀 있으나, 좌측 전경에 프롬프트에 없는 인물의 뒷모습이 등장함.",
      "hard_violations": [
       "invented people (프롬프트가 허용하지 않은 좌측 전경의 인물 등장)"
      ],
      "physics": "전택수의 두 발이 계단 표면을 딛고 체중을 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 화면 좌측 하단 프레임 밖을 향함.",
      "built_space": "경찰서 내부 계단과 복도가 보이며, 벽면 구조와 안내판 위치가 정확함.",
      "entities": "화면에 전택수만 등장하며(추가 인물 없음), 캐릭터 레퍼런스와 복장이 일치하고 안내판의 '2F 형사과' 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "아래쪽 발에 체중을 싣고 계단에 멈춰 선 자세가 바닥에 의해 안정적으로 지지됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프레임 밖으로 제외하라고 명시된 인물들을 완벽히 배제하고 전택수 단독 프레이밍과 안내판 텍스트를 정확히 구현하여 지시사항을 훌륭하게 이행함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레임 밖에 있어야 할 인물의 뒷모습을 전경에 배치하여 등장인물 제한 지시를 심각하게 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 프레임 밖 왼쪽 아래쪽을 향하고 있음.",
      "built_space": "이전 컷과 동일한 형태의 계단과 우측 난간, 벽면에 부착된 안내판이 보임.",
      "entities": "전택수의 얼굴, 체형, 복장(수사 모자, 남색 재킷, 체크 셔츠, 바지)이 레퍼런스와 정확히 일치함. 벽면 안내판에 '2F 형사과' 텍스트가 명확하게 쓰여 있음.",
      "hard_violations": [],
      "physics": "전택수는 계단 중간에 멈춰 서서 아래쪽 발에 체중을 싣고 위쪽 발은 뒤에 둔 자세로 계단 표면에 안정적으로 지탱되어 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 프레임 왼쪽 전경에 위치한 인물의 뒷모습을 향하고 있음.",
      "built_space": "동일한 형태의 계단, 난간, 벽면 안내판이 묘사됨.",
      "entities": "전택수의 외형과 복장, 안내판 텍스트는 정확하게 묘사되었으나, 프롬프트에서 프레임 밖으로 제외할 것을 명시한 신원 미상의 인물이 왼쪽 전경에 나타남.",
      "hard_violations": [
       "invented people or extra bodies: 프롬프트에서 프레임 내 등장인물을 전택수 1인으로 제한하고 나머지 인물은 프레임 밖에 두라고 명시했으나, 전경에 추가 인물을 렌더링함."
      ],
      "physics": "전택수는 계단에 서 있는 상태로 물리적으로 자연스럽게 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프레임 밖으로 제외하라고 명시된 인물들을 완벽히 배제하고 전택수 단독 프레이밍과 안내판 텍스트를 정확히 구현하여 지시사항을 훌륭하게 이행함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레임 밖에 있어야 할 인물의 뒷모습을 전경에 배치하여 등장인물 제한 지시를 심각하게 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전택수의 시선은 프레임 밖 왼쪽 아래쪽을 향하고 있음.",
      "built_space": "이전 컷과 동일한 형태의 계단과 우측 난간, 벽면에 부착된 안내판이 보임.",
      "entities": "전택수의 얼굴, 체형, 복장(수사 모자, 남색 재킷, 체크 셔츠, 바지)이 레퍼런스와 정확히 일치함. 벽면 안내판에 '2F 형사과' 텍스트가 명확하게 쓰여 있음.",
      "hard_violations": [],
      "physics": "전택수는 계단 중간에 멈춰 서서 아래쪽 발에 체중을 싣고 위쪽 발은 뒤에 둔 자세로 계단 표면에 안정적으로 지탱되어 있음."
     },
     {
      "label": "A",
      "direction": "전택수의 시선은 프레임 왼쪽 전경에 위치한 인물의 뒷모습을 향하고 있음.",
      "built_space": "동일한 형태의 계단, 난간, 벽면 안내판이 묘사됨.",
      "entities": "전택수의 외형과 복장, 안내판 텍스트는 정확하게 묘사되었으나, 프롬프트에서 프레임 밖으로 제외할 것을 명시한 신원 미상의 인물이 왼쪽 전경에 나타남.",
      "hard_violations": [
       "invented people or extra bodies: 프롬프트에서 프레임 내 등장인물을 전택수 1인으로 제한하고 나머지 인물은 프레임 밖에 두라고 명시했으나, 전경에 추가 인물을 렌더링함."
      ],
      "physics": "전택수는 계단에 서 있는 상태로 물리적으로 자연스럽게 지탱됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S10sh13"
  }
 },
 "S51sh1::cine": {
  "applied": true,
  "fingerprint": "0bac42d3203916a26fa0c25e3a0867d9d194599198bc6ab5da4ac1b22fd79f0e",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S51sh1_sel.png",
  "source_sha256": "d36613f053b02aae0bc2c0bf29f59b5b24ca5dd85bb3e10d44186005771bc3f4",
  "file": "S51sh1_cine.png",
  "latency_ms": 13305
 },
 "S51sh4::signage": {
  "fp": "41ec7429536fb7fb",
  "inscriptions": [
   {
    "surface_native": "사무실 현판",
    "text_native": "강력계",
    "reason_ko": "경찰서 복도 및 강력계 사무실 부근이라는 공간적 배경을 사실적으로 나타내기 위해 문 위에 부착된 부서 표시판이 필요합니다."
   }
  ]
 },
 "S51sh4": {
  "input_fingerprint": "b54ae7296c2d6777",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 쪽을 향해 손가락질을 하며 굳은 표정으로 입을 벌린 나주경찰서장의 전경\n\nLOCATION (lock): In the main police-station corridor near the stairway and violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At upper-torso height along the corridor, let the lateral track settle into a medium view of 나주경찰서장 in the foreground, angled across his body rather than straight on. He plants his forward foot, points firmly toward 전택수 just beyond the left frame edge, and fixes his gaze there while speaking with a set expression.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-right of the frame, foreground, points to 전택수 beyond the left frame edge.\n- KEY BACKGROUND ELEMENTS: police-station corridor (In active use during the exchange); used as The receding corridor line carries the pointing gesture toward 전택수's attempted retreat.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination is kept subdued and neutral-to-cool, with controlled contrast on the chief's firm expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor and stairwell geometry, daylight, and police-station finishes. Exclude the senior officer's solitary stair pose and add the chief and reporter approaching him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 현판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 쪽을 향해 손가락질을 하며 굳은 표정으로 입을 벌린 나주경찰서장의 전경\n\nLOCATION (lock): In the main police-station corridor near the stairway and violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At upper-torso height along the corridor, let the lateral track settle into a medium view of 나주경찰서장 in the foreground, angled across his body rather than straight on. He plants his forward foot, points firmly toward 전택수 just beyond the left frame edge, and fixes his gaze there while speaking with a set expression.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-right of the frame, foreground, points to 전택수 beyond the left frame edge.\n- KEY BACKGROUND ELEMENTS: police-station corridor (In active use during the exchange); used as The receding corridor line carries the pointing gesture toward 전택수's attempted retreat.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination is kept subdued and neutral-to-cool, with controlled contrast on the chief's firm expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor and stairwell geometry, daylight, and police-station finishes. Exclude the senior officer's solitary stair pose and add the chief and reporter approaching him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 현판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 쪽을 향해 손가락질을 하며 굳은 표정으로 입을 벌린 나주경찰서장의 전경\n\nLOCATION (lock): In the main police-station corridor near the stairway and violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At upper-torso height along the corridor, let the lateral track settle into a medium view of 나주경찰서장 in the foreground, angled across his body rather than straight on. He plants his forward foot, points firmly toward 전택수 just beyond the left frame edge, and fixes his gaze there while speaking with a set expression.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 나주경찰서장 in the middle-right of the frame, foreground, points to 전택수 beyond the left frame edge.\n- KEY BACKGROUND ELEMENTS: police-station corridor (In active use during the exchange); used as The receding corridor line carries the pointing gesture toward 전택수's attempted retreat.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination is kept subdued and neutral-to-cool, with controlled contrast on the chief's firm expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor and stairwell geometry, daylight, and police-station finishes. Exclude the senior officer's solitary stair pose and add the chief and reporter approaching him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 현판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.625,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.125,
    "B": 1.75
   },
   "violations": {
    "A": [
     "[gemini-pro] invented objects (제복 가슴의 은색 배지 및 추가 마크)",
     "[gemini-pro] leaked text ('2F 형사과' 텍스트 포함)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1125
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "레퍼런스의 인물 외모와 복장을 완벽하게 구현했으며, 불필요한 텍스트 없이 지시된 '강력계' 현판과 정확한 시선 및 지시 동작을 훌륭하게 표현했습니다."
   },
   {
    "label": "A",
    "score": 1125,
    "verdict_ko": "요구된 동작과 표정은 잘 나타났으나, 레퍼런스에 없는 은색 가슴 배지가 추가되었고 지시되지 않은 텍스트('2F 형사과')가 포함되었습니다.  ★위반: [gemini-pro] invented objects (제복 가슴의 은색 배지 및 추가 마크) / [gemini-pro] leaked text ('2F 형사과' 텍스트 포함)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S51sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832602>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "오른쪽 어깨(화면 좌측) 견장의 별 장식이 2개로 렌더링되어, 3개인 왼쪽 견장 및 기준 이미지와 일치하지 않습니다.",
     "fix_en": "Redraw the right shoulder epaulet (viewer's left) to feature three star insignias, preserving the character's face, pose, uniform, and the current background.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "지시문에 명시된 '뒤로 멀어지는 복도 선(receding corridor line)'이 배경에 나타나지 않으며, 화면 왼쪽이 계단 밑 벽으로 막혀 있습니다.",
     "fix_en": "Redraw the left background to show a receding corridor space, maintaining the character, his pointing pose, and the lighting.",
     "severity": "minor",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "왼쪽 방향을 가리키는 오른손의 엄지손가락 관절과 솟아오른 각도가 해부학적으로 다소 어색합니다.",
     "fix_en": "Redraw the pointing right hand to correct the thumb's anatomical structure and angle, keeping the character's face, uniform, overall pose, and the background unchanged.",
     "severity": "minor",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "오른쪽 어깨(화면 좌측) 견장의 별 장식이 2개로 렌더링되어, 3개인 왼쪽 견장 및 기준 이미지와 일치하지 않습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "넥타이 하단 끝이 뾰족하지 않고 평평하게 일자로 잘린 형태로 렌더링되어 벨트와 부자연스럽게 맞닿아 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "지시문에 명시된 '뒤로 멀어지는 복도 선(receding corridor line)'이 배경에 나타나지 않으며, 화면 왼쪽이 계단 밑 벽으로 막혀 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "왼쪽 방향을 가리키는 오른손의 엄지손가락 관절과 솟아오른 각도가 해부학적으로 다소 어색합니다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "오른쪽 어깨(화면 좌측) 견장의 별 장식이 2개로 렌더링되어, 3개인 왼쪽 견장 및 기준 이미지와 일치하지 않습니다.",
    "fix_en": "Redraw the right shoulder epaulet (viewer's left) to feature three star insignias, preserving the character's face, pose, uniform, and the current background.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "지시문에 명시된 '뒤로 멀어지는 복도 선(receding corridor line)'이 배경에 나타나지 않으며, 화면 왼쪽이 계단 밑 벽으로 막혀 있습니다.",
    "fix_en": "Redraw the left background to show a receding corridor space, maintaining the character, his pointing pose, and the lighting.",
    "severity": "minor",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "왼쪽 방향을 가리키는 오른손의 엄지손가락 관절과 솟아오른 각도가 해부학적으로 다소 어색합니다.",
    "fix_en": "Redraw the pointing right hand to correct the thumb's anatomical structure and angle, keeping the character's face, uniform, overall pose, and the background unchanged.",
    "severity": "minor",
    "observation_index": 3
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S51sh1"
  }
 },
 "S51sh4::cine": {
  "applied": true,
  "fingerprint": "6a352779ec6e5971d15b497776163579f32022b262cccc648117094a7933f795",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S51sh4_sel.png",
  "source_sha256": "7e103293c327131a7287dc3f1a7a20ad95885fbe0b32d0112b17db4cacabb3fb",
  "file": "S51sh4_cine.png",
  "latency_ms": 10745
 },
 "S51sh7::signage": {
  "fp": "9805ef0dcaaa05a4",
  "inscriptions": [
   {
    "surface_native": "사무실 문옆 표지판",
    "text_native": "강력계",
    "reason_ko": "사건이 일어나는 경찰서 강력계 복도라는 공간적 배경을 확실히 나타내기 위해 사무실 표지판이 필요합니다."
   }
  ]
 },
 "S51sh7": {
  "input_fingerprint": "61e1f47d622f13d9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나주경찰서장과 30대 한국인 남성 기자를 향해 단호한 눈빛으로 고개를 한쪽으로 비튼 전택수의 상체\n\nLOCATION (lock): In the corridor immediately outside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height just beyond the shared shoulder line of 나주경찰서장 and the reporter, finish the restrained dolly-in on 전택수's three-quarter upper body, leaving only a soft sliver of the intervening shoulder at the frame edge. 전택수 twists his head back over a body still angled toward the office, chin canted to one side and eyes locked firmly on 나주경찰서장 as he refuses the interview.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 나주경찰서장 beside the camera.\n- KEY BACKGROUND ELEMENTS: corridor toward the violent-crimes office (Open as his intended path); used as The corridor direction behind 전택수 preserves the route he intends to continue along; soft shoulder edge (Mostly excluded from frame); used as A minimal foreground occlusion establishes the conversational side without introducing a second full subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent natural daytime corridor illumination maintains a restrained neutral-to-cool procedural mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor, stairwell background, daylight, and all three people's clothing. Exclude the chief's pointing gesture and frame the senior officer responding firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 문옆 표지판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나주경찰서장과 30대 한국인 남성 기자를 향해 단호한 눈빛으로 고개를 한쪽으로 비튼 전택수의 상체\n\nLOCATION (lock): In the corridor immediately outside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height just beyond the shared shoulder line of 나주경찰서장 and the reporter, finish the restrained dolly-in on 전택수's three-quarter upper body, leaving only a soft sliver of the intervening shoulder at the frame edge. 전택수 twists his head back over a body still angled toward the office, chin canted to one side and eyes locked firmly on 나주경찰서장 as he refuses the interview.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 나주경찰서장 beside the camera.\n- KEY BACKGROUND ELEMENTS: corridor toward the violent-crimes office (Open as his intended path); used as The corridor direction behind 전택수 preserves the route he intends to continue along; soft shoulder edge (Mostly excluded from frame); used as A minimal foreground occlusion establishes the conversational side without introducing a second full subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent natural daytime corridor illumination maintains a restrained neutral-to-cool procedural mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor, stairwell background, daylight, and all three people's clothing. Exclude the chief's pointing gesture and frame the senior officer responding firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 문옆 표지판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나주경찰서장과 30대 한국인 남성 기자를 향해 단호한 눈빛으로 고개를 한쪽으로 비튼 전택수의 상체\n\nLOCATION (lock): In the corridor immediately outside the violent-crimes office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height just beyond the shared shoulder line of 나주경찰서장 and the reporter, finish the restrained dolly-in on 전택수's three-quarter upper body, leaving only a soft sliver of the intervening shoulder at the frame edge. 전택수 twists his head back over a body still angled toward the office, chin canted to one side and eyes locked firmly on 나주경찰서장 as he refuses the interview.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, midground, looks toward 나주경찰서장 beside the camera.\n- KEY BACKGROUND ELEMENTS: corridor toward the violent-crimes office (Open as his intended path); used as The corridor direction behind 전택수 preserves the route he intends to continue along; soft shoulder edge (Mostly excluded from frame); used as A minimal foreground occlusion establishes the conversational side without introducing a second full subject.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent natural daytime corridor illumination maintains a restrained neutral-to-cool procedural mood.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor, stairwell background, daylight, and all three people's clothing. Exclude the chief's pointing gesture and frame the senior officer responding firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 사무실 문옆 표지판: \"강력계\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "메인 인물의 시선이 화면 왼쪽 전경에 위치한 상대방의 어깨를 정확히 향함.",
    "built_space": "복도, 계단, 벽면 표지판이 카메라 시점에 맞게 자연스럽게 배치됨.",
    "entities": "표지판의 '강력계' 텍스트가 일치함. 단, 주인공이 레퍼런스의 사복 대신 경찰 제복을 입고 있음.",
    "hard_violations": [],
    "physics": "인물의 서 있는 자세가 바닥에 안정적으로 지지되어 물리적 모순이 없음."
   },
   {
    "label": "B",
    "direction": "인물의 시선이 화면 왼쪽 전경의 대화 상대를 향함.",
    "built_space": "복도와 계단 구조가 자연스럽고 표지판 위치가 일관됨.",
    "entities": "표지판 텍스트가 일치하나 인물의 외형 및 의상이 캐릭터 레퍼런스와 다름.",
    "hard_violations": [
     "물리적으로 불가능한 아티팩트 (화면 왼쪽 전경 어깨 부근 허공에 떠 있는 날카롭고 투명한 사선)"
    ],
    "physics": "화면 왼쪽에 어떤 지지대도 없는 정체불명의 투명한 사선이 허공에 물리적으로 불가능하게 떠 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 6,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "지정된 구도와 샷 스케일을 잘 구현했으나, 캐릭터 레퍼런스 대신 이전 컷의 인물 외형과 의상을 가져온 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "구도와 배경은 적절하나, 화면 왼쪽 전경에 물리적으로 불가능한 아티팩트가 떠 있어 치명적인 오류로 작용했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S51sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "주요 인물이 캐릭터 레퍼런스로 지정된 '전택수'(남색 수사 재킷, 체크 셔츠 착용)가 아닌 이전 컷에 등장했던 경찰서장(단기팔 제복 착용)의 얼굴과 복장으로 렌더링되었습니다.",
     "fix_en": "Redraw the central figure as Jeon Taek-su from the character reference, with his square face, wearing a navy blue investigative jacket over a plaid shirt. Preserve his exact pose, the background staircase, the corridor, the wall sign, and the foreground elements.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "왼쪽 전경에 다른 인물의 뒤통수와 어깨가 크게 들어와 허용된 어깨 슬iver가 아니라 두 번째 인물이 되었다",
     "fix_en": "Shrink the left foreground occlusion to a soft, narrow shoulder sliver, filling the revealed space with the plain white corridor wall. Preserve the central character, his clothing, his pose, and the background staircase.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "주요 인물이 캐릭터 레퍼런스로 지정된 '전택수'(남색 수사 재킷, 체크 셔츠 착용)가 아닌 이전 컷에 등장했던 경찰서장(단기팔 제복 착용)의 얼굴과 복장으로 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 중앙 인물이 전택수 레퍼런스와 다른 얼굴·머리·체형이며 이전 스틸의 제복 경찰 얼굴을 그대로 썼다",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수가 레퍼런스의 수사 점퍼·체크셔츠·모자가 아니라 이전 컷과 같은 경찰 정복을 입고 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 전경에 다른 인물의 뒤통수와 어깨가 크게 들어와 허용된 어깨 슬iver가 아니라 두 번째 인물이 되었다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 전경에 다른 인물의 뒤통수와 어깨가 크게 들어와 허용된 어깨 슬iver가 아니라 두 번째 인물이 되었다",
    "fix_en": "Shrink the left foreground occlusion to a soft, narrow shoulder sliver, filling the revealed space with the plain white corridor wall. Preserve the central character, his clothing, his pose, and the background staircase.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the central figure as Jeon Taek-su from the character reference, with his square face, wearing a navy blue investigative jacket over a plaid shirt. Preserve his exact pose, the background staircase, the corridor, the wall sign, and the foreground elements.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1150,
      "verdict_ko": "고개를 돌리는 핵심 포즈와 구도는 텍스트에 맞게 정확히 연출되었으나, 레퍼런스의 전택수가 아닌 이전 샷 경찰서장의 얼굴과 제복이 그대로 렌더링되어 신원 기준을 심각하게 위반함.  ★위반: [openrouter:x-ai/grok-4.6] 전택수 자리에 이전 스틸 제복 경찰을 얼굴·헤어·의상까지 이식한 인물 교체"
     },
     {
      "label": "B",
      "score": 1800,
      "verdict_ko": "복장은 레퍼런스의 점퍼로 수정되었으나 얼굴이 여전히 경찰서장이며, 몸을 틀고 고개를 돌리는 핵심 포즈 지시를 잃고 정면을 향해 서 있어 연출과 신원 모두 실패함."
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.4,
      "B": 1.8
     },
     "adjusted": {
      "A": 1.15,
      "B": 1.8
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 전택수 자리에 이전 스틸 제복 경찰을 얼굴·헤어·의상까지 이식한 인물 교체"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.6,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1150,
      "verdict_ko": "고개를 돌리는 핵심 포즈와 구도는 텍스트에 맞게 정확히 연출되었으나, 레퍼런스의 전택수가 아닌 이전 샷 경찰서장의 얼굴과 제복이 그대로 렌더링되어 신원 기준을 심각하게 위반함.  ★위반: [openrouter:x-ai/grok-4.6] 전택수 자리에 이전 스틸 제복 경찰을 얼굴·헤어·의상까지 이식한 인물 교체"
     },
     {
      "label": "B",
      "score": 1800,
      "verdict_ko": "복장은 레퍼런스의 점퍼로 수정되었으나 얼굴이 여전히 경찰서장이며, 몸을 틀고 고개를 돌리는 핵심 포즈 지시를 잃고 정면을 향해 서 있어 연출과 신원 모두 실패함."
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "캐릭터 레퍼런스의 의상은 정확히 반영했으나, 얼굴을 이전 컷 스틸의 인물로 잘못 생성하여 치명적인 신원 위반이 발생함."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "캐릭터 레퍼런스의 얼굴과 의상을 모두 무시하고 이전 컷 스틸의 인물을 그대로 복사하여 심각한 신원 위반으로 탈락함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수가 프레임 좌측 전경에 걸친 어깨(경찰서장)를 향해 시선을 고정하고 있음.",
      "built_space": "이전 컷과 동일한 복도, 우측 계단, 벽면의 '강력계' 표지판이 정확히 렌더링됨.",
      "entities": "의상(남색 재킷, 체크 셔츠)은 캐릭터 레퍼런스와 일치하나, 얼굴이 레퍼런스가 아닌 이전 컷 스틸의 인물임. 표지판 텍스트 '강력계'는 일치함.",
      "hard_violations": [
       "등장해서는 안 되는 이전 컷 스틸 인물의 얼굴을 그대로 차용하여 캐릭터 레퍼런스의 신원을 위반함"
      ],
      "physics": "바닥에 안정적으로 서 있으며, 자세나 구조에 물리적 오류가 없음."
     },
     {
      "label": "B",
      "direction": "몸을 돌린 상태에서 프레임 좌측 전경의 어깨(경찰서장)를 향해 시선을 던지고 있음.",
      "built_space": "이전 컷과 동일한 복도 및 우측 계단, 벽면의 '강력계' 표지판이 위치함.",
      "entities": "전택수의 얼굴과 의상(경찰 제복) 모두 캐릭터 레퍼런스를 무시하고 이전 컷 스틸의 인물을 그대로 복사함.",
      "hard_violations": [
       "등장해서는 안 되는 이전 컷 스틸 인물의 얼굴과 의상을 그대로 차용하여 캐릭터 레퍼런스의 신원을 위반함"
      ],
      "physics": "자연스럽게 서 있는 자세로 물리적 오류나 지지되지 않은 객체는 없음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "캐릭터 레퍼런스의 의상은 정확히 반영했으나, 얼굴을 이전 컷 스틸의 인물로 잘못 생성하여 치명적인 신원 위반이 발생함."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "캐릭터 레퍼런스의 얼굴과 의상을 모두 무시하고 이전 컷 스틸의 인물을 그대로 복사하여 심각한 신원 위반으로 탈락함."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "전택수가 프레임 좌측 전경에 걸친 어깨(경찰서장)를 향해 시선을 고정하고 있음.",
      "built_space": "이전 컷과 동일한 복도, 우측 계단, 벽면의 '강력계' 표지판이 정확히 렌더링됨.",
      "entities": "의상(남색 재킷, 체크 셔츠)은 캐릭터 레퍼런스와 일치하나, 얼굴이 레퍼런스가 아닌 이전 컷 스틸의 인물임. 표지판 텍스트 '강력계'는 일치함.",
      "hard_violations": [
       "등장해서는 안 되는 이전 컷 스틸 인물의 얼굴을 그대로 차용하여 캐릭터 레퍼런스의 신원을 위반함"
      ],
      "physics": "바닥에 안정적으로 서 있으며, 자세나 구조에 물리적 오류가 없음."
     },
     {
      "label": "A",
      "direction": "몸을 돌린 상태에서 프레임 좌측 전경의 어깨(경찰서장)를 향해 시선을 던지고 있음.",
      "built_space": "이전 컷과 동일한 복도 및 우측 계단, 벽면의 '강력계' 표지판이 위치함.",
      "entities": "전택수의 얼굴과 의상(경찰 제복) 모두 캐릭터 레퍼런스를 무시하고 이전 컷 스틸의 인물을 그대로 복사함.",
      "hard_violations": [
       "등장해서는 안 되는 이전 컷 스틸 인물의 얼굴과 의상을 그대로 차용하여 캐릭터 레퍼런스의 신원을 위반함"
      ],
      "physics": "자연스럽게 서 있는 자세로 물리적 오류나 지지되지 않은 객체는 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1152,
     "B": 1804
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S51sh4"
  }
 },
 "S51sh7::cine": {
  "applied": true,
  "fingerprint": "5b657f212d93e5d56867df69d3674e7328e9af664a47de6d53aefe0f21b4846d",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S51sh7_sel.png",
  "source_sha256": "c0a8cabbcfe139ddfa2978a2b2ec6bb133374ffc0cbbfe634567aa97033a6674",
  "file": "S51sh7_cine.png",
  "latency_ms": 10660
 },
 "S52sh1::signage": {
  "fp": "95e5690921c7d06a",
  "inscriptions": [
   {
    "surface_native": "TV 화면 자막",
    "text_native": "뉴스 특보",
    "reason_ko": "켜진 TV 화면에 시대적 배경과 현장감을 더하기 위해 한국 방송 자막을 재현합니다."
   }
  ]
 },
 "S52sh1": {
  "input_fingerprint": "4ac3155dbaed78a0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 소파에 앉아 켜진 TV 화면을 응시하는 이미경 너머로, 주방에서 커피포트를 기울인 채 서 있는 30대 초반 한국인 남성이 보이는 거실 전경\n\nLOCATION (lock): In the one-room apartment's combined living room and kitchenette, with the television switched on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From behind and slightly beside 이미경 at standing chest height, begin a slow push through a wide composition: her seated figure occupies the lower-left foreground, the active TV holds her sightline ahead, and 30대 초반 남성 appears in the kitchen at upper right. He tilts the coffee pot at the counter while looking down, creating a quiet background action against 이미경's increasingly absorbed stillness.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이미경 in the lower-left of the frame, foreground, looks toward television ahead; 30대 초반 남성 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: television screen (On and broadcasting the news) — The active display face is visible on 이미경's sightline and shows the landscape of the Deudeul River; used as Carries the news image that fixes 이미경's attention and establishes the narrative trigger; sofa (Occupied by 이미경) — Its back runs across the lower foreground with 이미경 seated toward the television; used as Supports 이미경 in the foreground and defines the route for the later move around its end; coffee pot (Held and tilted by 30대 초반 남성) — Its spout is angled downward over the kitchen counter; used as Provides the fiancé's restrained background action and keeps his attention divided from the news; kitchen area (Occupied by 30대 초반 남성); used as Separates the foreground living area from the fiancé's background position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime interior illumination is joined by the active television's restrained screen light, maintaining a neutral-to-cool and naturalistic range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 30대 초반 한국인 남성 right now — 30대 초반 한국인 남성's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 30대 초반 한국인 남성 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리); 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- TV 화면 자막: \"뉴스 특보\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 소파에 앉아 켜진 TV 화면을 응시하는 이미경 너머로, 주방에서 커피포트를 기울인 채 서 있는 30대 초반 한국인 남성이 보이는 거실 전경\n\nLOCATION (lock): In the one-room apartment's combined living room and kitchenette, with the television switched on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From behind and slightly beside 이미경 at standing chest height, begin a slow push through a wide composition: her seated figure occupies the lower-left foreground, the active TV holds her sightline ahead, and 30대 초반 남성 appears in the kitchen at upper right. He tilts the coffee pot at the counter while looking down, creating a quiet background action against 이미경's increasingly absorbed stillness.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이미경 in the lower-left of the frame, foreground, looks toward television ahead; 30대 초반 남성 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: television screen (On and broadcasting the news) — The active display face is visible on 이미경's sightline and shows the landscape of the Deudeul River; used as Carries the news image that fixes 이미경's attention and establishes the narrative trigger; sofa (Occupied by 이미경) — Its back runs across the lower foreground with 이미경 seated toward the television; used as Supports 이미경 in the foreground and defines the route for the later move around its end; coffee pot (Held and tilted by 30대 초반 남성) — Its spout is angled downward over the kitchen counter; used as Provides the fiancé's restrained background action and keeps his attention divided from the news; kitchen area (Occupied by 30대 초반 남성); used as Separates the foreground living area from the fiancé's background position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime interior illumination is joined by the active television's restrained screen light, maintaining a neutral-to-cool and naturalistic range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 30대 초반 한국인 남성 right now — 30대 초반 한국인 남성's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 30대 초반 한국인 남성 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리); 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- TV 화면 자막: \"뉴스 특보\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 소파에 앉아 켜진 TV 화면을 응시하는 이미경 너머로, 주방에서 커피포트를 기울인 채 서 있는 30대 초반 한국인 남성이 보이는 거실 전경\n\nLOCATION (lock): In the one-room apartment's combined living room and kitchenette, with the television switched on. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From behind and slightly beside 이미경 at standing chest height, begin a slow push through a wide composition: her seated figure occupies the lower-left foreground, the active TV holds her sightline ahead, and 30대 초반 남성 appears in the kitchen at upper right. He tilts the coffee pot at the counter while looking down, creating a quiet background action against 이미경's increasingly absorbed stillness.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 이미경 in the lower-left of the frame, foreground, looks toward television ahead; 30대 초반 남성 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: television screen (On and broadcasting the news) — The active display face is visible on 이미경's sightline and shows the landscape of the Deudeul River; used as Carries the news image that fixes 이미경's attention and establishes the narrative trigger; sofa (Occupied by 이미경) — Its back runs across the lower foreground with 이미경 seated toward the television; used as Supports 이미경 in the foreground and defines the route for the later move around its end; coffee pot (Held and tilted by 30대 초반 남성) — Its spout is angled downward over the kitchen counter; used as Provides the fiancé's restrained background action and keeps his attention divided from the news; kitchen area (Occupied by 30대 초반 남성); used as Separates the foreground living area from the fiancé's background position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime interior illumination is joined by the active television's restrained screen light, maintaining a neutral-to-cool and naturalistic range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 30대 초반 한국인 남성 right now — 30대 초반 한국인 남성's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 30대 초반 한국인 남성 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리); 30대 초반 남성 (Korean 남성, 30대 초반, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- TV 화면 자막: \"뉴스 특보\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.625,
    "B": 1.75
   },
   "adjusted": {
    "A": 1.625,
    "B": 1.75
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1750,
   "A": 1625
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1750,
    "verdict_ko": "이전 샷의 배경(창문, 흰색 냉장고)과 열린 주방 구조를 정확히 유지하며, 남성의 작업복 얼룩 등 레퍼런스를 매우 충실하게 반영했습니다."
   },
   {
    "label": "A",
    "score": 1625,
    "verdict_ko": "인물의 배치와 행동은 지시사항을 따랐으나, 이전 샷에 명확히 존재하던 냉장고가 사라지고 주방 사이에 가벽이 생기는 등 로케이션 일관성을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S31sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   },
   {
    "label": "CHARACTER REFERENCE — 30대 초반 남성: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:762177>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 샷 스틸에서 젖어 있던 이미경의 머리카락이 생성된 사진에서는 완전히 마르고 단정한 상태로 묘사되어 머리카락 상태 고정(LOCKED) 조건을 위반함.",
     "fix_en": "Render Lee Mi-kyung's hair as wet, stringy, and clumped together, matching the wet appearance established in the previous shot still. Preserve her current position, her clothing, the sofa, the television with its broadcast, the background man and his action, and all lighting and set details.",
     "severity": "major",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이전 샷 스틸에서 젖어 있던 이미경의 머리카락이 생성된 사진에서는 완전히 마르고 단정한 상태로 묘사되어 머리카락 상태 고정(LOCKED) 조건을 위반함.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 샷 스틸에서 젖어 있던 이미경의 머리카락이 생성된 사진에서는 완전히 마르고 단정한 상태로 묘사되어 머리카락 상태 고정(LOCKED) 조건을 위반함.",
    "fix_en": "Render Lee Mi-kyung's hair as wet, stringy, and clumped together, matching the wet appearance established in the previous shot still. Preserve her current position, her clothing, the sofa, the television with its broadcast, the background man and his action, and all lighting and set details.",
    "severity": "major",
    "observation_index": 0
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S31sh2"
  }
 },
 "S52sh1::cine": {
  "applied": true,
  "fingerprint": "603b2980ed94942fca5a073ed44876516f7256972548dddda0f2f1edc37d65cf",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S52sh1_sel.png",
  "source_sha256": "0b96b8b901a919f6d0a8587285765d3ee2620022880ecb50bfec6f7149530a28",
  "file": "S52sh1_cine.png",
  "latency_ms": 10736
 },
 "S52sh5::signage": {
  "fp": "985e5973f95787e6",
  "inscriptions": []
 },
 "S52sh5": {
  "input_fingerprint": "031c12a9ce71c5a4",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 화면을 향해 목을 뻗은 채 동공이 미세하게 확장된 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): At the sofa in the apartment's living area, facing the switched-on television. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the sofa on 이미경's television-facing side, complete the push into a slightly elevated, off-axis facial close-up rather than aligning directly with the screen. Her neck extends subtly toward the television outside frame, pupils minutely widened and facial muscles held still, while a narrow lower-edge trace of the sofa keeps the close view physically grounded.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward television outside frame.\n- KEY BACKGROUND ELEMENTS: sofa edge (Supporting 이미경) — The upper edge passes beneath 이미경's near shoulder; used as A small lower-frame reference preserves her seated position without competing with the facial reaction; television screen (On and broadcasting the case report) — Its active display face is directed toward 이미경 outside the camera frame and continues the news broadcast; used as Kept outside frame but motivates her fixed eyeline and minute reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The active television lends restrained screen illumination to the otherwise subdued nighttime interior, keeping skin tones natural within the neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same compact living room, active television glow, kitchen background, and night lighting. Exclude the fiancé from the close frame and isolate the woman's fixed reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 화면을 향해 목을 뻗은 채 동공이 미세하게 확장된 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): At the sofa in the apartment's living area, facing the switched-on television. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the sofa on 이미경's television-facing side, complete the push into a slightly elevated, off-axis facial close-up rather than aligning directly with the screen. Her neck extends subtly toward the television outside frame, pupils minutely widened and facial muscles held still, while a narrow lower-edge trace of the sofa keeps the close view physically grounded.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward television outside frame.\n- KEY BACKGROUND ELEMENTS: sofa edge (Supporting 이미경) — The upper edge passes beneath 이미경's near shoulder; used as A small lower-frame reference preserves her seated position without competing with the facial reaction; television screen (On and broadcasting the case report) — Its active display face is directed toward 이미경 outside the camera frame and continues the news broadcast; used as Kept outside frame but motivates her fixed eyeline and minute reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The active television lends restrained screen illumination to the otherwise subdued nighttime interior, keeping skin tones natural within the neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same compact living room, active television glow, kitchen background, and night lighting. Exclude the fiancé from the close frame and isolate the woman's fixed reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 화면을 향해 목을 뻗은 채 동공이 미세하게 확장된 이미경의 굳은 얼굴 클로즈업\n\nLOCATION (lock): At the sofa in the apartment's living area, facing the switched-on television. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the sofa on 이미경's television-facing side, complete the push into a slightly elevated, off-axis facial close-up rather than aligning directly with the screen. Her neck extends subtly toward the television outside frame, pupils minutely widened and facial muscles held still, while a narrow lower-edge trace of the sofa keeps the close view physically grounded.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward television outside frame.\n- KEY BACKGROUND ELEMENTS: sofa edge (Supporting 이미경) — The upper edge passes beneath 이미경's near shoulder; used as A small lower-frame reference preserves her seated position without competing with the facial reaction; television screen (On and broadcasting the case report) — Its active display face is directed toward 이미경 outside the camera frame and continues the news broadcast; used as Kept outside frame but motivates her fixed eyeline and minute reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The active television lends restrained screen illumination to the otherwise subdued nighttime interior, keeping skin tones natural within the neutral-to-cool palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same compact living room, active television glow, kitchen background, and night lighting. Exclude the fiancé from the close frame and isolate the woman's fixed reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.375,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.625,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.625,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1625,
   "B": 1875
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1625,
    "verdict_ko": "목을 화면 쪽으로 뻗은 긴장된 자세와 굳은 표정을 잘 구현했으며, TV 화면의 불빛이 얼굴에 자연스럽게 묘사되었습니다."
   },
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "인물과 표정 묘사는 양호하나, A에 비해 목을 앞으로 뻗은 느낌이나 긴장감이 다소 약합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S52sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라가 인물의 정면(TV 방향) 쪽에서 촬영하는 구도임에도 인물의 가슴 앞(화면 하단 전경)에 소파의 등받이나 팔걸이가 시야를 가로막듯 수평으로 배치되어 있어, 인물이 소파에 정상적으로 앉지 않고 거꾸로 앉거나 등받이 뒤에 서 있는 것처럼 물리적으로 불가능한 자세로 묘사되었습니다.",
     "fix_en": "Replace the horizontal sofa backrest running across the bottom foreground with Lee Mi-kyung's continuous lower torso, extending her pink cardigan and white blouse downward, so she appears seated facing the camera without front obstruction, preserving her face, hair, the kitchen background, and the nighttime lighting.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 중앙의 이미경이 목을 화면 쪽으로 뻗지 않고 고개를 곧게 세운 채 거의 정면을 향한다",
     "fix_en": "Extend Lee Mi-kyung's neck forward to show her reaching toward the off-screen television, preserving her face, hair, clothing, the sofa, and the background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "이미경의 얼굴이 굳은 표정이 아니라 부드러운 일상 표정이다",
     "fix_en": "Tense Lee Mi-kyung's facial muscles into a stiff, hardened expression with slightly widened pupils, preserving her identity, hair, clothing, posture, and the background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "이미경의 어깨와 상체가 렌즈에 정면으로 맞춰져 초상처럼 보인다",
     "fix_en": "Angle Lee Mi-kyung's shoulders and torso slightly off-axis so she is not perfectly squared to the lens, preserving her face, hair, clothing, the sofa, and the background.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "이미경의 이목구비가 캐릭터 레퍼런스와 다른 사람이다",
     "fix_en": "Adjust Lee Mi-kyung's facial features to match the provided character reference, preserving her expression, hair, clothing, posture, and the background.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라가 인물의 정면(TV 방향) 쪽에서 촬영하는 구도임에도 인물의 가슴 앞(화면 하단 전경)에 소파의 등받이나 팔걸이가 시야를 가로막듯 수평으로 배치되어 있어, 인물이 소파에 정상적으로 앉지 않고 거꾸로 앉거나 등받이 뒤에 서 있는 것처럼 물리적으로 불가능한 자세로 묘사되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 중앙의 이미경이 목을 화면 쪽으로 뻗지 않고 고개를 곧게 세운 채 거의 정면을 향한다",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 얼굴이 굳은 표정이 아니라 부드러운 일상 표정이다",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 어깨와 상체가 렌즈에 정면으로 맞춰져 초상처럼 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 이목구비가 캐릭터 레퍼런스와 다른 사람이다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 중앙의 이미경이 목을 화면 쪽으로 뻗지 않고 고개를 곧게 세운 채 거의 정면을 향한다",
    "fix_en": "Extend Lee Mi-kyung's neck forward to show her reaching toward the off-screen television, preserving her face, hair, clothing, the sofa, and the background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "이미경의 얼굴이 굳은 표정이 아니라 부드러운 일상 표정이다",
    "fix_en": "Tense Lee Mi-kyung's facial muscles into a stiff, hardened expression with slightly widened pupils, preserving her identity, hair, clothing, posture, and the background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "이미경의 어깨와 상체가 렌즈에 정면으로 맞춰져 초상처럼 보인다",
    "fix_en": "Angle Lee Mi-kyung's shoulders and torso slightly off-axis so she is not perfectly squared to the lens, preserving her face, hair, clothing, the sofa, and the background.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "이미경의 이목구비가 캐릭터 레퍼런스와 다른 사람이다",
    "fix_en": "Adjust Lee Mi-kyung's facial features to match the provided character reference, preserving her expression, hair, clothing, posture, and the background.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the horizontal sofa backrest running across the bottom foreground with Lee Mi-kyung's continuous lower torso, extending her pink cardigan and white blouse downward, so she appears seated facing the camera without front obstruction, preserving her face, hair, the kitchen background, and the nighttime lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요청된 얼굴 클로즈업 샷 스케일을 정확히 구현했으며, 화면 하단에 소파 엣지를 좁게 배치하라는 지시도 충실히 따랐습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "명확한 클로즈업 지시에도 불구하고 화면을 과도하게 넓혀 상반신과 치마, 넓은 소파 면적을 보여준 점이 큰 감점 요인입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선이 화면 밖 왼쪽(TV 위치)을 향해 고정되어 있음.",
      "built_space": "화면 맨 아래쪽에 소파 윗부분이 좁게 걸쳐 있으며, 배경에 이전 샷의 창문과 주방이 아웃포커스로 보임.",
      "entities": "레퍼런스와 일치하는 외모의 이미경이 분홍색 가디건과 하얀 리본 블라우스를 입고 있음.",
      "hard_violations": [],
      "physics": "소파 엣지 아래로 몸이 자연스럽게 이어져 착석 상태를 지탱함."
     },
     {
      "label": "B",
      "direction": "시선이 화면 밖 왼쪽(TV 위치)을 향함.",
      "built_space": "화면 하단의 절반 가까이를 소파가 차지하고 있으며, 배경 구조는 이전 샷과 일치함.",
      "entities": "이미경 본인이며, 가디건, 블라우스, 민트색 치마까지 착장이 모두 확인됨.",
      "hard_violations": [],
      "physics": "소파 위에 앉아 무릎에 손을 모으고 몸을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요청된 얼굴 클로즈업 샷 스케일을 정확히 구현했으며, 화면 하단에 소파 엣지를 좁게 배치하라는 지시도 충실히 따랐습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "명확한 클로즈업 지시에도 불구하고 화면을 과도하게 넓혀 상반신과 치마, 넓은 소파 면적을 보여준 점이 큰 감점 요인입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 화면 밖 왼쪽(TV 위치)을 향해 고정되어 있음.",
      "built_space": "화면 맨 아래쪽에 소파 윗부분이 좁게 걸쳐 있으며, 배경에 이전 샷의 창문과 주방이 아웃포커스로 보임.",
      "entities": "레퍼런스와 일치하는 외모의 이미경이 분홍색 가디건과 하얀 리본 블라우스를 입고 있음.",
      "hard_violations": [],
      "physics": "소파 엣지 아래로 몸이 자연스럽게 이어져 착석 상태를 지탱함."
     },
     {
      "label": "B",
      "direction": "시선이 화면 밖 왼쪽(TV 위치)을 향함.",
      "built_space": "화면 하단의 절반 가까이를 소파가 차지하고 있으며, 배경 구조는 이전 샷과 일치함.",
      "entities": "이미경 본인이며, 가디건, 블라우스, 민트색 치마까지 착장이 모두 확인됨.",
      "hard_violations": [],
      "physics": "소파 위에 앉아 무릎에 손을 모으고 몸을 지탱하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프롬프트가 요구한 '얼굴 클로즈업'과 '소파의 좁은 하단 가장자리' 프레이밍을 정확히 구현했으며, 화면을 향해 목을 뻗은 경직된 표정 연기도 훌륭하게 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "캐릭터와 배경의 일치도는 높으나, 프레이밍이 클로즈업이 아닌 미디엄 샷에 가깝게 넓어져 프롬프트의 지시를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선은 프레임 밖 왼쪽(TV가 있는 방향)을 향하고 있으며, 동공이 약간 확장된 상태로 집중하고 있음.",
      "built_space": "이전 샷과 동일한 거실 배경(왼쪽 램프, 창문, 오른쪽 주방)이 아웃포커싱되어 보이며, 인물 하단에 소파 등받이 윗부분이 좁게 걸쳐 있음.",
      "entities": "이미경(30대 초반 한국인 여성, 검은색 중간 길이 머리, 핑크 카디건, 흰색 리본 블라우스)이 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "인물의 자세와 프레임 하단의 소파 가장자리가 자연스럽게 맞닿아 있어 앉아있는 상태를 물리적으로 잘 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "시선은 프레임 밖 왼쪽(TV 방향)을 향하고 있음.",
      "built_space": "이전 샷의 거실과 주방 구조가 배경에 동일하게 배치되어 있으며, 인물은 소파 중앙에 위치함.",
      "entities": "이미경의 외모와 의상(핑크 카디건, 흰색 리본 블라우스, 민트색 치마)이 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "소파 위에 앉아 허벅지와 손목 부근까지 드러나며 자세가 안정적으로 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트가 요구한 '얼굴 클로즈업'과 '소파의 좁은 하단 가장자리' 프레이밍을 정확히 구현했으며, 화면을 향해 목을 뻗은 경직된 표정 연기도 훌륭하게 표현되었습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "캐릭터와 배경의 일치도는 높으나, 프레이밍이 클로즈업이 아닌 미디엄 샷에 가깝게 넓어져 프롬프트의 지시를 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 프레임 밖 왼쪽(TV가 있는 방향)을 향하고 있으며, 동공이 약간 확장된 상태로 집중하고 있음.",
      "built_space": "이전 샷과 동일한 거실 배경(왼쪽 램프, 창문, 오른쪽 주방)이 아웃포커싱되어 보이며, 인물 하단에 소파 등받이 윗부분이 좁게 걸쳐 있음.",
      "entities": "이미경(30대 초반 한국인 여성, 검은색 중간 길이 머리, 핑크 카디건, 흰색 리본 블라우스)이 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "인물의 자세와 프레임 하단의 소파 가장자리가 자연스럽게 맞닿아 있어 앉아있는 상태를 물리적으로 잘 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "시선은 프레임 밖 왼쪽(TV 방향)을 향하고 있음.",
      "built_space": "이전 샷의 거실과 주방 구조가 배경에 동일하게 배치되어 있으며, 인물은 소파 중앙에 위치함.",
      "entities": "이미경의 외모와 의상(핑크 카디건, 흰색 리본 블라우스, 민트색 치마)이 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "소파 위에 앉아 허벅지와 손목 부근까지 드러나며 자세가 안정적으로 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 8
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S52sh1"
  }
 },
 "S52sh5::cine": {
  "applied": true,
  "fingerprint": "f0ce81d30bfa06fdb7324769227594ec26d0f5d9378f1197c64b3cf564e72578",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S52sh5_sel.png",
  "source_sha256": "00265775d214e675f4a081a02e52729b0f26a3bf3847eae256b0dc24ac6a9b52",
  "file": "S52sh5_cine.png",
  "latency_ms": 10588
 },
 "S52sh8::signage": {
  "fp": "306e447003687397",
  "inscriptions": []
 },
 "S52sh8": {
  "input_fingerprint": "965715eeac412168",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 30대 초반 한국인 남성을 향해 단호한 표정으로 턱을 치켜든 이미경의 상체\n\nLOCATION (lock): In the living area of the one-room apartment, between the sofa and adjoining kitchenette. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 30대 초반 남성's side at 이미경's shoulder height, hold a static oblique three-quarter medium close view with only a soft fragment of his shoulder near the frame edge. 이미경 sits more firmly into the sofa, raises her chin into the available headroom, and directs an unwavering look toward him as she tells him to leave.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: sofa (Occupied by 이미경) — The seat and back align beneath and behind 이미경 as she addresses her fiancé; used as Anchors 이미경's seated refusal and keeps her upper-body posture legible; fiancé shoulder edge (Mostly excluded from frame); used as A minimal foreground trace establishes the fiancé as her addressee without turning the shot into a full two-shot; television screen (On and continuing the broadcast) — The display face is turned away from the camera angle and is no longer the visual focus, though the news remains on; used as Remains peripheral context for the news that prompted her emotional withdrawal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime interior illumination and restrained television spill preserve the same neutral-to-cool realism as her attention shifts to her fiancé.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same living-room layout, television light, kitchen area, and both characters' clothing. Exclude the coffee-pouring action and frame the woman addressing the man firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 30대 초반 한국인 남성을 향해 단호한 표정으로 턱을 치켜든 이미경의 상체\n\nLOCATION (lock): In the living area of the one-room apartment, between the sofa and adjoining kitchenette. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 30대 초반 남성's side at 이미경's shoulder height, hold a static oblique three-quarter medium close view with only a soft fragment of his shoulder near the frame edge. 이미경 sits more firmly into the sofa, raises her chin into the available headroom, and directs an unwavering look toward him as she tells him to leave.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: sofa (Occupied by 이미경) — The seat and back align beneath and behind 이미경 as she addresses her fiancé; used as Anchors 이미경's seated refusal and keeps her upper-body posture legible; fiancé shoulder edge (Mostly excluded from frame); used as A minimal foreground trace establishes the fiancé as her addressee without turning the shot into a full two-shot; television screen (On and continuing the broadcast) — The display face is turned away from the camera angle and is no longer the visual focus, though the news remains on; used as Remains peripheral context for the news that prompted her emotional withdrawal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime interior illumination and restrained television spill preserve the same neutral-to-cool realism as her attention shifts to her fiancé.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same living-room layout, television light, kitchen area, and both characters' clothing. Exclude the coffee-pouring action and frame the woman addressing the man firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 30대 초반 한국인 남성을 향해 단호한 표정으로 턱을 치켜든 이미경의 상체\n\nLOCATION (lock): In the living area of the one-room apartment, between the sofa and adjoining kitchenette. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 30대 초반 남성's side at 이미경's shoulder height, hold a static oblique three-quarter medium close view with only a soft fragment of his shoulder near the frame edge. 이미경 sits more firmly into the sofa, raises her chin into the available headroom, and directs an unwavering look toward him as she tells him to leave.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: sofa (Occupied by 이미경) — The seat and back align beneath and behind 이미경 as she addresses her fiancé; used as Anchors 이미경's seated refusal and keeps her upper-body posture legible; fiancé shoulder edge (Mostly excluded from frame); used as A minimal foreground trace establishes the fiancé as her addressee without turning the shot into a full two-shot; television screen (On and continuing the broadcast) — The display face is turned away from the camera angle and is no longer the visual focus, though the news remains on; used as Remains peripheral context for the news that prompted her emotional withdrawal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime interior illumination and restrained television spill preserve the same neutral-to-cool realism as her attention shifts to her fiancé.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same living-room layout, television light, kitchen area, and both characters' clothing. Exclude the coffee-pouring action and frame the woman addressing the man firmly.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "이미경은 왼쪽 전경에 위치한 남성의 어깨 쪽을 향해 단호한 시선을 보내고 있습니다.",
    "built_space": "소파에 인물이 앉아 있으며, 왼쪽 배경에는 창문과 장식장(액자와 화분 유지됨)이, 오른쪽에는 주방과 측면으로 비껴간 TV 화면이 위치합니다.",
    "entities": "이미경의 얼굴, 헤어스타일, 의상(분홍색 가디건, 리본 블라우스)이 레퍼런스와 정확히 일치하며, 왼쪽 가장자리에 30대 남성의 어깨가 보입니다.",
    "hard_violations": [],
    "physics": "이미경은 소파에 안정적으로 앉아 체중을 지탱하고 있습니다."
   },
   {
    "label": "A",
    "direction": "이미경은 왼쪽 전경의 남성을 향해 시선을 고정하고 있습니다.",
    "built_space": "소파 위 인물 뒤로 창문과 주방이 보이나, 왼쪽 장식장 위에 TV가 정면을 향해 배치되어 공간의 연속성을 깨뜨렸습니다.",
    "entities": "이미경의 외양과 의상이 레퍼런스와 일치하며, 전경에 남성의 뒷모습 일부가 존재합니다.",
    "hard_violations": [
     "이전 컷에서 고정된 배경 요소(창문 앞 장식장의 액자와 화분)가 있어야 할 위치에 TV를 임의로 생성함"
    ],
    "physics": "인물들은 소파와 바닥을 통해 정상적으로 지지받고 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 8,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "이전 컷의 배경 요소(액자, 화분 등)를 완벽하게 유지하면서, 프롬프트의 요구대로 카메라를 비껴간 TV와 인물의 구도를 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "이전 컷에 있던 배경 요소(장식장 위 액자와 화분)를 무시하고 그 자리에 정면을 향한 TV를 배치하여 위치 고정 및 TV 방향 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S52sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프레임 가장자리에 남성의 어깨 일부만 부드럽게 걸쳐져야 한다는 카메라 프레이밍 지시와 달리, 프레임 왼쪽 화면의 3분의 1가량을 남성의 뒷모습(머리와 어깨 전체)이 크게 차지하고 있습니다.",
     "fix_en": "Reduce the man in the left foreground to only a minimal, out-of-focus sliver of his right shoulder at the very edge of the frame, revealing the background wall and window behind where his head currently is; maintain the woman, her clothing, the lighting, and the rest of the room unchanged.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "단호한 표정으로 턱을 치켜들어야 한다는 액션 지시와 달리, 여성의 턱이 들려 있지 않으며 표정 또한 단호하기보다는 다소 슬프거나 무기력해 보입니다.",
     "fix_en": "Raise the woman's chin and adjust her facial expression to be firm and resolute, maintaining her exact identity, clothing, the foreground man's shoulder, the lighting, and the background unchanged.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "오른쪽 TV 화면이 카메라를 향해 보여 화면이 외면되어야 한다는 지시를 어김",
     "fix_en": "Rotate the television set in the right background so its screen faces away from the camera, leaving only its back or side visible, while keeping the woman's pose, the foreground man, lighting, and all other background elements unchanged.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프레임 가장자리에 남성의 어깨 일부만 부드럽게 걸쳐져야 한다는 카메라 프레이밍 지시와 달리, 프레임 왼쪽 화면의 3분의 1가량을 남성의 뒷모습(머리와 어깨 전체)이 크게 차지하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "단호한 표정으로 턱을 치켜들어야 한다는 액션 지시와 달리, 여성의 턱이 들려 있지 않으며 표정 또한 단호하기보다는 다소 슬프거나 무기력해 보입니다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 전경에 남성의 머리·목·등이 크게 들어와 어깨 가장자리만 남긴 최소 전경이 아니라 오버숄더 투샷이 됨",
     "severity": "major"
    },
    {
     "issue_ko": "이미경이 턱을 치켜들지 않았고 표정이 단호하지 않음",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 TV 화면이 카메라를 향해 보여 화면이 외면되어야 한다는 지시를 어김",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "프레임 가장자리에 남성의 어깨 일부만 부드럽게 걸쳐져야 한다는 카메라 프레이밍 지시와 달리, 프레임 왼쪽 화면의 3분의 1가량을 남성의 뒷모습(머리와 어깨 전체)이 크게 차지하고 있습니다.",
    "fix_en": "Reduce the man in the left foreground to only a minimal, out-of-focus sliver of his right shoulder at the very edge of the frame, revealing the background wall and window behind where his head currently is; maintain the woman, her clothing, the lighting, and the rest of the room unchanged.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "단호한 표정으로 턱을 치켜들어야 한다는 액션 지시와 달리, 여성의 턱이 들려 있지 않으며 표정 또한 단호하기보다는 다소 슬프거나 무기력해 보입니다.",
    "fix_en": "Raise the woman's chin and adjust her facial expression to be firm and resolute, maintaining her exact identity, clothing, the foreground man's shoulder, the lighting, and the background unchanged.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "오른쪽 TV 화면이 카메라를 향해 보여 화면이 외면되어야 한다는 지시를 어김",
    "fix_en": "Rotate the television set in the right background so its screen faces away from the camera, leaving only its back or side visible, while keeping the woman's pose, the foreground man, lighting, and all other background elements unchanged.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S52sh5"
  }
 },
 "S52sh8::cine": {
  "applied": true,
  "fingerprint": "8dd91d68c8105241d326a25d04cfc95b59a87034904e6e667a3d4aebb992a7b5",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S52sh8_sel.png",
  "source_sha256": "1847e4713a422582ce5b937990f8ccfd7b4c3e7ee78eb935c0db98eb1867ae0e",
  "file": "S52sh8_cine.png",
  "latency_ms": 10203
 },
 "S53sh5::signage": {
  "fp": "51662b3a16a3f582",
  "inscriptions": []
 },
 "era_assess::169b1f2aca8fb63d": {
  "subjects": [
   {
    "subject_native": "대한민국 검찰 조사실 (2000년대~2010년대)",
    "search_terms_native": [
     "검찰 조사실",
     "피의자 신문실",
     "검찰청 조사실 내부",
     "지검 조사실"
    ],
    "language_lock_native": "이 검색어는 반드시 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "서구식 취조실(일방향 거울과 철제 탁자)과 달리, 한국의 검찰 조사실은 컴퓨터, 산더미 같은 서류 상자, 캐비닛, 녹화 장비가 혼재된 특유의 사무실 형태를 띠고 있어 고증이 필수적입니다."
   }
  ]
 },
 "era_ref::ce6d4cb4d01fa126": {
  "subject": "대한민국 검찰 조사실 (2000년대~2010년대)",
  "terms": [
   "검찰 조사실",
   "피의자 신문실",
   "검찰청 조사실 내부",
   "지검 조사실"
  ],
  "queries": [
   [
    "검찰 조사실 피의자 신문실 검찰청 조사실 내부 지검 조사실",
    "대한민국 검찰 조사실 2000년대 2010년대"
   ],
   [
    "검찰청 조사실 내부 2000년대",
    "지방검찰청 피의자 신문실 내부 2010년대"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://d216kn0te1eemj.cloudfront.net/post/1752394078110_23c4d604-ea4a-4e34-8186-1dcbfd4e45d2",
  "picked_reason_ko": "1번은 ‘조사실’ 표기가 있는 실제 한국식 조사실 내부를 주제로 삼아 책상·좌석·컴퓨터·벽체와 출입문의 구성 및 재료가 가장 명확하게 보인다.",
  "sha256": "0b108abe313a92c7b93c4f4b83b0fe419cc36877fed4b2dd7a6199a079d6d306",
  "file": "eraref_ce6d4cb4d01fa126.png"
 },
 "S53sh5::bgfirst_bg": {
  "input_fingerprint": "b7519710ea29d2b4",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 장원섭의 손에 들린 살짝 열린 종이봉투 안, 노란 구운 은행알들이 가득 담긴 모습 클로즈업\n\nLOCATION (lock): Over the table inside the prosecution interview room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just above table height beside 장원섭, the camera continues its restrained dolly-in at a downward angle, isolating his hand and the slightly opened envelope while excluding both faces. The envelope rests near center frame with its opening angled toward the lens; the yellow roasted ginkgo nuts remain visible inside, while the table provides scale around the detail.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: slightly opened paper envelope (Slightly open in 장원섭's hand) — The open mouth is angled upward and toward the camera, exposing the interior; used as Central focal object held just above the table, with the opening revealing the contents; roasted ginkgo nuts (Filling the opened envelope); used as Visible contents providing the shot's narrative detail; interrogation-room table (Between the two seated men); used as Surrounding surface that keeps the envelope at a realistic scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality with moderate contrast and naturalistic detail on the hand, envelope, and yellow ginkgo nuts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 검찰 조사실 (2000년대~2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 장원섭의 손에 들린 살짝 열린 종이봉투 안, 노란 구운 은행알들이 가득 담긴 모습 클로즈업\n\nLOCATION (lock): Over the table inside the prosecution interview room.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just above table height beside 장원섭, the camera continues its restrained dolly-in at a downward angle, isolating his hand and the slightly opened envelope while excluding both faces. The envelope rests near center frame with its opening angled toward the lens; the yellow roasted ginkgo nuts remain visible inside, while the table provides scale around the detail.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: slightly opened paper envelope (Slightly open in 장원섭's hand) — The open mouth is angled upward and toward the camera, exposing the interior; used as Central focal object held just above the table, with the opening revealing the contents; roasted ginkgo nuts (Filling the opened envelope); used as Visible contents providing the shot's narrative detail; interrogation-room table (Between the two seated men); used as Surrounding surface that keeps the envelope at a realistic scale.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality with moderate contrast and naturalistic detail on the hand, envelope, and yellow ginkgo nuts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 검찰 조사실 (2000년대~2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S53sh5__bgfirst_bg.png",
  "asset_id": "2bdfc18c-5ae9-4091-866a-c2f9e838c27b",
  "input_asset_ids": [
   "6971483d-9229-4dcf-a784-27e9f87d79f1",
   "9aead8ed-6c8f-4fb0-8065-f8c1744fe0c5"
  ],
  "era_research": {
   "subject": "대한민국 검찰 조사실 (2000년대~2010년대)",
   "queries": [
    [
     "검찰 조사실 피의자 신문실 검찰청 조사실 내부 지검 조사실",
     "대한민국 검찰 조사실 2000년대 2010년대"
    ],
    [
     "검찰청 조사실 내부 2000년대",
     "지방검찰청 피의자 신문실 내부 2010년대"
    ]
   ],
   "picked_url": "https://d216kn0te1eemj.cloudfront.net/post/1752394078110_23c4d604-ea4a-4e34-8186-1dcbfd4e45d2",
   "sha256": "0b108abe313a92c7b93c4f4b83b0fe419cc36877fed4b2dd7a6199a079d6d306",
   "file": "eraref_ce6d4cb4d01fa126.png"
  }
 },
 "S53sh5": {
  "input_fingerprint": "d96b838051cb37f8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 장원섭의 손에 들린 살짝 열린 종이봉투 안, 노란 구운 은행알들이 가득 담긴 모습 클로즈업\n\nLOCATION (lock): Over the table inside the prosecution interview room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above table height beside 장원섭, the camera continues its restrained dolly-in at a downward angle, isolating his hand and the slightly opened envelope while excluding both faces. The envelope rests near center frame with its opening angled toward the lens; the yellow roasted ginkgo nuts remain visible inside, while the table provides scale around the detail.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: slightly opened paper envelope (Slightly open in 장원섭's hand) — The open mouth is angled upward and toward the camera, exposing the interior; used as Central focal object held just above the table, with the opening revealing the contents; roasted ginkgo nuts (Filling the opened envelope); used as Visible contents providing the shot's narrative detail; interrogation-room table (Between the two seated men); used as Surrounding surface that keeps the envelope at a realistic scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality with moderate contrast and naturalistic detail on the hand, envelope, and yellow ginkgo nuts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The opened paper bag contains roasted ginkgo nuts and is being placed at the center of the interrogation table.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 장원섭의 손에 들린 살짝 열린 종이봉투 안, 노란 구운 은행알들이 가득 담긴 모습 클로즈업\n\nLOCATION (lock): Over the table inside the prosecution interview room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above table height beside 장원섭, the camera continues its restrained dolly-in at a downward angle, isolating his hand and the slightly opened envelope while excluding both faces. The envelope rests near center frame with its opening angled toward the lens; the yellow roasted ginkgo nuts remain visible inside, while the table provides scale around the detail.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: slightly opened paper envelope (Slightly open in 장원섭's hand) — The open mouth is angled upward and toward the camera, exposing the interior; used as Central focal object held just above the table, with the opening revealing the contents; roasted ginkgo nuts (Filling the opened envelope); used as Visible contents providing the shot's narrative detail; interrogation-room table (Between the two seated men); used as Surrounding surface that keeps the envelope at a realistic scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality with moderate contrast and naturalistic detail on the hand, envelope, and yellow ginkgo nuts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The opened paper bag contains roasted ginkgo nuts and is being placed at the center of the interrogation table.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 장원섭의 손에 들린 살짝 열린 종이봉투 안, 노란 구운 은행알들이 가득 담긴 모습 클로즈업\n\nLOCATION (lock): Over the table inside the prosecution interview room. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above table height beside 장원섭, the camera continues its restrained dolly-in at a downward angle, isolating his hand and the slightly opened envelope while excluding both faces. The envelope rests near center frame with its opening angled toward the lens; the yellow roasted ginkgo nuts remain visible inside, while the table provides scale around the detail.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: slightly opened paper envelope (Slightly open in 장원섭's hand) — The open mouth is angled upward and toward the camera, exposing the interior; used as Central focal object held just above the table, with the opening revealing the contents; roasted ginkgo nuts (Filling the opened envelope); used as Visible contents providing the shot's narrative detail; interrogation-room table (Between the two seated men); used as Surrounding surface that keeps the envelope at a realistic scale.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality with moderate contrast and naturalistic detail on the hand, envelope, and yellow ginkgo nuts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The opened paper bag contains roasted ginkgo nuts and is being placed at the center of the interrogation table.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S53sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S53sh5.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L54B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 인서트 클로즈업 앵글을 완벽히 구현했으며, 종이봉투의 구겨진 질감과 구운 은행알의 형태가 매우 사실적으로 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레이밍과 앵글은 지시사항을 따랐으나, 종이봉투가 인위적인 상자처럼 보이고 은행알의 표면이 지나치게 매끄러워 사실감이 떨어집니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "봉투의 입구가 위를 향하고 렌즈 쪽으로 살짝 기울어져 내부를 보여줍니다.",
      "built_space": "위치 레퍼런스와 일치하는 짙은 색 나무 테이블 위에 세팅되었으며, 배경에 창문과 의자가 보입니다.",
      "entities": "회색 정장 소매를 입은 손이 등장하며, 봉투는 종이 상자처럼 각지고 뻣뻣하게 묘사되었고, 은행알은 질감 없이 매끄러운 구 형태로 나타납니다.",
      "hard_violations": [],
      "physics": "손이 봉투의 측면을 잡고 있으며, 봉투 하단은 테이블 위에 놓여 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "봉투의 열린 입구가 카메라 렌즈 방향을 향해 기울어져 은행알이 명확하게 보입니다.",
      "built_space": "레퍼런스와 동일한 짙은 색 나무 테이블, 뒷배경의 창문 및 수납장이 공간의 맥락을 정확히 보여줍니다.",
      "entities": "레퍼런스와 일치하는 짙은 회색 정장 소매를 입은 손이 보이며, 현실적인 주름을 가진 종이봉투와 자연스러운 타원형태의 구운 은행알이 묘사되었습니다.",
      "hard_violations": [],
      "physics": "손이 위에서부터 봉투 측면을 감싸 쥐고 있으며, 봉투는 테이블 표면 위에 안정적으로 놓여 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 인서트 클로즈업 앵글을 완벽히 구현했으며, 종이봉투의 구겨진 질감과 구운 은행알의 형태가 매우 사실적으로 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "프레이밍과 앵글은 지시사항을 따랐으나, 종이봉투가 인위적인 상자처럼 보이고 은행알의 표면이 지나치게 매끄러워 사실감이 떨어집니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "봉투의 입구가 위를 향하고 렌즈 쪽으로 살짝 기울어져 내부를 보여줍니다.",
      "built_space": "위치 레퍼런스와 일치하는 짙은 색 나무 테이블 위에 세팅되었으며, 배경에 창문과 의자가 보입니다.",
      "entities": "회색 정장 소매를 입은 손이 등장하며, 봉투는 종이 상자처럼 각지고 뻣뻣하게 묘사되었고, 은행알은 질감 없이 매끄러운 구 형태로 나타납니다.",
      "hard_violations": [],
      "physics": "손이 봉투의 측면을 잡고 있으며, 봉투 하단은 테이블 위에 놓여 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "봉투의 열린 입구가 카메라 렌즈 방향을 향해 기울어져 은행알이 명확하게 보입니다.",
      "built_space": "레퍼런스와 동일한 짙은 색 나무 테이블, 뒷배경의 창문 및 수납장이 공간의 맥락을 정확히 보여줍니다.",
      "entities": "레퍼런스와 일치하는 짙은 회색 정장 소매를 입은 손이 보이며, 현실적인 주름을 가진 종이봉투와 자연스러운 타원형태의 구운 은행알이 묘사되었습니다.",
      "hard_violations": [],
      "physics": "손이 위에서부터 봉투 측면을 감싸 쥐고 있으며, 봉투는 테이블 표면 위에 안정적으로 놓여 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "카메라가 인물 옆에 위치하라는 지시를 잘 따랐으며, 구겨진 종이봉투의 질감과 열린 입구의 방향도 프롬프트와 일치합니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "카메라 앵글이 인물의 옆이 아닌 1인칭 시점에 가깝게 연출되었고, 종이봉투가 종이상자처럼 부자연스럽게 표현되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "종이봉투의 열린 입구가 카메라 렌즈 방향을 향하고 있음.",
      "built_space": "취조실 테이블 위. 배경에 기준 이미지와 일치하는 창문과 수납장이 있으며, 카메라는 인물의 측면에 위치함.",
      "entities": "어두운 정장과 하늘색 셔츠를 입은 손, 구겨진 종이봉투, 가득 담긴 노란 은행알들 모두 지시사항과 일치함.",
      "hard_violations": [],
      "physics": "손이 테이블에 놓인 종이봉투를 안정적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "종이봉투의 입구가 위쪽을 향해 열려 있음.",
      "built_space": "취조실 테이블 위. 카메라는 인물의 옆이 아닌 팔이 화면 하단에서 뻗어 나오는 1인칭 시점(POV) 구도를 취함.",
      "entities": "정장 소매와 손, 노란 은행알이 보이나, 종이봉투가 각진 종이상자의 형태로 묘사됨.",
      "hard_violations": [],
      "physics": "손이 테이블 위의 상자형 봉투 가장자리를 잡고 지탱함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "카메라가 인물 옆에 위치하라는 지시를 잘 따랐으며, 구겨진 종이봉투의 질감과 열린 입구의 방향도 프롬프트와 일치합니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "카메라 앵글이 인물의 옆이 아닌 1인칭 시점에 가깝게 연출되었고, 종이봉투가 종이상자처럼 부자연스럽게 표현되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "종이봉투의 열린 입구가 카메라 렌즈 방향을 향하고 있음.",
      "built_space": "취조실 테이블 위. 배경에 기준 이미지와 일치하는 창문과 수납장이 있으며, 카메라는 인물의 측면에 위치함.",
      "entities": "어두운 정장과 하늘색 셔츠를 입은 손, 구겨진 종이봉투, 가득 담긴 노란 은행알들 모두 지시사항과 일치함.",
      "hard_violations": [],
      "physics": "손이 테이블에 놓인 종이봉투를 안정적으로 쥐고 있음."
     },
     {
      "label": "A",
      "direction": "종이봉투의 입구가 위쪽을 향해 열려 있음.",
      "built_space": "취조실 테이블 위. 카메라는 인물의 옆이 아닌 팔이 화면 하단에서 뻗어 나오는 1인칭 시점(POV) 구도를 취함.",
      "entities": "정장 소매와 손, 노란 은행알이 보이나, 종이봉투가 각진 종이상자의 형태로 묘사됨.",
      "hard_violations": [],
      "physics": "손이 테이블 위의 상자형 봉투 가장자리를 잡고 지탱함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "봉투의 입구가 위를 향하고 렌즈 쪽으로 살짝 기울어져 내부를 보여줍니다.",
    "built_space": "위치 레퍼런스와 일치하는 짙은 색 나무 테이블 위에 세팅되었으며, 배경에 창문과 의자가 보입니다.",
    "entities": "회색 정장 소매를 입은 손이 등장하며, 봉투는 종이 상자처럼 각지고 뻣뻣하게 묘사되었고, 은행알은 질감 없이 매끄러운 구 형태로 나타납니다.",
    "hard_violations": [],
    "physics": "손이 봉투의 측면을 잡고 있으며, 봉투 하단은 테이블 위에 놓여 지탱되고 있습니다."
   },
   {
    "label": "B",
    "direction": "봉투의 열린 입구가 카메라 렌즈 방향을 향해 기울어져 은행알이 명확하게 보입니다.",
    "built_space": "레퍼런스와 동일한 짙은 색 나무 테이블, 뒷배경의 창문 및 수납장이 공간의 맥락을 정확히 보여줍니다.",
    "entities": "레퍼런스와 일치하는 짙은 회색 정장 소매를 입은 손이 보이며, 현실적인 주름을 가진 종이봉투와 자연스러운 타원형태의 구운 은행알이 묘사되었습니다.",
    "hard_violations": [],
    "physics": "손이 위에서부터 봉투 측면을 감싸 쥐고 있으며, 봉투는 테이블 표면 위에 안정적으로 놓여 있습니다."
   }
  ],
  "totals": {
   "A": 8,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 인서트 클로즈업 앵글을 완벽히 구현했으며, 종이봉투의 구겨진 질감과 구운 은행알의 형태가 매우 사실적으로 표현되었습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "프레이밍과 앵글은 지시사항을 따랐으나, 종이봉투가 인위적인 상자처럼 보이고 은행알의 표면이 지나치게 매끄러워 사실감이 떨어집니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L54B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "봉투 안의 내용물이 표면이 매끄러운 은행알이 아니라, 깊은 주름과 이음새를 가진 피스타치오나 호두 형태의 전혀 다른 견과류로 렌더링되었습니다.",
     "fix_en": "Replace the ridged nuts in the bag with smooth, oval yellow roasted ginkgo nuts. Preserve the hand, paper bag, suit, table, and background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "종이봉투를 쥐고 있는 손가락들(특히 아래쪽 두 손가락)이 봉투 표면 위에 자연스럽게 놓이지 않고 종이 재질과 뭉개져 융합되어 있습니다.",
     "fix_en": "Render clear boundaries for the fingertips resting on the bag so they do not blend into the paper. Preserve the hand pose, bag, nuts, table, and background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "손과 봉투만 고립한 인서트 클로즈업이 아니라 왼쪽 상반신과 뒤쪽 방 전체가 넓게 보인다",
     "fix_en": "Apply a slight vignette to darken the torso and background, focusing attention on the hand and bag. Preserve all existing elements.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 중앙 종이봉투가 살짝 열린 정도가 아니라 입구가 크게 벌어져 있다",
     "fix_en": "Narrow the paper bag's opening so it is only slightly open. Preserve the hand, nuts, table, and background.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "봉투 안의 내용물이 표면이 매끄러운 은행알이 아니라, 깊은 주름과 이음새를 가진 피스타치오나 호두 형태의 전혀 다른 견과류로 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "종이봉투를 쥐고 있는 손가락들(특히 아래쪽 두 손가락)이 봉투 표면 위에 자연스럽게 놓이지 않고 종이 재질과 뭉개져 융합되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "손과 봉투만 고립한 인서트 클로즈업이 아니라 왼쪽 상반신과 뒤쪽 방 전체가 넓게 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙 종이봉투가 살짝 열린 정도가 아니라 입구가 크게 벌어져 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "종이봉투를 쥐고 있는 손가락들(특히 아래쪽 두 손가락)이 봉투 표면 위에 자연스럽게 놓이지 않고 종이 재질과 뭉개져 융합되어 있습니다.",
    "fix_en": "Render clear boundaries for the fingertips resting on the bag so they do not blend into the paper. Preserve the hand pose, bag, nuts, table, and background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "손과 봉투만 고립한 인서트 클로즈업이 아니라 왼쪽 상반신과 뒤쪽 방 전체가 넓게 보인다",
    "fix_en": "Apply a slight vignette to darken the torso and background, focusing attention on the hand and bag. Preserve all existing elements.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "화면 중앙 종이봉투가 살짝 열린 정도가 아니라 입구가 크게 벌어져 있다",
    "fix_en": "Narrow the paper bag's opening so it is only slightly open. Preserve the hand, nuts, table, and background.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the ridged nuts in the bag with smooth, oval yellow roasted ginkgo nuts. Preserve the hand, paper bag, suit, table, and background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "요구된 인서트 클로즈업 앵글과 배경, 의상을 완벽히 구현했으며, 은행알의 표면이 매끄럽게 수정되어 실제 형태에 더 가깝습니다."
     },
     {
      "label": "A",
      "score": 1875,
      "verdict_ko": "지정된 구도와 로케이션, 의상은 잘 일치하나, 은행알의 표면에 부자연스러운 세로줄 질감이 있어 형태가 어색합니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.875,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.875,
      "B": 1.75
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "요구된 인서트 클로즈업 앵글과 배경, 의상을 완벽히 구현했으며, 은행알의 표면이 매끄럽게 수정되어 실제 형태에 더 가깝습니다."
     },
     {
      "label": "A",
      "score": 1875,
      "verdict_ko": "지정된 구도와 로케이션, 의상은 잘 일치하나, 은행알의 표면에 부자연스러운 세로줄 질감이 있어 형태가 어색합니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "구운 은행알 특유의 테두리와 형태를 현실적으로 묘사하여 질감과 디테일 요구사항을 훌륭하게 충족했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "구도와 배경은 정확하나, 봉투 안의 은행알이 너무 매끄러운 타원형으로 묘사되어 실제 사물로서의 디테일이 부족합니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "종이봉투의 열린 입구가 렌즈 방향을 향하고 있음.",
      "built_space": "배경에 취조실의 창문과 캐비닛이 보이며, 피사체 아래의 테이블 형태가 레퍼런스와 일치함.",
      "entities": "장원섭의 수트(짙은 회색)와 셔츠(연하늘색) 소매가 일치하며, 종이봉투 안의 내용물이 은행알의 형태(테두리 선 등)를 정확히 띠고 있음.",
      "hard_violations": [],
      "physics": "손이 봉투를 잡고 지탱하며, 봉투의 하단은 테이블 표면에 자연스럽게 놓여 있음."
     },
     {
      "label": "A",
      "direction": "종이봉투의 열린 입구가 렌즈 방향을 향하고 있음.",
      "built_space": "배경에 취조실의 창문과 캐비닛이 보이며, 피사체 아래의 테이블 형태가 레퍼런스와 일치함.",
      "entities": "장원섭의 수트와 셔츠 소매는 일치하나, 봉투 안의 은행알이 특징적인 테두리 없이 지나치게 매끄러운 형태를 띠고 있음.",
      "hard_violations": [],
      "physics": "손이 봉투를 잡고 지탱하며, 봉투의 하단은 테이블 표면에 자연스럽게 놓여 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구운 은행알 특유의 테두리와 형태를 현실적으로 묘사하여 질감과 디테일 요구사항을 훌륭하게 충족했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "구도와 배경은 정확하나, 봉투 안의 은행알이 너무 매끄러운 타원형으로 묘사되어 실제 사물로서의 디테일이 부족합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "종이봉투의 열린 입구가 렌즈 방향을 향하고 있음.",
      "built_space": "배경에 취조실의 창문과 캐비닛이 보이며, 피사체 아래의 테이블 형태가 레퍼런스와 일치함.",
      "entities": "장원섭의 수트(짙은 회색)와 셔츠(연하늘색) 소매가 일치하며, 종이봉투 안의 내용물이 은행알의 형태(테두리 선 등)를 정확히 띠고 있음.",
      "hard_violations": [],
      "physics": "손이 봉투를 잡고 지탱하며, 봉투의 하단은 테이블 표면에 자연스럽게 놓여 있음."
     },
     {
      "label": "B",
      "direction": "종이봉투의 열린 입구가 렌즈 방향을 향하고 있음.",
      "built_space": "배경에 취조실의 창문과 캐비닛이 보이며, 피사체 아래의 테이블 형태가 레퍼런스와 일치함.",
      "entities": "장원섭의 수트와 셔츠 소매는 일치하나, 봉투 안의 은행알이 특징적인 테두리 없이 지나치게 매끄러운 형태를 띠고 있음.",
      "hard_violations": [],
      "physics": "손이 봉투를 잡고 지탱하며, 봉투의 하단은 테이블 표면에 자연스럽게 놓여 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1882,
     "B": 1755
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S53sh5__bgfirst_bg.png",
   "bg_asset_id": "2bdfc18c-5ae9-4091-866a-c2f9e838c27b",
   "bg_record_key": "S53sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S53sh5::cine": {
  "applied": true,
  "fingerprint": "e8211dec748f44ea7a94f637c31d4657ae1f4d306768d0732b23678256a40b5f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S53sh5_sel.png",
  "source_sha256": "a6d1baf8d0dba7697997a683f541164fd57a3670d1af720da16d9c46ca4c7146",
  "file": "S53sh5_cine.png",
  "latency_ms": 13183
 },
 "S53sh11::signage": {
  "fp": "65191611118956b1",
  "inscriptions": []
 },
 "S53sh11": {
  "input_fingerprint": "66c7433b593d568c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 웃음기를 거둔 채 미간을 미세하게 찌푸리며 싸늘하게 굳어버린 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the suspect's side of the table inside the prosecution interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height near 지국현 (현재)'s side of the table, the tracking move finishes a few inches closer in a tight frontal three-quarter close-up. His face occupies the left-center of frame with open space toward 장원섭's off-screen position; he keeps chewing as his brow tightens and his eyes remain fixed on 장원섭.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현 (현재) in the middle-left of the frame, foreground, looks toward 장원섭 across the table.\n- KEY BACKGROUND ELEMENTS: interrogation-room table edge (Separating 지국현 (현재) from 장원섭); used as Narrow lower-frame spatial reference maintaining the interrogation-table axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality preserves the minute change from fading ease to guarded tension without heightened shadows.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open bag of roasted ginkgo nuts remains in the center of the table, with both men having eaten from it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 웃음기를 거둔 채 미간을 미세하게 찌푸리며 싸늘하게 굳어버린 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the suspect's side of the table inside the prosecution interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height near 지국현 (현재)'s side of the table, the tracking move finishes a few inches closer in a tight frontal three-quarter close-up. His face occupies the left-center of frame with open space toward 장원섭's off-screen position; he keeps chewing as his brow tightens and his eyes remain fixed on 장원섭.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현 (현재) in the middle-left of the frame, foreground, looks toward 장원섭 across the table.\n- KEY BACKGROUND ELEMENTS: interrogation-room table edge (Separating 지국현 (현재) from 장원섭); used as Narrow lower-frame spatial reference maintaining the interrogation-table axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality preserves the minute change from fading ease to guarded tension without heightened shadows.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open bag of roasted ginkgo nuts remains in the center of the table, with both men having eaten from it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 웃음기를 거둔 채 미간을 미세하게 찌푸리며 싸늘하게 굳어버린 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the suspect's side of the table inside the prosecution interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height near 지국현 (현재)'s side of the table, the tracking move finishes a few inches closer in a tight frontal three-quarter close-up. His face occupies the left-center of frame with open space toward 장원섭's off-screen position; he keeps chewing as his brow tightens and his eyes remain fixed on 장원섭.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현 (현재) in the middle-left of the frame, foreground, looks toward 장원섭 across the table.\n- KEY BACKGROUND ELEMENTS: interrogation-room table edge (Separating 지국현 (현재) from 장원섭); used as Narrow lower-frame spatial reference maintaining the interrogation-table axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality preserves the minute change from fading ease to guarded tension without heightened shadows.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open bag of roasted ginkgo nuts remains in the center of the table, with both men having eaten from it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "화면 밖 오른쪽을 응시하며 시선이 대상을 향함.",
    "built_space": "테이블 앞에 앉아 있으며, 배경의 창문과 가구가 위치 레퍼런스와 일치함.",
    "entities": "지국현의 얼굴과 죄수복(교도소 1234 명찰 포함)이 캐릭터 레퍼런스와 완벽히 일치하며, 테이블 위 은행알 봉투도 제대로 구현됨.",
    "hard_violations": [],
    "physics": "봉투는 테이블 위에 안정적으로 놓여 있고, 인물은 앉은 자세를 잘 유지함."
   },
   {
    "label": "A",
    "direction": "화면 밖 오른쪽을 응시하며 시선이 대상을 향함.",
    "built_space": "테이블 앞에 앉아 있으며, 배경 구조는 위치 레퍼런스와 일치함.",
    "entities": "지국현의 얼굴은 일치하나, 캐릭터 레퍼런스의 죄수복이 아닌 이전 샷의 정장을 입고 있음.",
    "hard_violations": [
     "이전 샷의 의상(정장)을 가져오지 말라는 지시를 어기고 레퍼런스의 죄수복을 누락함"
    ],
    "physics": "봉투는 테이블 위에 안정적으로 놓여 있고, 인물은 앉은 자세를 잘 유지함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "프롬프트의 지시대로 이전 샷의 의상을 배제하고 캐릭터 레퍼런스의 죄수복을 정확히 구현했으며, 표정과 배경도 잘 묘사되었습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "금지된 이전 샷의 의상(정장)을 인물에게 그대로 입혀 캐릭터 레퍼런스의 의상(죄수복) 조건을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S53sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우측 하단 종이봉투 안의 내용물이 레퍼런스의 테두리가 있는 은행알 형태와 달리 매끄럽고 둥근 알 모양으로 다르게 묘사되었습니다.",
     "fix_en": "Redraw the smooth yellow balls in the paper bag to be roasted ginkgo nuts with distinct center seams.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "지국현의 입이 굳게 다물어져 있어 프롬프트에 명시된 '계속 씹고 있는(keeps chewing)' 동작이 묘사되지 않았습니다.",
     "fix_en": "Modify the man's jawline and closed lips to show the muscle tension of actively chewing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "화면 왼쪽 인물의 얼굴·이목구비가 캐릭터 레퍼런스의 지국현과 전혀 다른 사람으로 보임",
     "fix_en": "Replace the man's face to exactly match the provided character reference of Ji Guk-hyun with a slimmer face and sharper eyes. Preserve the current framing, the man's pose, his blue prison uniform with its patches, the paper bag with nuts, the desk, the lighting, and the background room elements.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신과 테이블·은행 봉지가 크게 잡힌 미디엄에 가까운 구도임",
     "fix_en": "Change the camera framing from a medium shot to a tight frontal three-quarter close-up on the man's face.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우측 하단 종이봉투 안의 내용물이 레퍼런스의 테두리가 있는 은행알 형태와 달리 매끄럽고 둥근 알 모양으로 다르게 묘사되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "지국현의 입이 굳게 다물어져 있어 프롬프트에 명시된 '계속 씹고 있는(keeps chewing)' 동작이 묘사되지 않았습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 왼쪽 인물의 얼굴·이목구비가 캐릭터 레퍼런스의 지국현과 전혀 다른 사람으로 보임",
     "severity": "critical"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신과 테이블·은행 봉지가 크게 잡힌 미디엄에 가까운 구도임",
     "severity": "major"
    },
    {
     "issue_ko": "입이 다물어져 있어 지시된 씹는 동작이 보이지 않음",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "우측 하단 종이봉투 안의 내용물이 레퍼런스의 테두리가 있는 은행알 형태와 달리 매끄럽고 둥근 알 모양으로 다르게 묘사되었습니다.",
    "fix_en": "Redraw the smooth yellow balls in the paper bag to be roasted ginkgo nuts with distinct center seams.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "지국현의 입이 굳게 다물어져 있어 프롬프트에 명시된 '계속 씹고 있는(keeps chewing)' 동작이 묘사되지 않았습니다.",
    "fix_en": "Modify the man's jawline and closed lips to show the muscle tension of actively chewing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "얼굴 클로즈업이 아니라 상반신과 테이블·은행 봉지가 크게 잡힌 미디엄에 가까운 구도임",
    "fix_en": "Change the camera framing from a medium shot to a tight frontal three-quarter close-up on the man's face.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the man's face to exactly match the provided character reference of Ji Guk-hyun with a slimmer face and sharper eyes. Preserve the current framing, the man's pose, his blue prison uniform with its patches, the paper bag with nuts, the desk, the lighting, and the background room elements.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1550,
      "verdict_ko": "얼굴 윤곽을 갸름하게 수정하여 레퍼런스의 정체성과 훨씬 잘 일치하며, 차갑고 굳은 표정 연출도 안정적으로 유지됩니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 캐릭터 레퍼런스와 다른 얼굴의 인물로 대체됨"
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "얼굴형이 넓고 각진 형태로 묘사되어 캐릭터 레퍼런스의 '갸름한 얼굴' 설정과 인상 차이가 발생하며 동일인물로 보기 어렵습니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 캐릭터 레퍼런스와 다른 얼굴의 인물로 대체됨"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.8
     },
     "adjusted": {
      "A": 1.5,
      "B": 1.55
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 지국현이 캐릭터 레퍼런스와 다른 얼굴의 인물로 대체됨"
      ],
      "B": [
       "[openrouter:x-ai/grok-4.6] 지국현이 캐릭터 레퍼런스와 다른 얼굴의 인물로 대체됨"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.2,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1550,
      "verdict_ko": "얼굴 윤곽을 갸름하게 수정하여 레퍼런스의 정체성과 훨씬 잘 일치하며, 차갑고 굳은 표정 연출도 안정적으로 유지됩니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 캐릭터 레퍼런스와 다른 얼굴의 인물로 대체됨"
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "얼굴형이 넓고 각진 형태로 묘사되어 캐릭터 레퍼런스의 '갸름한 얼굴' 설정과 인상 차이가 발생하며 동일인물로 보기 어렵습니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 캐릭터 레퍼런스와 다른 얼굴의 인물로 대체됨"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "지시된 클로즈업 앵글, 캐릭터의 굳은 표정, 은행알 봉투 등 주요 요구사항을 충실히 구현함."
     },
     {
      "label": "B",
      "score": 2000,
      "verdict_ko": "후보 A와 시각적 차이를 찾기 어려우며, 지시문의 모든 요구사항을 동일하게 잘 반영함."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 2.0
     },
     "adjusted": {
      "A": 1.714,
      "B": 2.0
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "지시된 클로즈업 앵글, 캐릭터의 굳은 표정, 은행알 봉투 등 주요 요구사항을 충실히 구현함."
     },
     {
      "label": "A",
      "score": 2000,
      "verdict_ko": "후보 A와 시각적 차이를 찾기 어려우며, 지시문의 모든 요구사항을 동일하게 잘 반영함."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 3500,
     "B": 3264
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh5"
  }
 },
 "S53sh11::cine": {
  "applied": true,
  "fingerprint": "bce681485e9bcd8492a2813a1543a2767f87ca994c288537c76964bfc2c072d1",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S53sh11_sel.png",
  "source_sha256": "0489ab56cd09464e7245d0a5c69a35d038ad597392019a7511e78514e160ad31",
  "file": "S53sh11_cine.png",
  "latency_ms": 10905
 },
 "S53sh15::signage": {
  "fp": "db3b6d8b162f0b37",
  "inscriptions": [
   {
    "surface_native": "서류철 표지",
    "text_native": "수사기록",
    "reason_ko": "검사 테이블 위에 놓인 서류철에 '수사기록'이라는 공식 표기를 더하여 대한민국 검찰 조사실의 사실감을 높입니다."
   },
   {
    "surface_native": "탁상용 명패",
    "text_native": "검사 장원섭",
    "reason_ko": "검사석 테이블 위에 장원섭의 직책과 이름을 표시한 명패를 배치하여 극 중 인물의 신분과 대치 상황을 명확히 보여줍니다."
   }
  ]
 },
 "S53sh15": {
  "input_fingerprint": "c60bf9da8ea01eea",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현을 향해 정색한 표정으로 날카로운 경고의 눈빛을 쏘아보는 장원섭의 단호한 얼굴\n\nLOCATION (lock): At the prosecutor's side of the table inside the prosecution interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera ends its track just inside 지국현 (현재)'s near shoulder line, looking slightly upward across the table at 장원섭 from a narrow three-quarter angle. 지국현 (현재)'s shoulder remains a soft foreground edge while 장원섭's firm face occupies most of the middle frame, his eyes locked on 지국현 (현재) as he delivers the warning.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현 (현재) near shoulder in the middle-right of the frame, foreground, looks toward 장원섭 across the table; 장원섭 in the middle-center of the frame, midground, looks toward 지국현 (현재) across the table; interrogation table between them in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interrogation-room table (Positioned between the two seated men); used as Lower-frame divider that sustains the opposing eyeline.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime tonality and moderate contrast hold 장원섭's warning expression in naturalistic detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open ginkgo bag remains on the table as Ji Gukhyeon chews another nut and Wonseop warns him.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"수사기록\"\n- 탁상용 명패: \"검사 장원섭\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현을 향해 정색한 표정으로 날카로운 경고의 눈빛을 쏘아보는 장원섭의 단호한 얼굴\n\nLOCATION (lock): At the prosecutor's side of the table inside the prosecution interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera ends its track just inside 지국현 (현재)'s near shoulder line, looking slightly upward across the table at 장원섭 from a narrow three-quarter angle. 지국현 (현재)'s shoulder remains a soft foreground edge while 장원섭's firm face occupies most of the middle frame, his eyes locked on 지국현 (현재) as he delivers the warning.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현 (현재) near shoulder in the middle-right of the frame, foreground, looks toward 장원섭 across the table; 장원섭 in the middle-center of the frame, midground, looks toward 지국현 (현재) across the table; interrogation table between them in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interrogation-room table (Positioned between the two seated men); used as Lower-frame divider that sustains the opposing eyeline.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime tonality and moderate contrast hold 장원섭's warning expression in naturalistic detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open ginkgo bag remains on the table as Ji Gukhyeon chews another nut and Wonseop warns him.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"수사기록\"\n- 탁상용 명패: \"검사 장원섭\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현을 향해 정색한 표정으로 날카로운 경고의 눈빛을 쏘아보는 장원섭의 단호한 얼굴\n\nLOCATION (lock): At the prosecutor's side of the table inside the prosecution interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera ends its track just inside 지국현 (현재)'s near shoulder line, looking slightly upward across the table at 장원섭 from a narrow three-quarter angle. 지국현 (현재)'s shoulder remains a soft foreground edge while 장원섭's firm face occupies most of the middle frame, his eyes locked on 지국현 (현재) as he delivers the warning.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현 (현재) near shoulder in the middle-right of the frame, foreground, looks toward 장원섭 across the table; 장원섭 in the middle-center of the frame, midground, looks toward 지국현 (현재) across the table; interrogation table between them in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interrogation-room table (Positioned between the two seated men); used as Lower-frame divider that sustains the opposing eyeline.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained neutral-to-cool daytime tonality and moderate contrast hold 장원섭's warning expression in naturalistic detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The open ginkgo bag remains on the table as Ji Gukhyeon chews another nut and Wonseop warns him.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서류철 표지: \"수사기록\"\n- 탁상용 명패: \"검사 장원섭\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1714,
   "A": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "지국현의 어깨를 프레임 우측 전경에 배치하는 레이아웃 지시를 정확히 따랐으며, 명패와 서류철의 지정된 텍스트도 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "텍스트 소품과 인물의 외형은 훌륭하게 묘사되었으나, 프레임 우측에 있어야 할 지국현의 어깨가 좌측에 배치되어 화면 레이아웃 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S53sh11_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라를 등진 지국현의 뒷모습(오른쪽 어깨)에 앞가슴에 부착되어야 할 '교도소 1234' 패치가 잘못 생성되었습니다.",
     "fix_en": "Remove the white patch from the back of the blue uniform in the foreground, replacing it with plain blue fabric, keeping the foreground character's pose, the prosecutor, the table, and the background exactly as they are.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "테이블 위의 '수사기록' 서류철이 사용자인 장원섭이 아닌 지국현(카메라) 방향을 향해 역방향으로 놓여 있습니다.",
     "fix_en": "Redraw the beige folder on the table so its printed face is turned away from the camera toward the prosecutor in the grey suit, keeping both characters, the black nameplate, the bag of nuts, and the room lighting exactly as they are.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "카메라 앵글이 반전되었음에도 이전 샷에서 지국현 뒤에 있던 배경(창문과 갈색 캐비닛)이 장원섭의 뒤에 동일하게 나타납니다.",
     "fix_en": "Replace the window and brown cabinet in the background with a plain grey wall to resolve the spatial inconsistency, keeping both characters, their clothing, the table, and all props exactly as they are.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 왼쪽 장원섭의 얼굴이 캐릭터 레퍼런스가 아니라 이전 샷 수감자와 동일한 얼굴이다",
     "fix_en": "Replace the face of the prosecutor in the grey suit with the face from the character reference image, keeping his current stern expression, eye direction, posture, grey suit, the foreground character, the table, and the background exactly as they are.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "화면 오른쪽에 지국현의 등·머리·팔이 크게 들어와 장원섭만 보여야 하는 규정을 어겼다",
     "fix_en": "Darken and heavily blur the foreground character's head, back, and arm so they recede into a soft edge, keeping the prosecutor, the table props, and the background exactly as they are.",
     "severity": "critical",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라를 등진 지국현의 뒷모습(오른쪽 어깨)에 앞가슴에 부착되어야 할 '교도소 1234' 패치가 잘못 생성되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "테이블 위의 '수사기록' 서류철이 사용자인 장원섭이 아닌 지국현(카메라) 방향을 향해 역방향으로 놓여 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라 앵글이 반전되었음에도 이전 샷에서 지국현 뒤에 있던 배경(창문과 갈색 캐비닛)이 장원섭의 뒤에 동일하게 나타납니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 왼쪽 장원섭의 얼굴이 캐릭터 레퍼런스가 아니라 이전 샷 수감자와 동일한 얼굴이다",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 오른쪽에 지국현의 등·머리·팔이 크게 들어와 장원섭만 보여야 하는 규정을 어겼다",
     "severity": "critical"
    },
    {
     "issue_ko": "클로즈업이 아니며 장원섭 얼굴이 프레임 대부분을 차지하지 않고 넓은 투샷이다",
     "severity": "major"
    },
    {
     "issue_ko": "지국현이 중간-오른쪽 부드러운 어깨 가장자리가 아니라 오른쪽 큰 면적을 차지한다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "카메라를 등진 지국현의 뒷모습(오른쪽 어깨)에 앞가슴에 부착되어야 할 '교도소 1234' 패치가 잘못 생성되었습니다.",
    "fix_en": "Remove the white patch from the back of the blue uniform in the foreground, replacing it with plain blue fabric, keeping the foreground character's pose, the prosecutor, the table, and the background exactly as they are.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "테이블 위의 '수사기록' 서류철이 사용자인 장원섭이 아닌 지국현(카메라) 방향을 향해 역방향으로 놓여 있습니다.",
    "fix_en": "Redraw the beige folder on the table so its printed face is turned away from the camera toward the prosecutor in the grey suit, keeping both characters, the black nameplate, the bag of nuts, and the room lighting exactly as they are.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "카메라 앵글이 반전되었음에도 이전 샷에서 지국현 뒤에 있던 배경(창문과 갈색 캐비닛)이 장원섭의 뒤에 동일하게 나타납니다.",
    "fix_en": "Replace the window and brown cabinet in the background with a plain grey wall to resolve the spatial inconsistency, keeping both characters, their clothing, the table, and all props exactly as they are.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the face of the prosecutor in the grey suit with the face from the character reference image, keeping his current stern expression, eye direction, posture, grey suit, the foreground character, the table, and the background exactly as they are.\n- Darken and heavily blur the foreground character's head, back, and arm so they recede into a soft edge, keeping the prosecutor, the table props, and the background exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1125,
      "verdict_ko": "전경 인물의 등(오른쪽 어깨 뒤)에 이전 장면 참조 이미지의 앞가슴 명찰(\"교도소 1234\")이 비정상적으로 붙어 있고, 은행알을 쥔 손의 해부학적 구조가 어색한 치명적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 의상 구조: 전경 인물을 뒤에서 바라보는 구도임에도 불구하고, 이전 샷 참조 이미지에서 가슴에 있던 명찰('교도소 1234')이 인물의 오른쪽 등/어깨 뒤편에 잘못 인쇄됨."
     },
     {
      "label": "B",
      "score": 1429,
      "verdict_ko": "요구된 카메라 앵글과 오버더숄더 구도를 정확히 구현했으며, 장원섭의 외모와 표정, 지정된 텍스트(\"수사기록\", \"검사 장원섭\")를 물리적 오류 없이 훌륭하게 렌더링했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.375,
      "B": 1.429
     },
     "adjusted": {
      "A": 1.125,
      "B": 1.429
     },
     "violations": {
      "A": [
       "[gemini-pro] 물리적으로 불가능한 의상 구조: 전경 인물을 뒤에서 바라보는 구도임에도 불구하고, 이전 샷 참조 이미지에서 가슴에 있던 명찰('교도소 1234')이 인물의 오른쪽 등/어깨 뒤편에 잘못 인쇄됨."
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.571,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1125,
      "verdict_ko": "전경 인물의 등(오른쪽 어깨 뒤)에 이전 장면 참조 이미지의 앞가슴 명찰(\"교도소 1234\")이 비정상적으로 붙어 있고, 은행알을 쥔 손의 해부학적 구조가 어색한 치명적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 의상 구조: 전경 인물을 뒤에서 바라보는 구도임에도 불구하고, 이전 샷 참조 이미지에서 가슴에 있던 명찰('교도소 1234')이 인물의 오른쪽 등/어깨 뒤편에 잘못 인쇄됨."
     },
     {
      "label": "B",
      "score": 1429,
      "verdict_ko": "요구된 카메라 앵글과 오버더숄더 구도를 정확히 구현했으며, 장원섭의 외모와 표정, 지정된 텍스트(\"수사기록\", \"검사 장원섭\")를 물리적 오류 없이 훌륭하게 렌더링했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1625,
      "verdict_ko": "지국현의 푸른색 수의와 은행알을 씹는 동작, 명시된 소품의 텍스트와 장원섭의 외형을 모두 완벽하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "전경에 위치한 지국현의 의상이 이전 샷의 수의가 아닌 정장으로 바뀌었으며, 은행알을 먹는 지시 사항이 누락되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.625
     },
     "adjusted": {
      "A": 1.5,
      "B": 1.625
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.375,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1625,
      "verdict_ko": "지국현의 푸른색 수의와 은행알을 씹는 동작, 명시된 소품의 텍스트와 장원섭의 외형을 모두 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "전경에 위치한 지국현의 의상이 이전 샷의 수의가 아닌 정장으로 바뀌었으며, 은행알을 먹는 지시 사항이 누락되었습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2750,
     "B": 2929
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh11"
  }
 },
 "S53sh15::cine": {
  "applied": true,
  "fingerprint": "0dcafbc57d5aa5c18ba1fdf2114ee3d555b377592623580d34b69ba96a9a26b0",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S53sh15_sel.png",
  "source_sha256": "fa01b2ce65ebca994091135514c0fec0bf60d7d45c45fbed384104b0101fda2b",
  "file": "S53sh15_cine.png",
  "latency_ms": 11226
 },
 "S54sh7::signage": {
  "fp": "2546646066785bd3",
  "inscriptions": [
   {
    "surface_native": "책상 위 명패",
    "text_native": "검사",
    "reason_ko": "검사의 개인 사무실이라는 공간적 배경을 명확히 하고 극중 권위적이고 차가운 분위기를 뒷받침하기 위해 책상 위의 '검사' 명패가 필요합니다."
   }
  ]
 },
 "S54sh7": {
  "input_fingerprint": "7f0d73812b505bfd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사무실 소파를 사이에 두고 거리를 둔 채 마주 앉은 장원섭과 전택수의 전경\n\nLOCATION (lock): In the sofa meeting area of the prosecutor's private office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: A static wide composition at seated chest height observes the sofa area from a three-quarter side axis, placing 장원섭 on one side and 전택수 opposite him with the table clearly separating them. Both men angle their torsos toward each other rather than toward the lens, and their cautious eye contact carries the formal distance of the meeting.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground, looks toward 전택수 across the table; 전택수 in the middle-right of the frame, midground, looks toward 장원섭 across the table; table separating the sofas in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: office sofas (Occupied by 장원섭 and 전택수) — The seating faces are oriented toward each other across the table; used as Opposing seating areas that establish the men's deliberate separation; table in front of the sofas (Positioned between 장원섭 and 전택수); used as Central spatial divider between the seated men.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime illumination appropriate to the office is treated in subdued neutral-to-cool tones with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecutor's office, formal furnishings, daylight, and both men's business attire. Exclude the earlier handshake at the desk and place them opposite each other on the sofas.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사무실 소파를 사이에 두고 거리를 둔 채 마주 앉은 장원섭과 전택수의 전경\n\nLOCATION (lock): In the sofa meeting area of the prosecutor's private office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: A static wide composition at seated chest height observes the sofa area from a three-quarter side axis, placing 장원섭 on one side and 전택수 opposite him with the table clearly separating them. Both men angle their torsos toward each other rather than toward the lens, and their cautious eye contact carries the formal distance of the meeting.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground, looks toward 전택수 across the table; 전택수 in the middle-right of the frame, midground, looks toward 장원섭 across the table; table separating the sofas in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: office sofas (Occupied by 장원섭 and 전택수) — The seating faces are oriented toward each other across the table; used as Opposing seating areas that establish the men's deliberate separation; table in front of the sofas (Positioned between 장원섭 and 전택수); used as Central spatial divider between the seated men.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime illumination appropriate to the office is treated in subdued neutral-to-cool tones with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecutor's office, formal furnishings, daylight, and both men's business attire. Exclude the earlier handshake at the desk and place them opposite each other on the sofas.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 사무실 소파를 사이에 두고 거리를 둔 채 마주 앉은 장원섭과 전택수의 전경\n\nLOCATION (lock): In the sofa meeting area of the prosecutor's private office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: A static wide composition at seated chest height observes the sofa area from a three-quarter side axis, placing 장원섭 on one side and 전택수 opposite him with the table clearly separating them. Both men angle their torsos toward each other rather than toward the lens, and their cautious eye contact carries the formal distance of the meeting.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground, looks toward 전택수 across the table; 전택수 in the middle-right of the frame, midground, looks toward 장원섭 across the table; table separating the sofas in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: office sofas (Occupied by 장원섭 and 전택수) — The seating faces are oriented toward each other across the table; used as Opposing seating areas that establish the men's deliberate separation; table in front of the sofas (Positioned between 장원섭 and 전택수); used as Central spatial divider between the seated men.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime illumination appropriate to the office is treated in subdued neutral-to-cool tones with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecutor's office, formal furnishings, daylight, and both men's business attire. Exclude the earlier handshake at the desk and place them opposite each other on the sofas.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "두 인물이 테이블을 사이에 두고 서로를 응시함.",
    "built_space": "마주 보는 소파와 중앙 테이블, 이전 샷과 동일한 사무실 배경이 올바르게 배치됨.",
    "entities": "왼쪽은 장원섭, 오른쪽은 전택수의 얼굴과 일치하며 둘 다 정장을 착용함. 테이블 위에 '검사' 명패가 정확히 렌더링됨.",
    "hard_violations": [],
    "physics": "두 인물 모두 소파에 안정적으로 앉아 자세를 지탱하고 있음."
   },
   {
    "label": "A",
    "direction": "두 인물이 서로를 마주보고 있음.",
    "built_space": "사무실 배경과 소파, 테이블 배치가 지시된 공간 구조와 일치함.",
    "entities": "왼쪽 인물은 장원섭과 일치하나, 오른쪽 인물 역시 전택수가 아닌 장원섭의 얼굴로 렌더링됨. '검사' 명패는 정확함.",
    "hard_violations": [],
    "physics": "두 사람 모두 소파에 자연스럽게 앉아 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "두 인물의 신원과 복장, 지정된 사무실 배경 및 명패 텍스트를 모두 정확하게 구현함."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "오른쪽 인물에 전택수 대신 장원섭의 얼굴이 중복 적용되어 인물 지정 지시를 크게 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S35sh12_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 우측에 앉은 인물(전택수)의 얼굴이 레퍼런스와 전혀 다르게 좌측 인물(장원섭)과 거의 동일한 외모(복제)로 생성되었습니다.",
     "fix_en": "Change the face of the man sitting on the right to match the older, square-jawed features of the 전택수 character reference; preserve his dark suit, body posture, the man on the left, the sofas, the coffee table, and the lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "프롬프트에서 '책상 위 명패'로 지시했으나, 명패가 책상이 아닌 소파 사이의 탁자(커피 테이블) 위에 잘못 배치되었습니다.",
     "fix_en": "Remove the nameplate from the foreground coffee table, leaving a bare wooden surface, and draw a small nameplate on the desk in the background left; preserve both seated men, their clothing, the sofas, and the background.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 우측에 앉은 인물(전택수)의 얼굴이 레퍼런스와 전혀 다르게 좌측 인물(장원섭)과 거의 동일한 외모(복제)로 생성되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "프롬프트에서 '책상 위 명패'로 지시했으나, 명패가 책상이 아닌 소파 사이의 탁자(커피 테이블) 위에 잘못 배치되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 남성의 얼굴이 전택수 레퍼런스와 이목구비·인상부터 다르다",
     "severity": "major"
    },
    {
     "issue_ko": "지정된 책상 위 ‘검사’ 명패가 없고 같은 글자가 소파 사이 테이블 위에 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트에서 '책상 위 명패'로 지시했으나, 명패가 책상이 아닌 소파 사이의 탁자(커피 테이블) 위에 잘못 배치되었습니다.",
    "fix_en": "Remove the nameplate from the foreground coffee table, leaving a bare wooden surface, and draw a small nameplate on the desk in the background left; preserve both seated men, their clothing, the sofas, and the background.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the face of the man sitting on the right to match the older, square-jawed features of the 전택수 character reference; preserve his dark suit, body posture, the man on the left, the sofas, the coffee table, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 와이드 샷 구도, 두 인물의 좌우 배치 및 시선 교환, 이전 샷에서 지정된 의상과 사무실 배경을 모두 충실히 구현하여 프롬프트의 의도를 정확하게 따랐습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "요구된 투샷 구도를 완전히 무시하고 이전 샷의 프레이밍을 그대로 반복했으며, 인물 한 명이 누락되고 신원과 의상이 뒤섞이는 치명적인 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "좌측의 장원섭과 우측의 전택수가 테이블을 사이에 두고 서로를 응시함.",
      "built_space": "두 개의 소파가 마주보고 배치되었고 중간에 테이블이 있음. 창문, 수납장 등 이전 샷의 공간 구조가 잘 유지됨.",
      "entities": "장원섭(좌)은 이전 샷의 의상(도트 넥타이, 행거치프)과 얼굴이 일치함. 전택수(우) 역시 참조 이미지의 얼굴과 일치하며 정장을 착용함. 테이블 위 명패에 '검사' 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "두 사람 모두 소파에 안정적으로 앉아 있으며 손과 발의 지지가 자연스러움."
     },
     {
      "label": "B",
      "direction": "화면 속 단일 인물이 카메라 렌즈를 정면으로 응시함.",
      "built_space": "마주 보는 소파 공간이 구현되지 않고 이전 샷의 책상 위치 구도가 그대로 사용됨.",
      "entities": "장원섭이 누락됨. 등장인물은 전택수의 얼굴을 하고 있으나 장원섭의 이전 샷 의상을 입고 있음. 명패 텍스트도 지시와 다름.",
      "hard_violations": [
       "필수 인물 누락 (장원섭 없음)",
       "프레이밍 및 구도 지시 완전 위반 (와이드 샷 및 투샷 아님)",
       "신원 및 의상 잠금 지시 위반 (전택수 얼굴에 장원섭 의상 결합)"
      ],
      "physics": "인물이 하단 구조물에 안정적으로 기대어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 와이드 샷 구도, 두 인물의 좌우 배치 및 시선 교환, 이전 샷에서 지정된 의상과 사무실 배경을 모두 충실히 구현하여 프롬프트의 의도를 정확하게 따랐습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "요구된 투샷 구도를 완전히 무시하고 이전 샷의 프레이밍을 그대로 반복했으며, 인물 한 명이 누락되고 신원과 의상이 뒤섞이는 치명적인 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "좌측의 장원섭과 우측의 전택수가 테이블을 사이에 두고 서로를 응시함.",
      "built_space": "두 개의 소파가 마주보고 배치되었고 중간에 테이블이 있음. 창문, 수납장 등 이전 샷의 공간 구조가 잘 유지됨.",
      "entities": "장원섭(좌)은 이전 샷의 의상(도트 넥타이, 행거치프)과 얼굴이 일치함. 전택수(우) 역시 참조 이미지의 얼굴과 일치하며 정장을 착용함. 테이블 위 명패에 '검사' 텍스트가 정확함.",
      "hard_violations": [],
      "physics": "두 사람 모두 소파에 안정적으로 앉아 있으며 손과 발의 지지가 자연스러움."
     },
     {
      "label": "B",
      "direction": "화면 속 단일 인물이 카메라 렌즈를 정면으로 응시함.",
      "built_space": "마주 보는 소파 공간이 구현되지 않고 이전 샷의 책상 위치 구도가 그대로 사용됨.",
      "entities": "장원섭이 누락됨. 등장인물은 전택수의 얼굴을 하고 있으나 장원섭의 이전 샷 의상을 입고 있음. 명패 텍스트도 지시와 다름.",
      "hard_violations": [
       "필수 인물 누락 (장원섭 없음)",
       "프레이밍 및 구도 지시 완전 위반 (와이드 샷 및 투샷 아님)",
       "신원 및 의상 잠금 지시 위반 (전택수 얼굴에 장원섭 의상 결합)"
      ],
      "physics": "인물이 하단 구조물에 안정적으로 기대어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "샷 텍스트가 요구한 두 인물의 소파 대면 구도를 완전히 무시하고 이전 컷을 그대로 복사하여 인물 1명이 누락된 치명적인 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "두 인물이 소파에 마주 앉은 전경 구도와 의상, 명패 텍스트까지 프롬프트의 지시사항을 매우 정확하게 구현했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "남자가 카메라 렌즈를 정면으로 응시함.",
      "built_space": "사무실 책상 앞에 앉아 있는 단일 구조.",
      "entities": "인물이 1명만 존재함. 의상과 구도는 이전 컷의 장원섭과 같으나 얼굴은 전택수와 유사함. 명패에 '검사 장원섭' 표기됨.",
      "hard_violations": [
       "지시된 인물 2명 중 1명 누락",
       "소파 구도가 아닌 이전 컷의 책상 구도 복사"
      ],
      "physics": "의자에 앉아 책상에 팔을 올리고 있는 안정적인 자세."
     },
     {
      "label": "B",
      "direction": "두 남자가 테이블을 사이에 두고 서로의 얼굴을 향해 시선을 맞춤.",
      "built_space": "마주 보는 두 개의 소파와 그 사이를 가르는 테이블이 배치된 사무실 공간.",
      "entities": "좌측에 장원섭(회색 정장 및 레퍼런스 일치), 우측에 전택수(비즈니스 정장 착용)가 배치됨. 테이블 위에 '검사' 명패 존재.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파에 체중을 싣고 앉아 무릎 쪽에 손을 얹은 자연스러운 지지 상태."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "샷 텍스트가 요구한 두 인물의 소파 대면 구도를 완전히 무시하고 이전 컷을 그대로 복사하여 인물 1명이 누락된 치명적인 오류가 있습니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "두 인물이 소파에 마주 앉은 전경 구도와 의상, 명패 텍스트까지 프롬프트의 지시사항을 매우 정확하게 구현했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "남자가 카메라 렌즈를 정면으로 응시함.",
      "built_space": "사무실 책상 앞에 앉아 있는 단일 구조.",
      "entities": "인물이 1명만 존재함. 의상과 구도는 이전 컷의 장원섭과 같으나 얼굴은 전택수와 유사함. 명패에 '검사 장원섭' 표기됨.",
      "hard_violations": [
       "지시된 인물 2명 중 1명 누락",
       "소파 구도가 아닌 이전 컷의 책상 구도 복사"
      ],
      "physics": "의자에 앉아 책상에 팔을 올리고 있는 안정적인 자세."
     },
     {
      "label": "A",
      "direction": "두 남자가 테이블을 사이에 두고 서로의 얼굴을 향해 시선을 맞춤.",
      "built_space": "마주 보는 두 개의 소파와 그 사이를 가르는 테이블이 배치된 사무실 공간.",
      "entities": "좌측에 장원섭(회색 정장 및 레퍼런스 일치), 우측에 전택수(비즈니스 정장 착용)가 배치됨. 테이블 위에 '검사' 명패 존재.",
      "hard_violations": [],
      "physics": "두 인물 모두 소파에 체중을 싣고 앉아 무릎 쪽에 손을 얹은 자연스러운 지지 상태."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S35sh12"
  }
 },
 "S54sh7::cine": {
  "applied": true,
  "fingerprint": "8725fb970a0484ca927c413403f65ef83fb88dd53390f07d53f22107c6e4ea14",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S54sh7_sel.png",
  "source_sha256": "f678d7e67efa6f937473f231b7efa7bd73d3271833c6852c23b650ed10911f83",
  "file": "S54sh7_cine.png",
  "latency_ms": 11302
 },
 "S54sh12::signage": {
  "fp": "f0628a471700cc60",
  "inscriptions": [
   {
    "surface_native": "편지봉투",
    "text_native": "친전",
    "reason_ko": "검사실에서 오가는 오래된 편지봉투에 사실성을 부여하고 중요한 개인적 편지임을 나타내기 위해 '친전'이라는 한글 표기가 필요합니다."
   }
  ]
 },
 "S54sh12": {
  "input_fingerprint": "4c935f90c586de53",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 손이 낡은 편지봉투를 소파 앞 테이블 위로 막 내려놓는 순간 클로즈업\n\nLOCATION (lock): Over the low table between the sofas in the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above tabletop height on 전택수's diagonal, the dolly-in catches his hand at the instant it releases the worn letter envelope in front of 장원섭. The hand and envelope occupy the central lower area without exceeding their natural scale, while portions of the table and the facing sofa area remain visible as context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: worn letter envelope (Just placed in front of 장원섭) — Its broad side is visible at a shallow angle as it meets the tabletop; used as Primary evidence object caught as it is placed on the table; table in front of the sofas (Between 장원섭 and 전택수); used as Scale reference and landing surface for the envelope; office sofas (Occupied by the two men) — The facing seating edges recede on opposite sides of the table; used as Partial background context retaining the meeting location.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality keeps the worn envelope and placing hand legible without dramatized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sofa table, office materials, daylight, and the seated men's positions. Exclude their faces from the insert and show only the old envelope being placed down.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu places the inmate's old tip-letter envelope on the table in front of Wonseop.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편지봉투: \"친전\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 손이 낡은 편지봉투를 소파 앞 테이블 위로 막 내려놓는 순간 클로즈업\n\nLOCATION (lock): Over the low table between the sofas in the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above tabletop height on 전택수's diagonal, the dolly-in catches his hand at the instant it releases the worn letter envelope in front of 장원섭. The hand and envelope occupy the central lower area without exceeding their natural scale, while portions of the table and the facing sofa area remain visible as context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: worn letter envelope (Just placed in front of 장원섭) — Its broad side is visible at a shallow angle as it meets the tabletop; used as Primary evidence object caught as it is placed on the table; table in front of the sofas (Between 장원섭 and 전택수); used as Scale reference and landing surface for the envelope; office sofas (Occupied by the two men) — The facing seating edges recede on opposite sides of the table; used as Partial background context retaining the meeting location.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality keeps the worn envelope and placing hand legible without dramatized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sofa table, office materials, daylight, and the seated men's positions. Exclude their faces from the insert and show only the old envelope being placed down.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu places the inmate's old tip-letter envelope on the table in front of Wonseop.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편지봉투: \"친전\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 손이 낡은 편지봉투를 소파 앞 테이블 위로 막 내려놓는 순간 클로즈업\n\nLOCATION (lock): Over the low table between the sofas in the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above tabletop height on 전택수's diagonal, the dolly-in catches his hand at the instant it releases the worn letter envelope in front of 장원섭. The hand and envelope occupy the central lower area without exceeding their natural scale, while portions of the table and the facing sofa area remain visible as context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: worn letter envelope (Just placed in front of 장원섭) — Its broad side is visible at a shallow angle as it meets the tabletop; used as Primary evidence object caught as it is placed on the table; table in front of the sofas (Between 장원섭 and 전택수); used as Scale reference and landing surface for the envelope; office sofas (Occupied by the two men) — The facing seating edges recede on opposite sides of the table; used as Partial background context retaining the meeting location.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued neutral-to-cool daytime tonality keeps the worn envelope and placing hand legible without dramatized highlights.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sofa table, office materials, daylight, and the seated men's positions. Exclude their faces from the insert and show only the old envelope being placed down.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu places the inmate's old tip-letter envelope on the table in front of Wonseop.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 편지봉투: \"친전\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.571,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.5
   },
   "violations": {
    "A": [
     "[openrouter:x-ai/grok-4.6] 이전 스틸의 다른 인물(회색 양복)의 몸통과 의복을 가져와 전택수가 아닌 인물을 프레임에 넣음"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1179,
   "B": 1500
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "이전 샷 스틸의 의상(회색 정장, 하늘색 셔츠) 연속성을 정확히 유지했으며, 테이블 위로 봉투를 내려놓는 클로즈업 앵글과 '친전' 텍스트를 충실히 구현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 이전 스틸의 다른 인물(회색 양복)의 몸통과 의복을 가져와 전택수가 아닌 인물을 프레임에 넣음"
   },
   {
    "label": "B",
    "score": 1500,
    "verdict_ko": "이전 샷 스틸의 의상을 유지하라는 지시를 무시하고 캐릭터 레퍼런스의 경찰복과 체크 셔츠를 입혀 장면의 연속성을 크게 훼손했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S54sh7_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "전택수의 팔뚝 소매가 이전 컷 사진에 고정된 어두운 정장, 흰 셔츠, 가죽 시계가 아닌 캐릭터 레퍼런스의 캐주얼 점퍼, 체크 셔츠, 메탈 시계로 잘못 렌더링되었습니다.",
     "fix_en": "Change the sleeve to a dark suit jacket and white shirt cuff, and replace the metal watch with a leather-strap watch, preserving the hand, envelope, plaque, table, and background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "카메라가 전택수의 대각선에 위치해야 함에도 그의 오른손이 화면 왼쪽에서 들어오고 있어 이전 컷의 좌석 배치 및 카메라 방향과 모순됩니다.",
     "fix_en": "Adjust the arm's entry point to the bottom right to reflect the correct seating position, preserving the table, plaque, and envelope.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이전 컷에서 테이블 측면을 향해 놓여 있던 '검사' 명패가, 카메라 앵글이 대각선으로 바뀌었음에도 여전히 화면을 정면으로 향하고 있습니다.",
     "fix_en": "Rotate the name plaque to sit at an angle relative to the camera, preserving the table, hand, envelope, and background.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "전택수의 팔뚝 소매가 이전 컷 사진에 고정된 어두운 정장, 흰 셔츠, 가죽 시계가 아닌 캐릭터 레퍼런스의 캐주얼 점퍼, 체크 셔츠, 메탈 시계로 잘못 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "카메라가 전택수의 대각선에 위치해야 함에도 그의 오른손이 화면 왼쪽에서 들어오고 있어 이전 컷의 좌석 배치 및 카메라 방향과 모순됩니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 컷에서 테이블 측면을 향해 놓여 있던 '검사' 명패가, 카메라 앵글이 대각선으로 바뀌었음에도 여전히 화면을 정면으로 향하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 전택수 손의 소매가 이전 장면의 정장이 아니라 남색 재킷과 체크 셔츠로 바뀌어 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "카메라가 전택수의 대각선에 위치해야 함에도 그의 오른손이 화면 왼쪽에서 들어오고 있어 이전 컷의 좌석 배치 및 카메라 방향과 모순됩니다.",
    "fix_en": "Adjust the arm's entry point to the bottom right to reflect the correct seating position, preserving the table, plaque, and envelope.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "이전 컷에서 테이블 측면을 향해 놓여 있던 '검사' 명패가, 카메라 앵글이 대각선으로 바뀌었음에도 여전히 화면을 정면으로 향하고 있습니다.",
    "fix_en": "Rotate the name plaque to sit at an angle relative to the camera, preserving the table, hand, envelope, and background.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the sleeve to a dark suit jacket and white shirt cuff, and replace the metal watch with a leather-strap watch, preserving the hand, envelope, plaque, table, and background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "캐릭터 레퍼런스에 제시된 전택수의 의상(남색 재킷, 체크무늬 셔츠, 메탈 시계)을 소매 디테일에 정확히 반영했으며, 봉투의 텍스트와 손의 동작이 물리적으로 매우 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "봉투의 텍스트와 배경 구성은 양호하나, 소매가 정장과 가죽 시계로 묘사되어 전택수의 캐릭터 레퍼런스 의상과 일치하지 않습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "화면 좌측에서 뻗어 나온 손이 테이블 위 지정된 위치로 봉투를 내려놓고 있음.",
      "built_space": "이전 샷 스틸의 구성을 따라 소파 사이의 낮은 테이블이 있으며, '검사' 명패가 카메라를 향해 올바르게 놓여 있음.",
      "entities": "봉투에 '친전' 텍스트가 정확히 표기됨. 손목의 의상(남색 점퍼, 체크 셔츠, 메탈 시계)이 전택수 캐릭터 레퍼런스와 완벽하게 일치함.",
      "hard_violations": [],
      "physics": "손가락이 봉투를 자연스럽게 파지하고 있으며, 봉투의 모서리가 테이블 표면에 닿아 중력과 지지점이 올바르게 표현됨."
     },
     {
      "label": "B",
      "direction": "화면 좌측에서 뻗어 나온 손이 테이블 위로 봉투를 향하게 하고 있음.",
      "built_space": "소파 사이의 테이블 및 '검사' 명패의 위치와 방향이 이전 샷을 잘 반영하고 있음.",
      "entities": "봉투 표면에 '친전'이 올바르게 적혀 있으나, 손목 부분의 의상이 캐릭터 레퍼런스(점퍼/체크셔츠)가 아닌 짙은 정장과 가죽 시계로 잘못 묘사됨.",
      "hard_violations": [],
      "physics": "손이 봉투를 쥐고 테이블에 내려놓는 동작의 접촉면과 지지가 정상적으로 묘사됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "캐릭터 레퍼런스에 제시된 전택수의 의상(남색 재킷, 체크무늬 셔츠, 메탈 시계)을 소매 디테일에 정확히 반영했으며, 봉투의 텍스트와 손의 동작이 물리적으로 매우 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "봉투의 텍스트와 배경 구성은 양호하나, 소매가 정장과 가죽 시계로 묘사되어 전택수의 캐릭터 레퍼런스 의상과 일치하지 않습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "화면 좌측에서 뻗어 나온 손이 테이블 위 지정된 위치로 봉투를 내려놓고 있음.",
      "built_space": "이전 샷 스틸의 구성을 따라 소파 사이의 낮은 테이블이 있으며, '검사' 명패가 카메라를 향해 올바르게 놓여 있음.",
      "entities": "봉투에 '친전' 텍스트가 정확히 표기됨. 손목의 의상(남색 점퍼, 체크 셔츠, 메탈 시계)이 전택수 캐릭터 레퍼런스와 완벽하게 일치함.",
      "hard_violations": [],
      "physics": "손가락이 봉투를 자연스럽게 파지하고 있으며, 봉투의 모서리가 테이블 표면에 닿아 중력과 지지점이 올바르게 표현됨."
     },
     {
      "label": "B",
      "direction": "화면 좌측에서 뻗어 나온 손이 테이블 위로 봉투를 향하게 하고 있음.",
      "built_space": "소파 사이의 테이블 및 '검사' 명패의 위치와 방향이 이전 샷을 잘 반영하고 있음.",
      "entities": "봉투 표면에 '친전'이 올바르게 적혀 있으나, 손목 부분의 의상이 캐릭터 레퍼런스(점퍼/체크셔츠)가 아닌 짙은 정장과 가죽 시계로 잘못 묘사됨.",
      "hard_violations": [],
      "physics": "손이 봉투를 쥐고 테이블에 내려놓는 동작의 접촉면과 지지가 정상적으로 묘사됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "이전 컷 스틸에 맞춰 정장 소매를 유지하려 했으나, 카메라가 장원섭의 자리(왼쪽)에 위치하여 '검사' 명패 앞면이 보이고 전택수의 손이 잘못된 좌석에서 등장하는 심각한 공간(Staging) 하드 비올레이션이 발생함.  ★위반: [gemini-pro] a person seated in a place the staging does not put them (카메라와 전택수의 손이 장원섭의 좌석 위치인 왼쪽에서 잘못 진입함)"
     },
     {
      "label": "B",
      "score": 1083,
      "verdict_ko": "A와 동일한 공간 왜곡(Staging) 하드 비올레이션이 발생했으며, 이전 컷 스틸로 고정(LOCKED)된 의상(정장) 지시를 무시하고 캐릭터 레퍼런스의 캐주얼 점퍼를 입혀 연속성을 완전히 훼손함.  ★위반: [gemini-pro] a person seated in a place the staging does not put them (카메라와 전택수의 손이 장원섭의 좌석 위치인 왼쪽에서 잘못 진입함)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.25,
      "B": 1.083
     },
     "violations": {
      "A": [
       "[gemini-pro] a person seated in a place the staging does not put them (카메라와 전택수의 손이 장원섭의 좌석 위치인 왼쪽에서 잘못 진입함)"
      ],
      "B": [
       "[gemini-pro] a person seated in a place the staging does not put them (카메라와 전택수의 손이 장원섭의 좌석 위치인 왼쪽에서 잘못 진입함)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1250,
      "verdict_ko": "이전 컷 스틸에 맞춰 정장 소매를 유지하려 했으나, 카메라가 장원섭의 자리(왼쪽)에 위치하여 '검사' 명패 앞면이 보이고 전택수의 손이 잘못된 좌석에서 등장하는 심각한 공간(Staging) 하드 비올레이션이 발생함.  ★위반: [gemini-pro] a person seated in a place the staging does not put them (카메라와 전택수의 손이 장원섭의 좌석 위치인 왼쪽에서 잘못 진입함)"
     },
     {
      "label": "A",
      "score": 1083,
      "verdict_ko": "A와 동일한 공간 왜곡(Staging) 하드 비올레이션이 발생했으며, 이전 컷 스틸로 고정(LOCKED)된 의상(정장) 지시를 무시하고 캐릭터 레퍼런스의 캐주얼 점퍼를 입혀 연속성을 완전히 훼손함.  ★위반: [gemini-pro] a person seated in a place the staging does not put them (카메라와 전택수의 손이 장원섭의 좌석 위치인 왼쪽에서 잘못 진입함)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1093,
     "B": 1256
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S54sh7"
  }
 },
 "S54sh12::cine": {
  "applied": true,
  "fingerprint": "37203dfc39ec3d87ab9bb5dd273113d83d5188844b0b28d4b33220fb63208d8a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S54sh12_sel.png",
  "source_sha256": "bbe63c3f674667f6b54e2a359f5f4b4363685e77a0da6dc16d7870733138acdf",
  "file": "S54sh12_cine.png",
  "latency_ms": 13529
 },
 "S54sh18::signage": {
  "fp": "0f22f851004f7563",
  "inscriptions": []
 },
 "S54sh18": {
  "input_fingerprint": "515b5606d59f4ea8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 허공에서 꽉 맞잡은 장원섭과 전택수의 두 손 클로즈업\n\nLOCATION (lock): Across the low table in the sofa meeting area of the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Near the table centerline, the camera completes a shallow downward tilt onto 장원섭 and 전택수's hands locking firmly in midair above the table. Their forearms enter from opposite sides of frame and converge at center, while the table and previously placed envelope remain secondary scale references below.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: table in front of the sofas (Directly below the handshake); used as Surface below the clasped hands, preserving depth and realistic scale; worn letter envelope (Resting on the table) — Its broad side remains visible at a shallow downward angle; used as Secondary evidence element remaining below the handshake.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic office daylight is rendered in restrained neutral-to-cool tones, emphasizing the joined hands without a celebratory gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa area, table, daylight, and both men's sleeves. Exclude the letter envelope as the focal element and show only their clasped hands above the table.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The inmate's letter and its envelope remain on the table after Wonseop has read it while he and Taeksu shake hands.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 허공에서 꽉 맞잡은 장원섭과 전택수의 두 손 클로즈업\n\nLOCATION (lock): Across the low table in the sofa meeting area of the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Near the table centerline, the camera completes a shallow downward tilt onto 장원섭 and 전택수's hands locking firmly in midair above the table. Their forearms enter from opposite sides of frame and converge at center, while the table and previously placed envelope remain secondary scale references below.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: table in front of the sofas (Directly below the handshake); used as Surface below the clasped hands, preserving depth and realistic scale; worn letter envelope (Resting on the table) — Its broad side remains visible at a shallow downward angle; used as Secondary evidence element remaining below the handshake.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic office daylight is rendered in restrained neutral-to-cool tones, emphasizing the joined hands without a celebratory gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa area, table, daylight, and both men's sleeves. Exclude the letter envelope as the focal element and show only their clasped hands above the table.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The inmate's letter and its envelope remain on the table after Wonseop has read it while he and Taeksu shake hands.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 허공에서 꽉 맞잡은 장원섭과 전택수의 두 손 클로즈업\n\nLOCATION (lock): Across the low table in the sofa meeting area of the prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Near the table centerline, the camera completes a shallow downward tilt onto 장원섭 and 전택수's hands locking firmly in midair above the table. Their forearms enter from opposite sides of frame and converge at center, while the table and previously placed envelope remain secondary scale references below.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: table in front of the sofas (Directly below the handshake); used as Surface below the clasped hands, preserving depth and realistic scale; worn letter envelope (Resting on the table) — Its broad side remains visible at a shallow downward angle; used as Secondary evidence element remaining below the handshake.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic office daylight is rendered in restrained neutral-to-cool tones, emphasizing the joined hands without a celebratory gloss.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office sofa area, table, daylight, and both men's sleeves. Exclude the letter envelope as the focal element and show only their clasped hands above the table.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The inmate's letter and its envelope remain on the table after Wonseop has read it while he and Taeksu shake hands.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.375,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.625
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.625
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1625,
   "A": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "두 인물의 지정된 레퍼런스 의상(회색 정장과 경찰 재킷)을 정확히 반영하였으며, 요구된 클로즈업 구도와 소품 디테일을 성공적으로 구현했습니다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "구도와 소품의 텍스트는 훌륭하게 묘사되었으나, 전택수의 지정된 의상(경찰 재킷)을 전혀 반영하지 못했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S54sh12_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 중앙의 맞잡은 두 손의 엄지손가락과 호구(엄지와 검지 사이)가 명확한 형태 없이 한 덩어리의 살로 융합되어 있습니다.",
     "fix_en": "Redraw the clasped hands to clearly separate the thumbs and webbing into correct human anatomy. Preserve the sleeves, watch, table, envelope, nameplate, and background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 컷에서 좌측 좌석을 향해 비스듬히 놓여 있던 '검사' 명패가 카메라 렌즈를 정면으로 바라보도록 테이블 위에서 회전되어 있습니다.",
     "fix_en": "Angle the nameplate diagonally toward the left seat. Preserve the hands, envelope, table, and background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "오른쪽 상단에 상반신과 넥타이가 넓게 보여 양옆에서 팔뚝만 모여드는 손 인서트 클로즈업이 아님",
     "fix_en": "Remove the right torso and tie, leaving only the sleeve entering the frame; fill the removed area with the blurred background. Preserve the hands, table, and left arm.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "하단 중앙 봉투가 화면을 크게 차지해 부차 참조가 아니라 맞잡은 손과 대등한 초점이 됨",
     "fix_en": "Reduce the envelope's size and place it lower. Preserve the hands, table, nameplate, and background.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 중앙의 맞잡은 두 손의 엄지손가락과 호구(엄지와 검지 사이)가 명확한 형태 없이 한 덩어리의 살로 융합되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 우측 끝 장원섭의 팔 아래 허공에 넥타이의 일부가 신체 구조와 맞지 않게 분리된 채 떠 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 컷에서 좌측 좌석을 향해 비스듬히 놓여 있던 '검사' 명패가 카메라 렌즈를 정면으로 바라보도록 테이블 위에서 회전되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 상단에 상반신과 넥타이가 넓게 보여 양옆에서 팔뚝만 모여드는 손 인서트 클로즈업이 아님",
     "severity": "major"
    },
    {
     "issue_ko": "하단 중앙 봉투가 화면을 크게 차지해 부차 참조가 아니라 맞잡은 손과 대등한 초점이 됨",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 컷에서 좌측 좌석을 향해 비스듬히 놓여 있던 '검사' 명패가 카메라 렌즈를 정면으로 바라보도록 테이블 위에서 회전되어 있습니다.",
    "fix_en": "Angle the nameplate diagonally toward the left seat. Preserve the hands, envelope, table, and background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "오른쪽 상단에 상반신과 넥타이가 넓게 보여 양옆에서 팔뚝만 모여드는 손 인서트 클로즈업이 아님",
    "fix_en": "Remove the right torso and tie, leaving only the sleeve entering the frame; fill the removed area with the blurred background. Preserve the hands, table, and left arm.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "하단 중앙 봉투가 화면을 크게 차지해 부차 참조가 아니라 맞잡은 손과 대등한 초점이 됨",
    "fix_en": "Reduce the envelope's size and place it lower. Preserve the hands, table, nameplate, and background.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the clasped hands to clearly separate the thumbs and webbing into correct human anatomy. Preserve the sleeves, watch, table, envelope, nameplate, and background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트가 요구한 두 사람의 악수 장면을 중앙에 잘 배치했으며, 이전 컷의 소매 디테일과 새 캐릭터의 의상을 정확히 반영하고 봉투를 탁자 위에 자연스럽게 남겨둔 완벽한 결과물입니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "프롬프트의 지시사항을 전혀 반영하지 않고 레퍼런스로 제공된 이전 컷을 그대로 복사하여 악수 장면을 누락한 심각한 실패작입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "양쪽 프레임에서 뻗어 나온 두 팔이 화면 중앙 탁자 위 허공에서 서로를 향해 뻗어 악수를 하며 손을 꽉 맞잡고 있음.",
      "built_space": "검사실 소파 미팅 영역의 낮고 넓은 탁자가 화면의 주요 배경을 구성하며, 그 위에 '검사' 명패가 올바른 원근으로 놓여 있음.",
      "entities": "왼쪽 팔은 이전 컷의 남색 정장 소매와 갈색 가죽 시계를 정확히 유지하고, 오른쪽 팔은 장원섭의 레퍼런스에 맞는 짙은 회색 정장과 연하늘색 셔츠를 착용함. 탁자 위에는 '친전'이라고 적힌 낡은 봉투가 지시대로 놓여 있음.",
      "hard_violations": [],
      "physics": "맞잡은 두 손은 양쪽에서 뻗어 나온 팔의 힘으로 허공에 안정적으로 지탱되고 있으며, 편지 봉투와 명패는 탁자 표면 위에 중력에 맞게 놓여 있음."
     },
     {
      "label": "B",
      "direction": "화면 왼쪽에서 뻗어 나온 한 손만이 봉투를 향해 있으며, 맞잡은 두 손의 방향성이나 악수의 초점이 전혀 존재하지 않음.",
      "built_space": "검사실 소파 앞 탁자와 '검사' 명패가 보임.",
      "entities": "지시된 장원섭의 손이나 의상이 전혀 등장하지 않으며, 바닥에 놓여야 할 봉투를 이전 컷과 동일하게 손에 들고 있음.",
      "hard_violations": [],
      "physics": "손이 봉투를 쥐고 공중에 떠 있으며, 지지 방식 자체는 물리적으로 가능하나 프롬프트의 상황과는 전혀 맞지 않음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트가 요구한 두 사람의 악수 장면을 중앙에 잘 배치했으며, 이전 컷의 소매 디테일과 새 캐릭터의 의상을 정확히 반영하고 봉투를 탁자 위에 자연스럽게 남겨둔 완벽한 결과물입니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "프롬프트의 지시사항을 전혀 반영하지 않고 레퍼런스로 제공된 이전 컷을 그대로 복사하여 악수 장면을 누락한 심각한 실패작입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "양쪽 프레임에서 뻗어 나온 두 팔이 화면 중앙 탁자 위 허공에서 서로를 향해 뻗어 악수를 하며 손을 꽉 맞잡고 있음.",
      "built_space": "검사실 소파 미팅 영역의 낮고 넓은 탁자가 화면의 주요 배경을 구성하며, 그 위에 '검사' 명패가 올바른 원근으로 놓여 있음.",
      "entities": "왼쪽 팔은 이전 컷의 남색 정장 소매와 갈색 가죽 시계를 정확히 유지하고, 오른쪽 팔은 장원섭의 레퍼런스에 맞는 짙은 회색 정장과 연하늘색 셔츠를 착용함. 탁자 위에는 '친전'이라고 적힌 낡은 봉투가 지시대로 놓여 있음.",
      "hard_violations": [],
      "physics": "맞잡은 두 손은 양쪽에서 뻗어 나온 팔의 힘으로 허공에 안정적으로 지탱되고 있으며, 편지 봉투와 명패는 탁자 표면 위에 중력에 맞게 놓여 있음."
     },
     {
      "label": "B",
      "direction": "화면 왼쪽에서 뻗어 나온 한 손만이 봉투를 향해 있으며, 맞잡은 두 손의 방향성이나 악수의 초점이 전혀 존재하지 않음.",
      "built_space": "검사실 소파 앞 탁자와 '검사' 명패가 보임.",
      "entities": "지시된 장원섭의 손이나 의상이 전혀 등장하지 않으며, 바닥에 놓여야 할 봉투를 이전 컷과 동일하게 손에 들고 있음.",
      "hard_violations": [],
      "physics": "손이 봉투를 쥐고 공중에 떠 있으며, 지지 방식 자체는 물리적으로 가능하나 프롬프트의 상황과는 전혀 맞지 않음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "프롬프트의 지시대로 테이블 위 허공에서 맞잡은 두 손을 정확하게 클로즈업으로 담아냈으며, 이전 컷의 서류 봉투를 테이블 위에 자연스럽게 배치하여 디테일을 훌륭히 살렸습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "두 사람의 맞잡은 손이라는 핵심 요구 사항을 무시하고 레퍼런스 이미지를 거의 그대로 복제하여 봉투를 쥔 한 손만 보여주어 지시를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "한 손이 서류 봉투를 잡고 있음.",
      "built_space": "테이블, 뒤쪽의 소파, 명패가 이전 컷과 동일한 위치에 있음.",
      "entities": "어두운 수트와 시계를 착용한 한 사람의 팔만 보이며, 서류 봉투가 화면의 중심에 있음. 장원섭의 팔은 보이지 않음.",
      "hard_violations": [
       "핵심 액션인 '두 사람의 맞잡은 손' 누락"
      ],
      "physics": "손이 봉투를 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "두 손이 테이블 위 허공에서 악수하며 맞잡고 있음.",
      "built_space": "테이블, 소파, 명패가 적절한 위치에 있으며 배경을 구성함.",
      "entities": "좌측은 어두운 수트와 시계를 찬 전택수의 팔(이전 컷과 일치), 우측은 회색 수트를 입은 장원섭의 팔(레퍼런스와 일치)이 보임. 테이블 위에는 봉투가 놓여 있음.",
      "hard_violations": [],
      "physics": "두 손이 허공에서 서로를 단단히 지탱하며 악수하고 있으며, 봉투는 테이블 위에 안정적으로 놓여 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트의 지시대로 테이블 위 허공에서 맞잡은 두 손을 정확하게 클로즈업으로 담아냈으며, 이전 컷의 서류 봉투를 테이블 위에 자연스럽게 배치하여 디테일을 훌륭히 살렸습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "두 사람의 맞잡은 손이라는 핵심 요구 사항을 무시하고 레퍼런스 이미지를 거의 그대로 복제하여 봉투를 쥔 한 손만 보여주어 지시를 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "한 손이 서류 봉투를 잡고 있음.",
      "built_space": "테이블, 뒤쪽의 소파, 명패가 이전 컷과 동일한 위치에 있음.",
      "entities": "어두운 수트와 시계를 착용한 한 사람의 팔만 보이며, 서류 봉투가 화면의 중심에 있음. 장원섭의 팔은 보이지 않음.",
      "hard_violations": [
       "핵심 액션인 '두 사람의 맞잡은 손' 누락"
      ],
      "physics": "손이 봉투를 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "두 손이 테이블 위 허공에서 악수하며 맞잡고 있음.",
      "built_space": "테이블, 소파, 명패가 적절한 위치에 있으며 배경을 구성함.",
      "entities": "좌측은 어두운 수트와 시계를 찬 전택수의 팔(이전 컷과 일치), 우측은 회색 수트를 입은 장원섭의 팔(레퍼런스와 일치)이 보임. 테이블 위에는 봉투가 놓여 있음.",
      "hard_violations": [],
      "physics": "두 손이 허공에서 서로를 단단히 지탱하며 악수하고 있으며, 봉투는 테이블 위에 안정적으로 놓여 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 18,
     "B": 3
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S54sh12"
  }
 },
 "S54sh18::cine": {
  "applied": true,
  "fingerprint": "1008928e49b017d0d6e8aa8de0fa814ef991dadef4bc4da1f99332bea1b67f11",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S54sh18_sel.png",
  "source_sha256": "d2b43a84aeb40b1de8d80f4e9b7753a12ae6c07e123cd3167170ac50c582025a",
  "file": "S54sh18_cine.png",
  "latency_ms": 11043
 },
 "S55sh1::signage": {
  "fp": "aa083bfeed884913",
  "inscriptions": [
   {
    "surface_native": "복도 벽면 표지판",
    "text_native": "정숙",
    "reason_ko": "교도소 복도의 엄숙하고 조용한 분위기를 사실적으로 전달하기 위해 벽면에 '정숙'이라는 안내판이 필요합니다."
   }
  ]
 },
 "era_assess::1f7358e6d83fcf7f": {
  "subjects": [
   {
    "subject_native": "2000년대 대한민국 교도관 제복",
    "search_terms_native": [
     "교도관 제복 2000년대",
     "교도관 유니폼 한국",
     "교도소 교도관 실내"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 교도관의 제복 디자인과 색상은 시대별로 고유하며 서구권의 일반적인 보안관이나 경찰 제복과는 시각적으로 크게 다릅니다."
   },
   {
    "subject_native": "2000년대 대한민국 교도소 내부 복도",
    "search_terms_native": [
     "교도소 복도 한국",
     "교도소 내부 복도",
     "한국 교도소 실내"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 작성해야 하며, 영어 등 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 교도소의 복도 벽면 도색, 수감실 철문 양식, 경고 표지판 등은 서구식 감옥과 다른 독특한 한국적 분위기를 지니고 있습니다."
   }
  ]
 },
 "era_ref::c840d67c85b0c7d6": {
  "subject": "2000년대 대한민국 교도관 제복",
  "terms": [
   "교도관 제복 2000년대",
   "교도관 유니폼 한국",
   "교도소 교도관 실내"
  ],
  "queries": [
   [
    "교도관 제복 2000년대 대한민국 교도관 유니폼 한국",
    "2000년대 대한민국 교도소 교도관 실내 근무복"
   ],
   [
    "대한민국 교도관 구형 제복 2000년대 법무부",
    "한국 교도관 겨울 제복 교도소 내부 2000년대"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://img3.daumcdn.net/thumb/R658x0.q70/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fnews%2F201812%2F05%2Fsegye%2F20181205130127254knnu.jpg",
  "picked_reason_ko": "사진 4는 2001~2017년 대한민국 교도관의 남녀 제복을 전신으로 선명하게 보여 주어 색상, 재단, 계급장, 표장, 넥타이와 허리띠까지 가장 정확히 읽을 수 있다.",
  "sha256": "2f6c1e41c1a9713e0f653c3fc9438ab2d9126949dfdb69c76fd31de2cb834224",
  "file": "eraref_c840d67c85b0c7d6.png"
 },
 "S55sh1::bgfirst_bg": {
  "input_fingerprint": "141d224071f1123e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 광주교도소 교도관의 등 뒤로 길고 차가운 복도를 나란히 걷는 mid-stride 자세로 한쪽 발이 바닥에서 막 떨어진 서의용, 나상혁, 검찰 수사관의 전경\n\nLOCATION (lock): Along a long interior corridor inside the prison, following a correctional officer.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At waist height several paces behind 광주교도소 교도관, the tracking camera follows the group down the long corridor with the guard anchoring the near foreground. Beyond him, 서의용, 나상혁, and 검찰 수사관 advance in a loose side-by-side formation, each caught at a different phase of the same stride so their lifted feet, shoulder angles, and spacing do not appear duplicated.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 광주교도소 교도관 leading the group in the lower-center of the frame, foreground, moves toward far end of the corridor; 서의용 following the guard in the middle-left of the frame, midground, moves toward far end of the corridor; 나상혁 within the moving group in the middle-right of the frame, midground, moves toward far end of the corridor.\n- KEY BACKGROUND ELEMENTS: long prison corridor (Occupied by the moving guard and investigators) — The corridor recedes away from the camera in the group's direction of travel; used as Strong receding spatial axis surrounding and directing the moving formation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the prison corridor is held in a subdued neutral-to-cool procedural palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 교도관 제복: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 광주교도소 교도관의 등 뒤로 길고 차가운 복도를 나란히 걷는 mid-stride 자세로 한쪽 발이 바닥에서 막 떨어진 서의용, 나상혁, 검찰 수사관의 전경\n\nLOCATION (lock): Along a long interior corridor inside the prison, following a correctional officer.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At waist height several paces behind 광주교도소 교도관, the tracking camera follows the group down the long corridor with the guard anchoring the near foreground. Beyond him, 서의용, 나상혁, and 검찰 수사관 advance in a loose side-by-side formation, each caught at a different phase of the same stride so their lifted feet, shoulder angles, and spacing do not appear duplicated.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 광주교도소 교도관 leading the group in the lower-center of the frame, foreground, moves toward far end of the corridor; 서의용 following the guard in the middle-left of the frame, midground, moves toward far end of the corridor; 나상혁 within the moving group in the middle-right of the frame, midground, moves toward far end of the corridor.\n- KEY BACKGROUND ELEMENTS: long prison corridor (Occupied by the moving guard and investigators) — The corridor recedes away from the camera in the group's direction of travel; used as Strong receding spatial axis surrounding and directing the moving formation.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the prison corridor is held in a subdued neutral-to-cool procedural palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 교도관 제복: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S55sh1__bgfirst_bg.png",
  "asset_id": "249270fd-de27-46a5-b130-6522d6ee6469",
  "input_asset_ids": [
   "dc79d6e7-8c65-48bf-a8ab-1a428321e4c0",
   "83fe623c-d896-4882-b478-030f63ffb76f"
  ],
  "era_research": {
   "subject": "2000년대 대한민국 교도관 제복",
   "queries": [
    [
     "교도관 제복 2000년대 대한민국 교도관 유니폼 한국",
     "2000년대 대한민국 교도소 교도관 실내 근무복"
    ],
    [
     "대한민국 교도관 구형 제복 2000년대 법무부",
     "한국 교도관 겨울 제복 교도소 내부 2000년대"
    ]
   ],
   "picked_url": "https://img3.daumcdn.net/thumb/R658x0.q70/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fnews%2F201812%2F05%2Fsegye%2F20181205130127254knnu.jpg",
   "sha256": "2f6c1e41c1a9713e0f653c3fc9438ab2d9126949dfdb69c76fd31de2cb834224",
   "file": "eraref_c840d67c85b0c7d6.png"
  }
 },
 "S55sh1": {
  "input_fingerprint": "f9a849574e1839d6",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 광주교도소 교도관의 등 뒤로 길고 차가운 복도를 나란히 걷는 mid-stride 자세로 한쪽 발이 바닥에서 막 떨어진 서의용, 나상혁, 검찰 수사관의 전경\n\nLOCATION (lock): Along a long interior corridor inside the prison, following a correctional officer. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height several paces behind 광주교도소 교도관, the tracking camera follows the group down the long corridor with the guard anchoring the near foreground. Beyond him, 서의용, 나상혁, and 검찰 수사관 advance in a loose side-by-side formation, each caught at a different phase of the same stride so their lifted feet, shoulder angles, and spacing do not appear duplicated.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 광주교도소 교도관 leading the group in the lower-center of the frame, foreground, moves toward far end of the corridor; 서의용 following the guard in the middle-left of the frame, midground, moves toward far end of the corridor; 나상혁 within the moving group in the middle-right of the frame, midground, moves toward far end of the corridor.\n- KEY BACKGROUND ELEMENTS: long prison corridor (Occupied by the moving guard and investigators) — The corridor recedes away from the camera in the group's direction of travel; used as Strong receding spatial axis surrounding and directing the moving formation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the prison corridor is held in a subdued neutral-to-cool procedural palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리); 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 복도 벽면 표지판: \"정숙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 광주교도소 교도관의 등 뒤로 길고 차가운 복도를 나란히 걷는 mid-stride 자세로 한쪽 발이 바닥에서 막 떨어진 서의용, 나상혁, 검찰 수사관의 전경\n\nLOCATION (lock): Along a long interior corridor inside the prison, following a correctional officer. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height several paces behind 광주교도소 교도관, the tracking camera follows the group down the long corridor with the guard anchoring the near foreground. Beyond him, 서의용, 나상혁, and 검찰 수사관 advance in a loose side-by-side formation, each caught at a different phase of the same stride so their lifted feet, shoulder angles, and spacing do not appear duplicated.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 광주교도소 교도관 leading the group in the lower-center of the frame, foreground, moves toward far end of the corridor; 서의용 following the guard in the middle-left of the frame, midground, moves toward far end of the corridor; 나상혁 within the moving group in the middle-right of the frame, midground, moves toward far end of the corridor.\n- KEY BACKGROUND ELEMENTS: long prison corridor (Occupied by the moving guard and investigators) — The corridor recedes away from the camera in the group's direction of travel; used as Strong receding spatial axis surrounding and directing the moving formation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the prison corridor is held in a subdued neutral-to-cool procedural palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리); 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 복도 벽면 표지판: \"정숙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 광주교도소 교도관의 등 뒤로 길고 차가운 복도를 나란히 걷는 mid-stride 자세로 한쪽 발이 바닥에서 막 떨어진 서의용, 나상혁, 검찰 수사관의 전경\n\nLOCATION (lock): Along a long interior corridor inside the prison, following a correctional officer. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height several paces behind 광주교도소 교도관, the tracking camera follows the group down the long corridor with the guard anchoring the near foreground. Beyond him, 서의용, 나상혁, and 검찰 수사관 advance in a loose side-by-side formation, each caught at a different phase of the same stride so their lifted feet, shoulder angles, and spacing do not appear duplicated.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 광주교도소 교도관 leading the group in the lower-center of the frame, foreground, moves toward far end of the corridor; 서의용 following the guard in the middle-left of the frame, midground, moves toward far end of the corridor; 나상혁 within the moving group in the middle-right of the frame, midground, moves toward far end of the corridor.\n- KEY BACKGROUND ELEMENTS: long prison corridor (Occupied by the moving guard and investigators) — The corridor recedes away from the camera in the group's direction of travel; used as Strong receding spatial axis surrounding and directing the moving formation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the prison corridor is held in a subdued neutral-to-cool procedural palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리); 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 복도 벽면 표지판: \"정숙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S55sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S55sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    },
    {
     "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:831079>"
    },
    {
     "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:861801>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L41B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:929743>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:747016>"
    },
    {
     "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:831079>"
    },
    {
     "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:861801>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "서의용 캐릭터가 누락되고 지시문에 없는 정장 차림의 인물이 등장했으며, 세 인물의 얼굴이 복제되었고 이동 방향이 프롬프트와 반대입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "서의용이 없고 정장 차림의 인물이 추가되었으며 얼굴이 복제된 데다, 이동 방향과 복도의 건축 구조(우측 창문)가 지시문 및 레퍼런스와 어긋납니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "교도관은 카메라를 등지고 복도 안쪽을 향해 걷고 있으며, 나머지 세 남성은 카메라와 교도관을 향해 걸어오고 있음.",
      "built_space": "양쪽에 감방 철문이 늘어선 실내 복도 구조로 레퍼런스와 유사하며, 우측 상단 벽면에 '정숙' 표지판이 부착되어 있음.",
      "entities": "교도관의 의상은 레퍼런스와 일치함. 마주 오는 세 명 중 좌우측 인물은 나상혁과 검찰 수사관의 의상과 유사하나, 서의용(비니, 가죽재킷)은 없고 대신 정장 차림의 인물이 중앙에 추가됨. 세 남성의 얼굴이 동일하게 복제됨.",
      "hard_violations": [
       "invented people",
       "duplicated faces"
      ],
      "physics": "모든 인물의 발이 복도 바닥에 닿아 지탱되고 있으며, 걷는 자세가 자연스러움."
     },
     {
      "label": "B",
      "direction": "교도관은 카메라를 등지고 서 있으며, 세 남성은 카메라를 향해 정면으로 걸어오고 있음.",
      "built_space": "좌측에는 철문이 있으나 우측에는 쇠창살이 있는 창문이 늘어서 있어 레퍼런스의 복도 구조와 다름. 좌측 벽면에 '정숙' 표지판이 있음.",
      "entities": "교도관의 의상은 일치함. 마주 오는 세 남성 중 서의용은 완전히 누락되었고, 중앙에 지시문에 없는 정장 차림의 인물이 등장함. 세 명 모두 완전히 동일한 얼굴을 하고 있음.",
      "hard_violations": [
       "invented people",
       "duplicated faces"
      ],
      "physics": "모든 인물이 바닥에 발을 딛고 서 있거나 걷고 있어 물리적으로 안정됨."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "서의용 캐릭터가 누락되고 지시문에 없는 정장 차림의 인물이 등장했으며, 세 인물의 얼굴이 복제되었고 이동 방향이 프롬프트와 반대입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "서의용이 없고 정장 차림의 인물이 추가되었으며 얼굴이 복제된 데다, 이동 방향과 복도의 건축 구조(우측 창문)가 지시문 및 레퍼런스와 어긋납니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "교도관은 카메라를 등지고 복도 안쪽을 향해 걷고 있으며, 나머지 세 남성은 카메라와 교도관을 향해 걸어오고 있음.",
      "built_space": "양쪽에 감방 철문이 늘어선 실내 복도 구조로 레퍼런스와 유사하며, 우측 상단 벽면에 '정숙' 표지판이 부착되어 있음.",
      "entities": "교도관의 의상은 레퍼런스와 일치함. 마주 오는 세 명 중 좌우측 인물은 나상혁과 검찰 수사관의 의상과 유사하나, 서의용(비니, 가죽재킷)은 없고 대신 정장 차림의 인물이 중앙에 추가됨. 세 남성의 얼굴이 동일하게 복제됨.",
      "hard_violations": [
       "invented people",
       "duplicated faces"
      ],
      "physics": "모든 인물의 발이 복도 바닥에 닿아 지탱되고 있으며, 걷는 자세가 자연스러움."
     },
     {
      "label": "B",
      "direction": "교도관은 카메라를 등지고 서 있으며, 세 남성은 카메라를 향해 정면으로 걸어오고 있음.",
      "built_space": "좌측에는 철문이 있으나 우측에는 쇠창살이 있는 창문이 늘어서 있어 레퍼런스의 복도 구조와 다름. 좌측 벽면에 '정숙' 표지판이 있음.",
      "entities": "교도관의 의상은 일치함. 마주 오는 세 남성 중 서의용은 완전히 누락되었고, 중앙에 지시문에 없는 정장 차림의 인물이 등장함. 세 명 모두 완전히 동일한 얼굴을 하고 있음.",
      "hard_violations": [
       "invented people",
       "duplicated faces"
      ],
      "physics": "모든 인물이 바닥에 발을 딛고 서 있거나 걷고 있어 물리적으로 안정됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 667,
      "verdict_ko": "서의용이 누락되고 프롬프트에 없는 회색 정장의 인물이 추가되었으며, 일행이 복도 끝이 아닌 카메라를 향해 걷고 있어 연출 방향이 어긋납니다.  ★위반: [gemini-pro] 발명된 사람(회색 정장을 입은 남성) / [openrouter:x-ai/grok-4.6] 샷에 없는 회색 정장 남성 추가 / [openrouter:x-ai/grok-4.6] 서의용(참조 의상·얼굴) 미등장 / [openrouter:x-ai/grok-4.6] 세 사람의 진행이 복도 먼 끝이 아니라 카메라 쪽"
     },
     {
      "label": "B",
      "score": 500,
      "verdict_ko": "서의용 누락 및 임의의 인물 추가뿐만 아니라, 가운데 남성 뒤에 여분의 팔다리가 노출되는 심각한 신체 중복 오류가 있습니다.  ★위반: [gemini-pro] 발명된 사람(회색 정장을 입은 남성) / [gemini-pro] 중복되거나 여분의 신체(가운데 남성 뒤로 보이는 또 다른 팔과 다리) / [openrouter:x-ai/grok-4.6] 샷에 없는 회색 정장 남성 추가 / [openrouter:x-ai/grok-4.6] 서의용(참조 의상·얼굴) 미등장 / [openrouter:x-ai/grok-4.6] 세 사람의 진행이 복도 먼 끝이 아니라 카메라 쪽"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.75
     },
     "adjusted": {
      "A": 0.667,
      "B": 0.5
     },
     "violations": {
      "A": [
       "[gemini-pro] 발명된 사람(회색 정장을 입은 남성)",
       "[openrouter:x-ai/grok-4.6] 샷에 없는 회색 정장 남성 추가",
       "[openrouter:x-ai/grok-4.6] 서의용(참조 의상·얼굴) 미등장",
       "[openrouter:x-ai/grok-4.6] 세 사람의 진행이 복도 먼 끝이 아니라 카메라 쪽"
      ],
      "B": [
       "[gemini-pro] 발명된 사람(회색 정장을 입은 남성)",
       "[gemini-pro] 중복되거나 여분의 신체(가운데 남성 뒤로 보이는 또 다른 팔과 다리)",
       "[openrouter:x-ai/grok-4.6] 샷에 없는 회색 정장 남성 추가",
       "[openrouter:x-ai/grok-4.6] 서의용(참조 의상·얼굴) 미등장",
       "[openrouter:x-ai/grok-4.6] 세 사람의 진행이 복도 먼 끝이 아니라 카메라 쪽"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.333,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 667,
      "verdict_ko": "서의용이 누락되고 프롬프트에 없는 회색 정장의 인물이 추가되었으며, 일행이 복도 끝이 아닌 카메라를 향해 걷고 있어 연출 방향이 어긋납니다.  ★위반: [gemini-pro] 발명된 사람(회색 정장을 입은 남성) / [openrouter:x-ai/grok-4.6] 샷에 없는 회색 정장 남성 추가 / [openrouter:x-ai/grok-4.6] 서의용(참조 의상·얼굴) 미등장 / [openrouter:x-ai/grok-4.6] 세 사람의 진행이 복도 먼 끝이 아니라 카메라 쪽"
     },
     {
      "label": "A",
      "score": 500,
      "verdict_ko": "서의용 누락 및 임의의 인물 추가뿐만 아니라, 가운데 남성 뒤에 여분의 팔다리가 노출되는 심각한 신체 중복 오류가 있습니다.  ★위반: [gemini-pro] 발명된 사람(회색 정장을 입은 남성) / [gemini-pro] 중복되거나 여분의 신체(가운데 남성 뒤로 보이는 또 다른 팔과 다리) / [openrouter:x-ai/grok-4.6] 샷에 없는 회색 정장 남성 추가 / [openrouter:x-ai/grok-4.6] 서의용(참조 의상·얼굴) 미등장 / [openrouter:x-ai/grok-4.6] 세 사람의 진행이 복도 먼 끝이 아니라 카메라 쪽"
     }
    ],
    "all_candidates_fail": true
   },
   "combined": {
    "totals": {
     "A": 502,
     "B": 669
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "교도관은 카메라를 등지고 복도 안쪽을 향해 걷고 있으며, 나머지 세 남성은 카메라와 교도관을 향해 걸어오고 있음.",
    "built_space": "양쪽에 감방 철문이 늘어선 실내 복도 구조로 레퍼런스와 유사하며, 우측 상단 벽면에 '정숙' 표지판이 부착되어 있음.",
    "entities": "교도관의 의상은 레퍼런스와 일치함. 마주 오는 세 명 중 좌우측 인물은 나상혁과 검찰 수사관의 의상과 유사하나, 서의용(비니, 가죽재킷)은 없고 대신 정장 차림의 인물이 중앙에 추가됨. 세 남성의 얼굴이 동일하게 복제됨.",
    "hard_violations": [
     "invented people",
     "duplicated faces"
    ],
    "physics": "모든 인물의 발이 복도 바닥에 닿아 지탱되고 있으며, 걷는 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "교도관은 카메라를 등지고 서 있으며, 세 남성은 카메라를 향해 정면으로 걸어오고 있음.",
    "built_space": "좌측에는 철문이 있으나 우측에는 쇠창살이 있는 창문이 늘어서 있어 레퍼런스의 복도 구조와 다름. 좌측 벽면에 '정숙' 표지판이 있음.",
    "entities": "교도관의 의상은 일치함. 마주 오는 세 남성 중 서의용은 완전히 누락되었고, 중앙에 지시문에 없는 정장 차림의 인물이 등장함. 세 명 모두 완전히 동일한 얼굴을 하고 있음.",
    "hard_violations": [
     "invented people",
     "duplicated faces"
    ],
    "physics": "모든 인물이 바닥에 발을 딛고 서 있거나 걷고 있어 물리적으로 안정됨."
   }
  ],
  "totals": {
   "A": 502,
   "B": 669
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "서의용 캐릭터가 누락되고 지시문에 없는 정장 차림의 인물이 등장했으며, 세 인물의 얼굴이 복제되었고 이동 방향이 프롬프트와 반대입니다."
   },
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "서의용이 없고 정장 차림의 인물이 추가되었으며 얼굴이 복제된 데다, 이동 방향과 복도의 건축 구조(우측 창문)가 지시문 및 레퍼런스와 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L41B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   },
   {
    "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:831079>"
   },
   {
    "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:861801>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "세 남자가 복도 끝이 아닌 카메라와 교도관을 향해 정면으로 걸어오고 있습니다.",
     "fix_en": "Turn the three men in the midground around so their backs are to the camera, walking toward the far end of the corridor to follow the guard. Preserve the guard in the foreground and the corridor architecture.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "비니와 가죽 재킷을 착용한 지정 캐릭터 '서의용'이 프레임에 존재하지 않습니다.",
     "fix_en": "Replace the man on the left with Seo Eui-yong, a Korean man wearing a navy beanie, dark leather jacket, grey t-shirt, and faded jeans. Preserve the other two men, the guard in the foreground, and the corridor setting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "프롬프트에 없는 정장과 넥타이 차림의 인물이 중앙에 생성되었습니다.",
     "fix_en": "Change the middle man's clothing from a suit and tie to a navy Harrington jacket over a light blue button-down shirt. Preserve his face, the other men, the guard, and the hallway environment.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "걸어오는 세 남자의 얼굴이 모두 동일한 인물의 얼굴로 중복 생성되었습니다.",
     "fix_en": "Modify the faces of the left and middle men so each has a distinct facial structure, avoiding duplicate identical faces in the group. Preserve their current clothing, poses, the guard, and the corridor.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "복도 우측 벽면에 로케이션 레퍼런스에 없는 대형 쇠창살 창문들이 연속해서 생성되었습니다.",
     "fix_en": "Change the large barred windows on the right wall into a solid wall with pale blue paint on the lower half and closed metal doors, matching the left side. Preserve all people in the corridor.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "세 남자가 복도 끝이 아닌 카메라와 교도관을 향해 정면으로 걸어오고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "비니와 가죽 재킷을 착용한 지정 캐릭터 '서의용'이 프레임에 존재하지 않습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "프롬프트에 없는 정장과 넥타이 차림의 인물이 중앙에 생성되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "걸어오는 세 남자의 얼굴이 모두 동일한 인물의 얼굴로 중복 생성되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "인물들이 카메라 렌즈 방향을 정면으로 응시하며 일렬로 포즈를 취한 듯한 구도입니다.",
     "severity": "major"
    },
    {
     "issue_ko": "복도 우측 벽면에 로케이션 레퍼런스에 없는 대형 쇠창살 창문들이 연속해서 생성되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "중경 가운데 정장 남성은 서의용이 아니며 가죽점퍼·비니의 서의용이 프레임에 없다",
     "severity": "critical"
    },
    {
     "issue_ko": "중경 세 사람이 교도관과 같은 방향으로 걷지 않고 카메라를 향해 걸어오며 렌즈를 보고 일렬로 정렬돼 있다",
     "severity": "major"
    },
    {
     "issue_ko": "복도의 문·창문·벽면 건축이 참조 장소 사진과 다르다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 6,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "복도 우측 벽면에 로케이션 레퍼런스에 없는 대형 쇠창살 창문들이 연속해서 생성되었습니다.",
    "fix_en": "Change the large barred windows on the right wall into a solid wall with pale blue paint on the lower half and closed metal doors, matching the left side. Preserve all people in the corridor.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 6,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Turn the three men in the midground around so their backs are to the camera, walking toward the far end of the corridor to follow the guard. Preserve the guard in the foreground and the corridor architecture.\n- Replace the man on the left with Seo Eui-yong, a Korean man wearing a navy beanie, dark leather jacket, grey t-shirt, and faded jeans. Preserve the other two men, the guard in the foreground, and the corridor setting.\n- Change the middle man's clothing from a suit and tie to a navy Harrington jacket over a light blue button-down shirt. Preserve his face, the other men, the guard, and the hallway environment.\n- Modify the faces of the left and middle men so each has a distinct facial structure, avoiding duplicate identical faces in the group. Preserve their current clothing, poses, the guard, and the corridor.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 인물들이 모두 등장하고 복장이 일치하나, 교도관이 카메라를 등지지 않고 정면을 향해 있어 구도 지시를 위반했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "서의용 캐릭터가 완전히 누락되고 프롬프트에 없는 인물이 추가되었으며, 인물들의 이동 방향이 반대로 연출되어 실격입니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "교도관은 카메라를 정면으로 바라보고 있으며, 세 명의 남성(서의용, 나상혁, 수사관)은 카메라를 등지고 복도 끝을 향해 걸어가고 있음.",
      "built_space": "양옆으로 다수의 문이 배치된 긴 실내 복도 구조임.",
      "entities": "교도관(정면, 얼굴과 유니폼 일치), 서의용(비니와 가죽 재킷 일치), 나상혁(남색 재킷 일치), 검찰 수사관(남색 가디건 일치). '정숙' 표지판은 보이지 않음.",
      "hard_violations": [
       "교도관이 복도 끝을 향하지 않고 카메라를 정면으로 바라보고 있어 '교도관의 등 뒤로'라는 카메라 위치 및 방향 지시 위반"
      ],
      "physics": "모든 인물의 발이 바닥 표면에 닿아 무게를 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "교도관은 복도 끝을 향하고 있으며, 세 남성은 카메라와 교도관을 향해 정면으로 걸어오고 있음.",
      "built_space": "좌측 벽면에 '정숙' 표지판과 철창이 있는 감방 문들이 보이고 긴 복도가 이어짐.",
      "entities": "교도관(뒷모습 일치), 나상혁(파란 재킷 일치), 검찰 수사관(남색 가디건 일치). 서의용이 누락되고 넥타이를 멘 정장 차림의 새로운 인물이 등장함. 벽에 '정숙' 텍스트 표지판이 있음.",
      "hard_violations": [
       "필수 캐릭터인 서의용 누락",
       "프롬프트에 명시되지 않은 정장 차림의 인물 추가",
       "세 인물이 복도 끝이 아닌 카메라 방향으로 걸어와 동선 지시 위반"
      ],
      "physics": "네 명 모두 바닥을 단단히 디디고 자연스럽게 걷는 자세를 유지함."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 인물들이 모두 등장하고 복장이 일치하나, 교도관이 카메라를 등지지 않고 정면을 향해 있어 구도 지시를 위반했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "서의용 캐릭터가 완전히 누락되고 프롬프트에 없는 인물이 추가되었으며, 인물들의 이동 방향이 반대로 연출되어 실격입니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "교도관은 카메라를 정면으로 바라보고 있으며, 세 명의 남성(서의용, 나상혁, 수사관)은 카메라를 등지고 복도 끝을 향해 걸어가고 있음.",
      "built_space": "양옆으로 다수의 문이 배치된 긴 실내 복도 구조임.",
      "entities": "교도관(정면, 얼굴과 유니폼 일치), 서의용(비니와 가죽 재킷 일치), 나상혁(남색 재킷 일치), 검찰 수사관(남색 가디건 일치). '정숙' 표지판은 보이지 않음.",
      "hard_violations": [
       "교도관이 복도 끝을 향하지 않고 카메라를 정면으로 바라보고 있어 '교도관의 등 뒤로'라는 카메라 위치 및 방향 지시 위반"
      ],
      "physics": "모든 인물의 발이 바닥 표면에 닿아 무게를 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "교도관은 복도 끝을 향하고 있으며, 세 남성은 카메라와 교도관을 향해 정면으로 걸어오고 있음.",
      "built_space": "좌측 벽면에 '정숙' 표지판과 철창이 있는 감방 문들이 보이고 긴 복도가 이어짐.",
      "entities": "교도관(뒷모습 일치), 나상혁(파란 재킷 일치), 검찰 수사관(남색 가디건 일치). 서의용이 누락되고 넥타이를 멘 정장 차림의 새로운 인물이 등장함. 벽에 '정숙' 텍스트 표지판이 있음.",
      "hard_violations": [
       "필수 캐릭터인 서의용 누락",
       "프롬프트에 명시되지 않은 정장 차림의 인물 추가",
       "세 인물이 복도 끝이 아닌 카메라 방향으로 걸어와 동선 지시 위반"
      ],
      "physics": "네 명 모두 바닥을 단단히 디디고 자연스럽게 걷는 자세를 유지함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "세 인물이 복도 끝을 향해 걷는 점과 복장이 부분적으로 일치하나, 교도관이 카메라를 향해 서 있고 장소 레퍼런스를 완전히 무시하여 지시된 연출을 심각하게 위반했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 텍스트 표지판이 존재하나, 수사관 일행이 카메라를 향해 역방향으로 걸어오며 서의용 캐릭터가 누락된 채 얼굴이 복제되어 완전히 실패한 결과물입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "교도관은 카메라를 향해 정면을 보고 서 있으며, 나머지 세 명의 인물은 카메라를 등지고 복도 끝을 향해 걷고 있음.",
      "built_space": "일반적인 사무실이나 병원 복도처럼 보이며, 레퍼런스 이미지의 교도소 철문 및 구조와 전혀 다름.",
      "entities": "교도관(정면), 서의용(뒷모습, 비니와 가죽 재킷), 나상혁(뒷모습), 검찰 수사관(뒷모습)이 확인되나, '정숙' 표지판은 없음.",
      "hard_violations": [
       "교도관이 복도 끝이 아닌 카메라를 향해 서 있음 (동선 및 카메라 구도 위반)",
       "지정된 장소(교도소 복도) 레퍼런스 불일치"
      ],
      "physics": "인물들은 바닥을 정상적으로 딛고 서 있거나 걷고 있음."
     },
     {
      "label": "B",
      "direction": "교도관은 카메라를 등지고 있으며, 나머지 세 명의 인물은 교도관 및 카메라를 향해 정면으로 걸어오고 있음.",
      "built_space": "교도소 복도 느낌을 내고 있으며 좌측 벽면에 '정숙' 표지판이 있음.",
      "entities": "교도관(뒷모습), 나상혁, 검찰 수사관이 보이나, 서의용(비니/가죽 재킷)이 없고 정장 차림의 발명된 인물이 추가됨. 세 명의 얼굴이 복제된 듯 유사함.",
      "hard_violations": [
       "수사관 일행이 복도 끝이 아닌 카메라를 향해 걸어옴 (동선 위반)",
       "명시된 인물(서의용) 결여 및 다른 인물로 대체됨",
       "인물 얼굴 중복/복제"
      ],
      "physics": "인물들은 바닥을 정상적으로 딛고 걷고 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "single_gemini-pro",
     "models": [
      "gemini-pro"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "세 인물이 복도 끝을 향해 걷는 점과 복장이 부분적으로 일치하나, 교도관이 카메라를 향해 서 있고 장소 레퍼런스를 완전히 무시하여 지시된 연출을 심각하게 위반했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지정된 텍스트 표지판이 존재하나, 수사관 일행이 카메라를 향해 역방향으로 걸어오며 서의용 캐릭터가 누락된 채 얼굴이 복제되어 완전히 실패한 결과물입니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "교도관은 카메라를 향해 정면을 보고 서 있으며, 나머지 세 명의 인물은 카메라를 등지고 복도 끝을 향해 걷고 있음.",
      "built_space": "일반적인 사무실이나 병원 복도처럼 보이며, 레퍼런스 이미지의 교도소 철문 및 구조와 전혀 다름.",
      "entities": "교도관(정면), 서의용(뒷모습, 비니와 가죽 재킷), 나상혁(뒷모습), 검찰 수사관(뒷모습)이 확인되나, '정숙' 표지판은 없음.",
      "hard_violations": [
       "교도관이 복도 끝이 아닌 카메라를 향해 서 있음 (동선 및 카메라 구도 위반)",
       "지정된 장소(교도소 복도) 레퍼런스 불일치"
      ],
      "physics": "인물들은 바닥을 정상적으로 딛고 서 있거나 걷고 있음."
     },
     {
      "label": "A",
      "direction": "교도관은 카메라를 등지고 있으며, 나머지 세 명의 인물은 교도관 및 카메라를 향해 정면으로 걸어오고 있음.",
      "built_space": "교도소 복도 느낌을 내고 있으며 좌측 벽면에 '정숙' 표지판이 있음.",
      "entities": "교도관(뒷모습), 나상혁, 검찰 수사관이 보이나, 서의용(비니/가죽 재킷)이 없고 정장 차림의 발명된 인물이 추가됨. 세 명의 얼굴이 복제된 듯 유사함.",
      "hard_violations": [
       "수사관 일행이 복도 끝이 아닌 카메라를 향해 걸어옴 (동선 위반)",
       "명시된 인물(서의용) 결여 및 다른 인물로 대체됨",
       "인물 얼굴 중복/복제"
      ],
      "physics": "인물들은 바닥을 정상적으로 딛고 걷고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 7
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S55sh1__bgfirst_bg.png",
   "bg_asset_id": "249270fd-de27-46a5-b130-6522d6ee6469",
   "bg_record_key": "S55sh1::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S55sh1::cine": {
  "applied": true,
  "fingerprint": "fadfa00a50e638cc8765f0133d47d10cf552afcb2ac14abc2d76da7f9238a43e",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S55sh1_sel.png",
  "source_sha256": "480bead845dabdfb8f72d07d61c02af0fab5580a4e5cdc2b413bf71a2859d779",
  "file": "S55sh1_cine.png",
  "latency_ms": 11718
 },
 "S55sh8::signage": {
  "fp": "5556723f2780eaba",
  "inscriptions": [
   {
    "surface_native": "접견실 유리창의 안내 스티커",
    "text_native": "경고: 녹음 및 녹화중",
    "reason_ko": "교도소 접견실 내부의 엄격하고 감시받는 분위기를 사실적으로 전달하기 위해 유리창에 붙은 경고 문구가 필요합니다."
   }
  ]
 },
 "era_assess::05bb85a98855076c": {
  "subjects": [
   {
    "subject_native": "대한민국 교도소 수의 (2001년~2017년)",
    "search_terms_native": [
     "교도소 수의",
     "재소자 수의",
     "교도소 죄수복"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어나 다른 언어로 번역하거나 추가적인 외국어 단어를 덧붙이지 마십시오.",
    "reason_ko": "한국 교도소의 수의는 미국식 오렌지색 점프수트나 흑백 줄무늬와 다르며, 당시 쓰이던 청색 또는 갈색 계열의 독특한 색상과 명찰 형태를 재현해야 하기 때문입니다."
   },
   {
    "subject_native": "대한민국 교도소 접견실 (2001년~2017년)",
    "search_terms_native": [
     "교도소 접견실",
     "교도소 면회실",
     "교도소 일반접견"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어나 다른 언어로 번역하거나 추가적인 외국어 단어를 덧붙이지 마십시오.",
    "reason_ko": "한국 교도소의 면회실 구조(유리 칸막이, 대화용 마이크, 특유의 테이블과 내부 마감재)는 서구식 교도소나 현대식 면회실과 시각적 특징이 전혀 다릅니다."
   }
  ]
 },
 "era_fail::2cc51f61abf22711": {
  "stage": "research",
  "subject": "대한민국 교도소 수의 (2001년~2017년)",
  "attempts": 3
 },
 "S55sh8::bgfirst_bg": {
  "input_fingerprint": "77183732fc27d2dd",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용 쪽을 향해 상체를 바짝 들이민 채 목소리를 낮추듯 입술을 뗀 40대 체격이 큰 한국인 남성의 측면\n\nLOCATION (lock): Inside a small prison interview room, across a simple table from the inmate.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Beside the interview-room table at seated shoulder height, the lateral move begins in a profile-dominant medium view of the 40대 체격이 큰 재소자 leaning close toward 서의용. The Korean inmate occupies the near side of frame with his lips just parting to lower his voice, while 서의용 remains across the table, his attention fixed on the disclosure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 40대 체격이 큰 재소자 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 40대 체격이 큰 재소자; interview-room table shared by both men in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interview-room table (Between the seated inmate and 서의용); used as Lower-frame divider and tracking axis connecting speaker and listener; small interview room (Occupied by 서의용 and the inmate); used as Tight spatial enclosure visible around the two-person interview.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the interview room retains subdued neutral-to-cool color and documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 교도소 수의 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용 쪽을 향해 상체를 바짝 들이민 채 목소리를 낮추듯 입술을 뗀 40대 체격이 큰 한국인 남성의 측면\n\nLOCATION (lock): Inside a small prison interview room, across a simple table from the inmate.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Beside the interview-room table at seated shoulder height, the lateral move begins in a profile-dominant medium view of the 40대 체격이 큰 재소자 leaning close toward 서의용. The Korean inmate occupies the near side of frame with his lips just parting to lower his voice, while 서의용 remains across the table, his attention fixed on the disclosure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 40대 체격이 큰 재소자 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 40대 체격이 큰 재소자; interview-room table shared by both men in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interview-room table (Between the seated inmate and 서의용); used as Lower-frame divider and tracking axis connecting speaker and listener; small interview room (Occupied by 서의용 and the inmate); used as Tight spatial enclosure visible around the two-person interview.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the interview room retains subdued neutral-to-cool color and documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 교도소 수의 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S55sh8__bgfirst_bg.png",
  "asset_id": "0018b552-00d8-4f85-9e23-19dc728bf2b2",
  "input_asset_ids": [
   "9e56d10e-6f79-4666-a863-0fb9a0b412f7",
   "83fe623c-d896-4882-b478-030f63ffb76f"
  ],
  "era_research": {
   "subject": "대한민국 교도소 수의 (2001년~2017년)",
   "queries": [
    [
     "대한민국 교도소 수 2001년 2017년 통계",
     "대한민국 재소자 수 2001년 2017년 통계",
     "대한민국 교도소 죄수복 2001년 2017년 수의 변화"
    ],
    [
     "법무부 교정통계연보 2018 교정시설 수 연도별 2001",
     "교정통계연보 2018 수용인원 2001 2017",
     "대한민국 교정시설 현황 2001년 교도소 구치소 수",
     "대한민국 수용자 복장 규칙 2001년 2017년 색상"
    ],
    [
     "\"2001\" \"1일 평균 수용인원\" 교정시설",
     "\"2001년\" \"수용인원\" \"교정시설\"",
     "\"2001\" \"수형자\" \"미결수용자\" 교정",
     "\"교정기관\" \"2001\" \"2017\" 교도소"
    ],
    [
     "\"교정기관 수\" \"2001\" 법무부",
     "\"교정시설 수\" \"2001\" \"2017\" 법무부",
     "\"교도소\" \"구치소\" \"지소\" \"2017\" \"개\" 교정시설",
     "\"전국 교정기관\" \"2001\" \"개\""
    ],
    [
     "대한민국 교도소 수용자복 2001년 2006년",
     "대한민국 교도소 수용자복 2007년 2017년"
    ],
    [
     "2017년 수용자 의류 색상 수형자 미결수용자 법무부령",
     "2017 교도소 수용자복 기결수 미결수 색상",
     "형의 집행 및 수용자의 처우에 관한 법률 시행규칙 2017 수용자 의류 색상"
    ]
   ],
   "picked_url": "https://yaejongshop.com/web/product/big/202302/da7209f63e3d2641076a4b9fa9a3921f.jpg",
   "sha256": "0fbc28c48b489e9721dffcedbc166992dfa88f1181effbc3579defa1d498848a",
   "file": "eraref_2cc51f61abf22711.png"
  }
 },
 "S55sh8": {
  "input_fingerprint": "f359be3327d79bee",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽을 향해 상체를 바짝 들이민 채 목소리를 낮추듯 입술을 뗀 40대 체격이 큰 한국인 남성의 측면\n\nLOCATION (lock): Inside a small prison interview room, across a simple table from the inmate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the interview-room table at seated shoulder height, the lateral move begins in a profile-dominant medium view of the 40대 체격이 큰 재소자 leaning close toward 서의용. The Korean inmate occupies the near side of frame with his lips just parting to lower his voice, while 서의용 remains across the table, his attention fixed on the disclosure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 40대 체격이 큰 재소자 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 40대 체격이 큰 재소자; interview-room table shared by both men in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interview-room table (Between the seated inmate and 서의용); used as Lower-frame divider and tracking axis connecting speaker and listener; small interview room (Occupied by 서의용 and the inmate); used as Tight spatial enclosure visible around the two-person interview.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the interview room retains subdued neutral-to-cool color and documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 접견실 유리창의 안내 스티커: \"경고: 녹음 및 녹화중\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽을 향해 상체를 바짝 들이민 채 목소리를 낮추듯 입술을 뗀 40대 체격이 큰 한국인 남성의 측면\n\nLOCATION (lock): Inside a small prison interview room, across a simple table from the inmate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the interview-room table at seated shoulder height, the lateral move begins in a profile-dominant medium view of the 40대 체격이 큰 재소자 leaning close toward 서의용. The Korean inmate occupies the near side of frame with his lips just parting to lower his voice, while 서의용 remains across the table, his attention fixed on the disclosure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 40대 체격이 큰 재소자 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 40대 체격이 큰 재소자; interview-room table shared by both men in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interview-room table (Between the seated inmate and 서의용); used as Lower-frame divider and tracking axis connecting speaker and listener; small interview room (Occupied by 서의용 and the inmate); used as Tight spatial enclosure visible around the two-person interview.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the interview room retains subdued neutral-to-cool color and documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 접견실 유리창의 안내 스티커: \"경고: 녹음 및 녹화중\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽을 향해 상체를 바짝 들이민 채 목소리를 낮추듯 입술을 뗀 40대 체격이 큰 한국인 남성의 측면\n\nLOCATION (lock): Inside a small prison interview room, across a simple table from the inmate. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the interview-room table at seated shoulder height, the lateral move begins in a profile-dominant medium view of the 40대 체격이 큰 재소자 leaning close toward 서의용. The Korean inmate occupies the near side of frame with his lips just parting to lower his voice, while 서의용 remains across the table, his attention fixed on the disclosure.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 40대 체격이 큰 재소자 in the middle-left of the frame, foreground, looks toward 서의용 across the table; 서의용 in the middle-right of the frame, midground, looks toward 40대 체격이 큰 재소자; interview-room table shared by both men in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: interview-room table (Between the seated inmate and 서의용); used as Lower-frame divider and tracking axis connecting speaker and listener; small interview room (Occupied by 서의용 and the inmate); used as Tight spatial enclosure visible around the two-person interview.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the interview room retains subdued neutral-to-cool color and documentary restraint.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 광주교도소 교도관 (Korean 남성, 성인, 단정한 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 접견실 유리창의 안내 스티커: \"경고: 녹음 및 녹화중\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S55sh8__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S55sh8.png"
    },
    {
     "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1232981>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1165975>"
    },
    {
     "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1217288>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1423703>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L41B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1232981>"
    },
    {
     "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1165975>"
    },
    {
     "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1217288>"
    },
    {
     "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1423703>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "구도와 지정된 텍스트는 잘 구현했으나, 오른쪽 서의용 자리에 다른 얼굴이 들어가고 왼쪽 재소자에게 서의용의 얼굴이 잘못 적용되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물 배치가 뒤바뀐 오류에 더해, 샷 텍스트에 없는 교도관을 임의로 추가하는 치명적인 위반을 범했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "왼쪽 남성은 상체를 약간 기울여 오른쪽 남성을 향해 시선을 맞추고 있으며, 오른쪽 남성 역시 왼쪽을 응시함.",
      "built_space": "두 사람 사이에 단순한 테이블이 놓여 있고, 뒤쪽으로 창살과 안내 텍스트가 붙은 유리창이 있는 접견실 구조.",
      "entities": "지정된 2인만 등장하나, 왼쪽 재소자에 서의용(Ref 4), 오른쪽 서의용 자리에 나상혁(Ref 2)의 얼굴이 잘못 할당됨. 창문에 '경고: 녹음 및 녹화중' 텍스트는 정확함.",
      "hard_violations": [],
      "physics": "두 남성 모두 테이블 위에 팔을 올려 상체를 안정적으로 지지하며 마주 앉아 있음."
     },
     {
      "label": "A",
      "direction": "왼쪽과 오른쪽 남성이 서로 마주보며 시선을 교환하고, 배경의 교도관은 정면을 향하고 있음.",
      "built_space": "마주 앉은 테이블과 뒤쪽 유리창 등 접견실의 기본적인 구조가 배치되어 있음.",
      "entities": "왼쪽 재소자에 서의용(Ref 4), 오른쪽에 나상혁(Ref 2) 얼굴이 잘못 배치되었고, 지시문에 없는 교도관(Ref 1)이 등장함. 창문에 '경고: 녹음 및 녹화중' 텍스트 존재함.",
      "hard_violations": [
       "샷 텍스트에 명시되지 않은 추가 인물(배경의 교도관) 등장"
      ],
      "physics": "왼쪽 남성의 팔이 테이블에 닿아 지지 중이며, 뒤쪽 교도관은 바닥을 딛고 서 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "구도와 지정된 텍스트는 잘 구현했으나, 오른쪽 서의용 자리에 다른 얼굴이 들어가고 왼쪽 재소자에게 서의용의 얼굴이 잘못 적용되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물 배치가 뒤바뀐 오류에 더해, 샷 텍스트에 없는 교도관을 임의로 추가하는 치명적인 위반을 범했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "왼쪽 남성은 상체를 약간 기울여 오른쪽 남성을 향해 시선을 맞추고 있으며, 오른쪽 남성 역시 왼쪽을 응시함.",
      "built_space": "두 사람 사이에 단순한 테이블이 놓여 있고, 뒤쪽으로 창살과 안내 텍스트가 붙은 유리창이 있는 접견실 구조.",
      "entities": "지정된 2인만 등장하나, 왼쪽 재소자에 서의용(Ref 4), 오른쪽 서의용 자리에 나상혁(Ref 2)의 얼굴이 잘못 할당됨. 창문에 '경고: 녹음 및 녹화중' 텍스트는 정확함.",
      "hard_violations": [],
      "physics": "두 남성 모두 테이블 위에 팔을 올려 상체를 안정적으로 지지하며 마주 앉아 있음."
     },
     {
      "label": "A",
      "direction": "왼쪽과 오른쪽 남성이 서로 마주보며 시선을 교환하고, 배경의 교도관은 정면을 향하고 있음.",
      "built_space": "마주 앉은 테이블과 뒤쪽 유리창 등 접견실의 기본적인 구조가 배치되어 있음.",
      "entities": "왼쪽 재소자에 서의용(Ref 4), 오른쪽에 나상혁(Ref 2) 얼굴이 잘못 배치되었고, 지시문에 없는 교도관(Ref 1)이 등장함. 창문에 '경고: 녹음 및 녹화중' 텍스트 존재함.",
      "hard_violations": [
       "샷 텍스트에 명시되지 않은 추가 인물(배경의 교도관) 등장"
      ],
      "physics": "왼쪽 남성의 팔이 테이블에 닿아 지지 중이며, 뒤쪽 교도관은 바닥을 딛고 서 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "수감자와 서의용의 역할을 무시하고 엉뚱한 레퍼런스(나상혁)를 배치한 심각한 신원 오류가 있으나, 불필요한 추가 인물은 없음."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물 신원 배치가 완전히 틀렸으며, 샷 텍스트에 명시되지 않은 교도관을 추가로 생성한 하드 위반이 있음."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽 남성이 오른쪽 남성을 향해 몸을 들이밀고 시선을 맞추며, 오른쪽 남성은 이를 주시함.",
      "built_space": "접견실 내부, 중앙에 테이블 1개가 놓여 있고 배경에 창문이 있음.",
      "entities": "지문과 달리 왼쪽에 '서의용', 오른쪽에 '나상혁' 레퍼런스 얼굴이 잘못 배치됨. 유리창의 '경고: 녹음 및 녹화중' 텍스트는 정확하게 렌더링됨.",
      "hard_violations": [],
      "physics": "두 사람 모두 의자에 앉아 테이블에 팔을 올리고 체중을 자연스럽게 지지함."
     },
     {
      "label": "B",
      "direction": "왼쪽 남성이 오른쪽 남성에게 상체를 기울이고, 뒤편 교도관은 정면을 응시함.",
      "built_space": "접견실 내부, 중앙 테이블 1개와 배경에 창문 및 문이 있음.",
      "entities": "왼쪽 수감자 역할에 '서의용', 오른쪽 역할에 '나상혁' 레퍼런스가 오용되었으며 텍스트에 없는 교도관이 배경에 나타남. 유리창 텍스트는 정확함.",
      "hard_violations": [
       "샷 텍스트에 없는 추가 인물(배경의 교도관) 등장"
      ],
      "physics": "앉아 있는 두 인물은 테이블에 팔을 의지하고, 배경의 인물은 바닥에 발을 지지해 서 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "수감자와 서의용의 역할을 무시하고 엉뚱한 레퍼런스(나상혁)를 배치한 심각한 신원 오류가 있으나, 불필요한 추가 인물은 없음."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물 신원 배치가 완전히 틀렸으며, 샷 텍스트에 명시되지 않은 교도관을 추가로 생성한 하드 위반이 있음."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "왼쪽 남성이 오른쪽 남성을 향해 몸을 들이밀고 시선을 맞추며, 오른쪽 남성은 이를 주시함.",
      "built_space": "접견실 내부, 중앙에 테이블 1개가 놓여 있고 배경에 창문이 있음.",
      "entities": "지문과 달리 왼쪽에 '서의용', 오른쪽에 '나상혁' 레퍼런스 얼굴이 잘못 배치됨. 유리창의 '경고: 녹음 및 녹화중' 텍스트는 정확하게 렌더링됨.",
      "hard_violations": [],
      "physics": "두 사람 모두 의자에 앉아 테이블에 팔을 올리고 체중을 자연스럽게 지지함."
     },
     {
      "label": "A",
      "direction": "왼쪽 남성이 오른쪽 남성에게 상체를 기울이고, 뒤편 교도관은 정면을 응시함.",
      "built_space": "접견실 내부, 중앙 테이블 1개와 배경에 창문 및 문이 있음.",
      "entities": "왼쪽 수감자 역할에 '서의용', 오른쪽 역할에 '나상혁' 레퍼런스가 오용되었으며 텍스트에 없는 교도관이 배경에 나타남. 유리창 텍스트는 정확함.",
      "hard_violations": [
       "샷 텍스트에 없는 추가 인물(배경의 교도관) 등장"
      ],
      "physics": "앉아 있는 두 인물은 테이블에 팔을 의지하고, 배경의 인물은 바닥에 발을 지지해 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 8
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "왼쪽 남성은 상체를 약간 기울여 오른쪽 남성을 향해 시선을 맞추고 있으며, 오른쪽 남성 역시 왼쪽을 응시함.",
    "built_space": "두 사람 사이에 단순한 테이블이 놓여 있고, 뒤쪽으로 창살과 안내 텍스트가 붙은 유리창이 있는 접견실 구조.",
    "entities": "지정된 2인만 등장하나, 왼쪽 재소자에 서의용(Ref 4), 오른쪽 서의용 자리에 나상혁(Ref 2)의 얼굴이 잘못 할당됨. 창문에 '경고: 녹음 및 녹화중' 텍스트는 정확함.",
    "hard_violations": [],
    "physics": "두 남성 모두 테이블 위에 팔을 올려 상체를 안정적으로 지지하며 마주 앉아 있음."
   },
   {
    "label": "A",
    "direction": "왼쪽과 오른쪽 남성이 서로 마주보며 시선을 교환하고, 배경의 교도관은 정면을 향하고 있음.",
    "built_space": "마주 앉은 테이블과 뒤쪽 유리창 등 접견실의 기본적인 구조가 배치되어 있음.",
    "entities": "왼쪽 재소자에 서의용(Ref 4), 오른쪽에 나상혁(Ref 2) 얼굴이 잘못 배치되었고, 지시문에 없는 교도관(Ref 1)이 등장함. 창문에 '경고: 녹음 및 녹화중' 텍스트 존재함.",
    "hard_violations": [
     "샷 텍스트에 명시되지 않은 추가 인물(배경의 교도관) 등장"
    ],
    "physics": "왼쪽 남성의 팔이 테이블에 닿아 지지 중이며, 뒤쪽 교도관은 바닥을 딛고 서 있음."
   }
  ],
  "totals": {
   "A": 6,
   "B": 8
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "구도와 지정된 텍스트는 잘 구현했으나, 오른쪽 서의용 자리에 다른 얼굴이 들어가고 왼쪽 재소자에게 서의용의 얼굴이 잘못 적용되었습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "인물 배치가 뒤바뀐 오류에 더해, 샷 텍스트에 없는 교도관을 임의로 추가하는 치명적인 위반을 범했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L41B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 광주교도소 교도관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1232981>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1165975>"
   },
   {
    "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1217288>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1423703>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "서의용이 테이블 우측(중경)에 있어야 한다는 지시와 달리, 우측에 나상혁의 레퍼런스 얼굴이 렌더링되었고 좌측 재소자 자리에 서의용의 레퍼런스 얼굴이 렌더링되어 등장인물이 완전히 잘못 배치되었습니다.",
     "fix_en": "Redraw both characters' heads to correct their identities. Give the left man a new, generic 40-something Korean face. Change the right man's face and hair to exactly match the reference for Seo Eui-yong (wide face, short cropped hair). Preserve their current bodies, poses, clothing, lighting, the table, and the room.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "방이 위치 사진의 접견실(나무 탁자·의자)과 달리 철창 창문과 금속 탁자로 되어 있다.",
     "fix_en": "Change the table surface to wood and the background walls to simple painted concrete without metal bars to better match the location reference. Preserve the characters, their clothing, poses, and the existing signage.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "서의용이 테이블 우측(중경)에 있어야 한다는 지시와 달리, 우측에 나상혁의 레퍼런스 얼굴이 렌더링되었고 좌측 재소자 자리에 서의용의 레퍼런스 얼굴이 렌더링되어 등장인물이 완전히 잘못 배치되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "오른쪽 서의용이 레퍼런스의 40대 중반 넓은 얼굴·짧은 머리가 아니라 젊은 가르마머리 남성이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 재소자가 서의용 레퍼런스와 동일한 얼굴이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "방이 위치 사진의 접견실(나무 탁자·의자)과 달리 철창 창문과 금속 탁자로 되어 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "방이 위치 사진의 접견실(나무 탁자·의자)과 달리 철창 창문과 금속 탁자로 되어 있다.",
    "fix_en": "Change the table surface to wood and the background walls to simple painted concrete without metal bars to better match the location reference. Preserve the characters, their clothing, poses, and the existing signage.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 6,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw both characters' heads to correct their identities. Give the left man a new, generic 40-something Korean face. Change the right man's face and hair to exactly match the reference for Seo Eui-yong (wide face, short cropped hair). Preserve their current bodies, poses, clothing, lighting, the table, and the room.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "요구된 측면 구도와 유리창의 텍스트 지시를 완벽히 구현했으나, 화자와 청자의 정체성(서의용의 위치 및 상대방 신원)이 프롬프트와 다르게 배치된 점이 아쉬움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 앵글과 인물 간의 상호작용을 완전히 무시하고 분할 화면(콜라주) 형태로 인물을 나열하여 치명적인 규정 위반임."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽 남성은 오른쪽 남성을 바라보고, 오른쪽 남성은 왼쪽 남성을 마주 보고 있음.",
      "built_space": "작은 접견실 내부. 두 사람 사이에 1개의 테이블이 배치되어 있고 배경에 유리창이 있음.",
      "entities": "왼쪽 인물은 '서의용', 오른쪽 인물은 '나상혁' 레퍼런스와 신원 및 복장이 일치함. 배경 유리창에 '경고: 녹음 및 녹화중' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 앉아 테이블 위에 팔을 기대어 상체를 안정적으로 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "두 인물 모두 서로를 향하지 않고 카메라 렌즈(정면)를 응시함.",
      "built_space": "물리적 공간의 연속성이 없는 단순 배경이며, 하단에 테이블 상판만 위치함.",
      "entities": "왼쪽 인물은 '나상혁', 오른쪽 인물은 '서의용' 레퍼런스와 일치함. 요구된 유리창과 안내 스티커 텍스트가 없음.",
      "hard_violations": [
       "collage/panels (서로 다른 두 이미지를 이어 붙인 분할 화면)"
      ],
      "physics": "인물들이 테이블 뒤에 위치하나, 하단이 잘려 있어 신체 지지 기반이나 상호작용을 확인할 수 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "요구된 측면 구도와 유리창의 텍스트 지시를 완벽히 구현했으나, 화자와 청자의 정체성(서의용의 위치 및 상대방 신원)이 프롬프트와 다르게 배치된 점이 아쉬움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 앵글과 인물 간의 상호작용을 완전히 무시하고 분할 화면(콜라주) 형태로 인물을 나열하여 치명적인 규정 위반임."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "왼쪽 남성은 오른쪽 남성을 바라보고, 오른쪽 남성은 왼쪽 남성을 마주 보고 있음.",
      "built_space": "작은 접견실 내부. 두 사람 사이에 1개의 테이블이 배치되어 있고 배경에 유리창이 있음.",
      "entities": "왼쪽 인물은 '서의용', 오른쪽 인물은 '나상혁' 레퍼런스와 신원 및 복장이 일치함. 배경 유리창에 '경고: 녹음 및 녹화중' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 앉아 테이블 위에 팔을 기대어 상체를 안정적으로 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "두 인물 모두 서로를 향하지 않고 카메라 렌즈(정면)를 응시함.",
      "built_space": "물리적 공간의 연속성이 없는 단순 배경이며, 하단에 테이블 상판만 위치함.",
      "entities": "왼쪽 인물은 '나상혁', 오른쪽 인물은 '서의용' 레퍼런스와 일치함. 요구된 유리창과 안내 스티커 텍스트가 없음.",
      "hard_violations": [
       "collage/panels (서로 다른 두 이미지를 이어 붙인 분할 화면)"
      ],
      "physics": "인물들이 테이블 뒤에 위치하나, 하단이 잘려 있어 신체 지지 기반이나 상호작용을 확인할 수 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "장소, 측면 구도, 그리고 안내 스티커 텍스트(경고: 녹음 및 녹화중)는 매우 훌륭하게 구현했으나, 프롬프트에 없는 인물이 등장하고 서의용의 역할이 뒤바뀌는 치명적 오류가 있음."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트에 없는 인물을 추가하고, 지정된 측면 구도와 장소를 완전히 무시한 채 정면 증명사진처럼 연출하여 핵심 지시를 모두 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 남성이 마주보지 않고 정면 렌즈를 똑바로 응시함.",
      "built_space": "배경이 없는 단색 공간이며, 하단에 좁은 테이블 윗면만 보임.",
      "entities": "왼쪽은 나상혁, 오른쪽은 서의용이며 안내 스티커 등 요구된 사물이 전혀 없음.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(나상혁) 등장",
       "지정된 장소(접견실) 누락",
       "측면 및 마주보는 상호작용 구도 완전 무시"
      ],
      "physics": "두 사람 모두 지지 기반이나 상호작용이 명확하지 않은 경직된 정면 자세임."
     },
     {
      "label": "B",
      "direction": "왼쪽 남성이 오른쪽 남성을 향해 시선을 고정하고 상체를 기울임.",
      "built_space": "교도소 접견실 내부에 테이블이 있고, 배경 유리창에 안내 스티커가 부착됨.",
      "entities": "왼쪽은 서의용, 오른쪽은 나상혁의 얼굴임. 창문에 '경고: 녹음 및 녹화중' 텍스트가 정확히 표기됨.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(나상혁) 등장",
       "지정된 인물(서의용)의 위치 및 역할 오류 (오른쪽 청자가 아닌 왼쪽 화자로 배치됨)"
      ],
      "physics": "테이블 위에 팔을 자연스럽게 올리고 상체를 기울인 자세를 안정적으로 지지함."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "장소, 측면 구도, 그리고 안내 스티커 텍스트(경고: 녹음 및 녹화중)는 매우 훌륭하게 구현했으나, 프롬프트에 없는 인물이 등장하고 서의용의 역할이 뒤바뀌는 치명적 오류가 있음."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트에 없는 인물을 추가하고, 지정된 측면 구도와 장소를 완전히 무시한 채 정면 증명사진처럼 연출하여 핵심 지시를 모두 위반함."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "두 남성이 마주보지 않고 정면 렌즈를 똑바로 응시함.",
      "built_space": "배경이 없는 단색 공간이며, 하단에 좁은 테이블 윗면만 보임.",
      "entities": "왼쪽은 나상혁, 오른쪽은 서의용이며 안내 스티커 등 요구된 사물이 전혀 없음.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(나상혁) 등장",
       "지정된 장소(접견실) 누락",
       "측면 및 마주보는 상호작용 구도 완전 무시"
      ],
      "physics": "두 사람 모두 지지 기반이나 상호작용이 명확하지 않은 경직된 정면 자세임."
     },
     {
      "label": "A",
      "direction": "왼쪽 남성이 오른쪽 남성을 향해 시선을 고정하고 상체를 기울임.",
      "built_space": "교도소 접견실 내부에 테이블이 있고, 배경 유리창에 안내 스티커가 부착됨.",
      "entities": "왼쪽은 서의용, 오른쪽은 나상혁의 얼굴임. 창문에 '경고: 녹음 및 녹화중' 텍스트가 정확히 표기됨.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(나상혁) 등장",
       "지정된 인물(서의용)의 위치 및 역할 오류 (오른쪽 청자가 아닌 왼쪽 화자로 배치됨)"
      ],
      "physics": "테이블 위에 팔을 자연스럽게 올리고 상체를 기울인 자세를 안정적으로 지지함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 11,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S55sh8__bgfirst_bg.png",
   "bg_asset_id": "0018b552-00d8-4f85-9e23-19dc728bf2b2",
   "bg_record_key": "S55sh8::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S55sh8::cine": {
  "applied": true,
  "fingerprint": "9b1dea7d39f36af3c7239cd59d09ef1ea52de9a455793013819a27782501bca4",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S55sh8_sel.png",
  "source_sha256": "7c866c9de859b6c0ea6a67fd5ae128474eaa2136efe6e97b1233e31b30459ef9",
  "file": "S55sh8_cine.png",
  "latency_ms": 14123
 },
 "S56sh4::signage": {
  "fp": "929600c425fea9ad",
  "inscriptions": [
   {
    "surface_native": "주민등록증 상단",
    "text_native": "주민등록증",
    "reason_ko": "신경희가 응시하고 있는 대상이 대한민국의 공식 신분증임을 보여주기 위해 주민등록증 타이틀이 필요합니다."
   },
   {
    "surface_native": "주민등록증 성명란",
    "text_native": "서의용",
    "reason_ko": "당황한 주인공이 확인한 신분증의 주인이 서의용임을 시청자에게 전달하기 위해 이름 표기가 필요합니다."
   }
  ]
 },
 "era_assess::e1b565d62ad07aaa": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 지방 다방의 배달용 오토바이와 보온통",
    "search_terms_native": [
     "다방 배달 보온통",
     "시티백 다방 배달",
     "다방 커피 배달 가방"
    ],
    "language_lock_native": "한국어로 작성된 검색어만 사용하여 검색 결과를 확인하고, 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "2000년대 한국의 전형적인 다방 배달용 오토바이(시티100 등)와 보자기에 싸인 전용 금속 보온통의 형태는 일반적인 이미지 생성 모델이 현대식 또는 서구식 배달 스쿠터로 오인하기 쉽습니다."
   }
  ]
 },
 "era_fail::c32c2848d77f5e21": {
  "stage": "research",
  "subject": "2000년대 한국 지방 다방의 배달용 오토바이와 보온통",
  "attempts": 1
 },
 "groupbg::한적한 다방 입구": {
  "input_fingerprint": "94b13810a9caef19",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "한적한 다방 입구",
    "tags": [
     "S56sh4",
     "S56sh6",
     "S56sh8"
    ]
   },
   "context_sig": "490a7acfe11526ee",
   "era_research_sha": "a277f963b7df89a199637f692890c7bbdf3f805332ede79bdb17646555ac34cf"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n양지다방 앞: 인적 드문 지방 소도시 거리의 옛날식 다방 출입구 주변이다. (특징: 복고풍 글씨체로 쓰인 오래된 '양지다방' 간판; 입구 앞에 세워진 배달용 스쿠터; 스쿠터 짐칸 쪽에 놓인 보자기로 싼 커피 배달용 보온통)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 양지다방 앞 - 낮\n- 한적한 거리. 다방 간판이 보인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 지방 다방의 배달용 오토바이와 보온통: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n양지다방 앞: 인적 드문 지방 소도시 거리의 옛날식 다방 출입구 주변이다. (특징: 복고풍 글씨체로 쓰인 오래된 '양지다방' 간판; 입구 앞에 세워진 배달용 스쿠터; 스쿠터 짐칸 쪽에 놓인 보자기로 싼 커피 배달용 보온통)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 양지다방 앞 - 낮\n- 한적한 거리. 다방 간판이 보인다.\n\nTIME OF DAY (lock): day.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 지방 다방의 배달용 오토바이와 보온통: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_한적한_다방_입구_838434.png",
  "asset_id": "b1118328-7462-4819-a2c3-a95353bf31b9",
  "input_asset_ids": [
   "d742d253-97d5-4387-ad93-2f0e339f5cae"
  ],
  "origin_tag": "S56sh4",
  "place_text": "On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter.",
  "origin_inputs": {
   "place_text": "On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter.",
   "time_of_day_en": "day",
   "conti_asset_id": "d742d253-97d5-4387-ad93-2f0e339f5cae"
  },
  "era_research": {
   "subject": "2000년대 한국 지방 다방의 배달용 오토바이와 보온통",
   "terms": [
    "다방 배달 보온통",
    "시티백 다방 배달",
    "다방 커피 배달 가방"
   ],
   "queries": [
    [
     "다방 배달 보온통 시티백 다방 배달",
     "다방 커피 배달 가방"
    ],
    [
     "다방 배달 보온통",
     "시티백 다방 배달",
     "다방 커피 배달 가방"
    ]
   ],
   "candidates": 4,
   "picked_index": 2,
   "picked_url": "https://shop1.daumcdn.net/shophow/p/E5263829000.jpg?opt=R500x500.q90&type=thumb&ut=20260226172048",
   "picked_reason_ko": "2번은 당시 다방 배달에서 쓰던 형태에 가까운 검은 보온 가방과 컵 고정 구조를 제품 전체가 잘 보이도록 명확하게 담았다.",
   "sha256": "a277f963b7df89a199637f692890c7bbdf3f805332ede79bdb17646555ac34cf",
   "file": "eraref_c32c2848d77f5e21.png",
   "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_c32c2848d77f5e21.png"
  }
 },
 "era_assess::46302c4902ef746b": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 길거리 스쿠터와 카페 입구",
    "search_terms_native": [
     "2000년대 한국 골목길 오토바이",
     "한국 카페 외관 2010년대",
     "대림 시티에이스 오토바이",
     "한국 길거리 스쿠터"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어 등 타 언어 단어를 혼용하거나 번역을 추가하지 마십시오.",
    "reason_ko": "2000년대에서 2010년대 사이 한국 골목길의 독특한 포장 상태, 카페 간판 및 외관 스타일, 그리고 당시 흔히 주차되어 있던 국산 오토바이(시티에이스 등)나 스쿠터의 세부 디테일은 일반적인 AI 모델이 서구식 거리나 현대식 스쿠터로 잘못 묘사할 가능성이 높습니다."
   }
  ]
 },
 "era_ref::b003756e904dba8c": {
  "subject": "2000년대~2010년대 한국의 길거리 스쿠터와 카페 입구",
  "terms": [
   "2000년대 한국 골목길 오토바이",
   "한국 카페 외관 2010년대",
   "대림 시티에이스 오토바이",
   "한국 길거리 스쿠터"
  ],
  "queries": [
   [
    "2000년대 한국 골목길 대림 시티에이스 오토바이 길거리 스쿠터",
    "2010년대 한국 카페 외관 입구 골목길 오토바이"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDA2MDNfMjMw/MDAxNzE3NDAzNTg5NTc4.9ejcG9-_LbqG_M1OLk7gtfILm4kHYeBVx-13MCaqkM8g.ueL-SsS60_cI_EJfgOVomziFTkMhybO5IPk6e-MKMv4g.JPEG/KakaoTalk_20240529_220819920_02.jpg?type=w400",
  "picked_reason_ko": "한국 골목의 일상적인 카페 출입구와 길가에 세워진 스쿠터가 함께 선명하게 보여, 2000~2010년대 형태·재료·색상·비례를 읽기 가장 좋은 참고 사진이다.",
  "sha256": "691dda3df7834ad72c1d3010d14e96cc2b93bd2794999250fe175e5db406772c",
  "file": "eraref_b003756e904dba8c.png"
 },
 "S56sh4::bgfirst_bg": {
  "input_fingerprint": "05a4df5b51078920",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용의 신분증을 향해 시선을 고정한 채 당황한 기색으로 눈을 동그랗게 뜬 신경희의 얼굴 클로즈업\n\nLOCATION (lock): On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Close behind 서의용's raised hand and right shoulder, the dolly-in settles slightly above 신경희's eye line on an oblique face close-up. Her widened eyes remain fixed on the identification at the edge of her eyeline, while 서의용's shoulder and raised hand form a narrow foreground obstruction without covering her startled expression.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 isolated in close-up in the middle-center of the frame, midground, looks toward identification held at the frame edge.\n- KEY BACKGROUND ELEMENTS: 서의용's identification (Raised in 서의용's hand) — The identification-bearing front face is presented toward 신경희 and appears obliquely at the frame edge; used as Identity-bearing surface held at the near edge of 신경희's fixed eyeline; coffee shop sign (Visible near the coffee shop entrance) — The sign's lettered face is oriented toward the street and remains partially visible behind the figures; used as Background location marker kept soft beyond the confrontation; coffee shop entrance (Approached by 신경희 before she is stopped) — The entrance faces the street behind the group; used as Deep background passage indicating where 신경희 was heading.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination is treated with restrained neutral-to-cool color, keeping 신경희's startled eyes clear without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 길거리 스쿠터와 카페 입구: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용의 신분증을 향해 시선을 고정한 채 당황한 기색으로 눈을 동그랗게 뜬 신경희의 얼굴 클로즈업\n\nLOCATION (lock): On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Close behind 서의용's raised hand and right shoulder, the dolly-in settles slightly above 신경희's eye line on an oblique face close-up. Her widened eyes remain fixed on the identification at the edge of her eyeline, while 서의용's shoulder and raised hand form a narrow foreground obstruction without covering her startled expression.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 isolated in close-up in the middle-center of the frame, midground, looks toward identification held at the frame edge.\n- KEY BACKGROUND ELEMENTS: 서의용's identification (Raised in 서의용's hand) — The identification-bearing front face is presented toward 신경희 and appears obliquely at the frame edge; used as Identity-bearing surface held at the near edge of 신경희's fixed eyeline; coffee shop sign (Visible near the coffee shop entrance) — The sign's lettered face is oriented toward the street and remains partially visible behind the figures; used as Background location marker kept soft beyond the confrontation; coffee shop entrance (Approached by 신경희 before she is stopped) — The entrance faces the street behind the group; used as Deep background passage indicating where 신경희 was heading.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination is treated with restrained neutral-to-cool color, keeping 신경희's startled eyes clear without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 길거리 스쿠터와 카페 입구: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S56sh4__bgfirst_bg.png",
  "asset_id": "808f3ba8-d525-45e4-8e54-64d9f542b583",
  "input_asset_ids": [
   "d742d253-97d5-4387-ad93-2f0e339f5cae",
   "b1118328-7462-4819-a2c3-a95353bf31b9"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국의 길거리 스쿠터와 카페 입구",
   "queries": [
    [
     "2000년대 한국 골목길 대림 시티에이스 오토바이 길거리 스쿠터",
     "2010년대 한국 카페 외관 입구 골목길 오토바이"
    ]
   ],
   "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDA2MDNfMjMw/MDAxNzE3NDAzNTg5NTc4.9ejcG9-_LbqG_M1OLk7gtfILm4kHYeBVx-13MCaqkM8g.ueL-SsS60_cI_EJfgOVomziFTkMhybO5IPk6e-MKMv4g.JPEG/KakaoTalk_20240529_220819920_02.jpg?type=w400",
   "sha256": "691dda3df7834ad72c1d3010d14e96cc2b93bd2794999250fe175e5db406772c",
   "file": "eraref_b003756e904dba8c.png"
  }
 },
 "S56sh4": {
  "input_fingerprint": "b7c106e233fadcb5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 신분증을 향해 시선을 고정한 채 당황한 기색으로 눈을 동그랗게 뜬 신경희의 얼굴 클로즈업\n\nLOCATION (lock): On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close behind 서의용's raised hand and right shoulder, the dolly-in settles slightly above 신경희's eye line on an oblique face close-up. Her widened eyes remain fixed on the identification at the edge of her eyeline, while 서의용's shoulder and raised hand form a narrow foreground obstruction without covering her startled expression.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 isolated in close-up in the middle-center of the frame, midground, looks toward identification held at the frame edge.\n- KEY BACKGROUND ELEMENTS: 서의용's identification (Raised in 서의용's hand) — The identification-bearing front face is presented toward 신경희 and appears obliquely at the frame edge; used as Identity-bearing surface held at the near edge of 신경희's fixed eyeline; coffee shop sign (Visible near the coffee shop entrance) — The sign's lettered face is oriented toward the street and remains partially visible behind the figures; used as Background location marker kept soft beyond the confrontation; coffee shop entrance (Approached by 신경희 before she is stopped) — The entrance faces the street behind the group; used as Deep background passage indicating where 신경희 was heading.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination is treated with restrained neutral-to-cool color, keeping 신경희's startled eyes clear without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong keeps his police identification presented toward Gyeong-hui as he identifies himself.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 주민등록증 상단: \"주민등록증\"\n- 주민등록증 성명란: \"서의용\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 신분증을 향해 시선을 고정한 채 당황한 기색으로 눈을 동그랗게 뜬 신경희의 얼굴 클로즈업\n\nLOCATION (lock): On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close behind 서의용's raised hand and right shoulder, the dolly-in settles slightly above 신경희's eye line on an oblique face close-up. Her widened eyes remain fixed on the identification at the edge of her eyeline, while 서의용's shoulder and raised hand form a narrow foreground obstruction without covering her startled expression.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 isolated in close-up in the middle-center of the frame, midground, looks toward identification held at the frame edge.\n- KEY BACKGROUND ELEMENTS: 서의용's identification (Raised in 서의용's hand) — The identification-bearing front face is presented toward 신경희 and appears obliquely at the frame edge; used as Identity-bearing surface held at the near edge of 신경희's fixed eyeline; coffee shop sign (Visible near the coffee shop entrance) — The sign's lettered face is oriented toward the street and remains partially visible behind the figures; used as Background location marker kept soft beyond the confrontation; coffee shop entrance (Approached by 신경희 before she is stopped) — The entrance faces the street behind the group; used as Deep background passage indicating where 신경희 was heading.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination is treated with restrained neutral-to-cool color, keeping 신경희's startled eyes clear without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong keeps his police identification presented toward Gyeong-hui as he identifies himself.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 주민등록증 상단: \"주민등록증\"\n- 주민등록증 성명란: \"서의용\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용의 신분증을 향해 시선을 고정한 채 당황한 기색으로 눈을 동그랗게 뜬 신경희의 얼굴 클로즈업\n\nLOCATION (lock): On the quiet street directly outside the coffeehouse entrance, beside the stopped scooter. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close behind 서의용's raised hand and right shoulder, the dolly-in settles slightly above 신경희's eye line on an oblique face close-up. Her widened eyes remain fixed on the identification at the edge of her eyeline, while 서의용's shoulder and raised hand form a narrow foreground obstruction without covering her startled expression.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 isolated in close-up in the middle-center of the frame, midground, looks toward identification held at the frame edge.\n- KEY BACKGROUND ELEMENTS: 서의용's identification (Raised in 서의용's hand) — The identification-bearing front face is presented toward 신경희 and appears obliquely at the frame edge; used as Identity-bearing surface held at the near edge of 신경희's fixed eyeline; coffee shop sign (Visible near the coffee shop entrance) — The sign's lettered face is oriented toward the street and remains partially visible behind the figures; used as Background location marker kept soft beyond the confrontation; coffee shop entrance (Approached by 신경희 before she is stopped) — The entrance faces the street behind the group; used as Deep background passage indicating where 신경희 was heading.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination is treated with restrained neutral-to-cool color, keeping 신경희's startled eyes clear without sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong keeps his police identification presented toward Gyeong-hui as he identifies himself.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 주민등록증 상단: \"주민등록증\"\n- 주민등록증 성명란: \"서의용\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S56sh4__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S56sh4.png"
    },
    {
     "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:843594>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_한적한_다방_입구_838434.png"
    },
    {
     "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:843594>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "로케이션과 캐릭터 레퍼런스를 완벽하게 구현하고 요구된 텍스트를 정확히 렌더링했으나, 신분증의 앞면이 신경희가 아닌 카메라를 향하고 있는 물리적 방향 오류가 아쉬움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 요구한 오른쪽 어깨 대신 왼쪽 어깨가 전경에 배치되어 카메라 구도를 위반했으며, 남성의 신분증에 여성 본인의 사진이 들어가는 결정적인 오류가 있음."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 전경에 들려 있는 신분증을 향해 고정되어 있음.",
      "built_space": "배경에 '양지다방' 간판과 출입구, 스쿠터가 로케이션 레퍼런스와 동일한 위치와 비례로 배치되어 있음.",
      "entities": "신경희의 얼굴, 헤어스타일, 의상이 레퍼런스와 정확히 일치함. 신분증에는 요구된 '주민등록증'과 '서의용' 텍스트가 명확히 적혀 있음. 전경에 남성의 오른쪽 어깨와 오른손이 구현됨.",
      "hard_violations": [],
      "physics": "신분증은 전경 인물의 오른손에 의해 안정적으로 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "신경희의 시선이 전경의 신분증을 향하고 있음.",
      "built_space": "배경의 다방 입구와 스쿠터의 위치가 레퍼런스와 다르게 좌우가 반전되거나 왜곡되어 배치됨.",
      "entities": "신경희의 캐릭터 레퍼런스는 일치하나, 서의용의 신분증에 신경희 본인의 증명사진이 들어간 치명적 오류가 있음. 전경에는 오른쪽 어깨가 아닌 왼쪽 어깨가 보임.",
      "hard_violations": [
       "지정된 인물의 신분증에 다른 인물의 사진이 들어간 발명된 객체(invented objects/details)",
       "명시된 카메라 구도(서의용의 오른쪽 어깨 뒤)를 위반하고 왼쪽 어깨를 렌더링함"
      ],
      "physics": "신분증이 전경 인물의 손에 쥐어져 지탱되고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "로케이션과 캐릭터 레퍼런스를 완벽하게 구현하고 요구된 텍스트를 정확히 렌더링했으나, 신분증의 앞면이 신경희가 아닌 카메라를 향하고 있는 물리적 방향 오류가 아쉬움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 요구한 오른쪽 어깨 대신 왼쪽 어깨가 전경에 배치되어 카메라 구도를 위반했으며, 남성의 신분증에 여성 본인의 사진이 들어가는 결정적인 오류가 있음."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 전경에 들려 있는 신분증을 향해 고정되어 있음.",
      "built_space": "배경에 '양지다방' 간판과 출입구, 스쿠터가 로케이션 레퍼런스와 동일한 위치와 비례로 배치되어 있음.",
      "entities": "신경희의 얼굴, 헤어스타일, 의상이 레퍼런스와 정확히 일치함. 신분증에는 요구된 '주민등록증'과 '서의용' 텍스트가 명확히 적혀 있음. 전경에 남성의 오른쪽 어깨와 오른손이 구현됨.",
      "hard_violations": [],
      "physics": "신분증은 전경 인물의 오른손에 의해 안정적으로 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "신경희의 시선이 전경의 신분증을 향하고 있음.",
      "built_space": "배경의 다방 입구와 스쿠터의 위치가 레퍼런스와 다르게 좌우가 반전되거나 왜곡되어 배치됨.",
      "entities": "신경희의 캐릭터 레퍼런스는 일치하나, 서의용의 신분증에 신경희 본인의 증명사진이 들어간 치명적 오류가 있음. 전경에는 오른쪽 어깨가 아닌 왼쪽 어깨가 보임.",
      "hard_violations": [
       "지정된 인물의 신분증에 다른 인물의 사진이 들어간 발명된 객체(invented objects/details)",
       "명시된 카메라 구도(서의용의 오른쪽 어깨 뒤)를 위반하고 왼쪽 어깨를 렌더링함"
      ],
      "physics": "신분증이 전경 인물의 손에 쥐어져 지탱되고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 오버더숄더 구도, 인물의 시선 방향, 신분증의 방향과 텍스트, 배경의 다방 간판과 스쿠터 배치를 모두 훌륭하게 구현한 완성도 높은 결과물입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "신분증이 인물이 아닌 카메라를 향해 잘못 배치되었고, 다른 사람의 신분증에 본인의 사진이 들어가는 등 소품과 상황의 논리적 오류가 큽니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "여자의 시선이 신분증이 아닌 약간 아래를 향함. 신분증의 앞면이 여자가 아닌 카메라 렌즈 쪽을 향하고 있음.",
      "built_space": "배경에 다방 입구와 스쿠터가 보이나 간판의 글자가 일부 잘림.",
      "entities": "신경희의 외모와 의상은 레퍼런스와 일치. 신분증 텍스트('주민등록증', '서의용')는 맞으나 서의용의 신분증에 신경희의 얼굴 사진이 들어감.",
      "hard_violations": [
       "신분증이 상대방이 아닌 카메라를 향해 제시됨 (소품 방향 및 논리 오류).",
       "타인의 신분증에 현재 인물의 사진이 삽입됨."
      ],
      "physics": "전경의 손이 신분증을 쥐고 있음. 스쿠터가 바닥에 서 있음."
     },
     {
      "label": "B",
      "direction": "여자의 시선이 프레임 가장자리의 신분증을 정확히 향함. 신분증 앞면이 여자를 향해 비스듬히 제시됨.",
      "built_space": "레퍼런스와 일치하는 양지다방 간판, 입구, 스쿠터가 올바른 위치에 배치됨.",
      "entities": "신경희의 외모 및 의상 일치. 남자의 오른쪽 어깨 너머 구도(오버더숄더)가 정확함. 신분증 텍스트('주민등록증', '서의용')가 올바름.",
      "hard_violations": [],
      "physics": "남자의 오른손이 신분증을 잡고 있음. 배경의 스쿠터가 지면에 닿아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 오버더숄더 구도, 인물의 시선 방향, 신분증의 방향과 텍스트, 배경의 다방 간판과 스쿠터 배치를 모두 훌륭하게 구현한 완성도 높은 결과물입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "신분증이 인물이 아닌 카메라를 향해 잘못 배치되었고, 다른 사람의 신분증에 본인의 사진이 들어가는 등 소품과 상황의 논리적 오류가 큽니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "여자의 시선이 신분증이 아닌 약간 아래를 향함. 신분증의 앞면이 여자가 아닌 카메라 렌즈 쪽을 향하고 있음.",
      "built_space": "배경에 다방 입구와 스쿠터가 보이나 간판의 글자가 일부 잘림.",
      "entities": "신경희의 외모와 의상은 레퍼런스와 일치. 신분증 텍스트('주민등록증', '서의용')는 맞으나 서의용의 신분증에 신경희의 얼굴 사진이 들어감.",
      "hard_violations": [
       "신분증이 상대방이 아닌 카메라를 향해 제시됨 (소품 방향 및 논리 오류).",
       "타인의 신분증에 현재 인물의 사진이 삽입됨."
      ],
      "physics": "전경의 손이 신분증을 쥐고 있음. 스쿠터가 바닥에 서 있음."
     },
     {
      "label": "A",
      "direction": "여자의 시선이 프레임 가장자리의 신분증을 정확히 향함. 신분증 앞면이 여자를 향해 비스듬히 제시됨.",
      "built_space": "레퍼런스와 일치하는 양지다방 간판, 입구, 스쿠터가 올바른 위치에 배치됨.",
      "entities": "신경희의 외모 및 의상 일치. 남자의 오른쪽 어깨 너머 구도(오버더숄더)가 정확함. 신분증 텍스트('주민등록증', '서의용')가 올바름.",
      "hard_violations": [],
      "physics": "남자의 오른손이 신분증을 잡고 있음. 배경의 스쿠터가 지면에 닿아 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "신경희의 시선은 전경에 들려 있는 신분증을 향해 고정되어 있음.",
    "built_space": "배경에 '양지다방' 간판과 출입구, 스쿠터가 로케이션 레퍼런스와 동일한 위치와 비례로 배치되어 있음.",
    "entities": "신경희의 얼굴, 헤어스타일, 의상이 레퍼런스와 정확히 일치함. 신분증에는 요구된 '주민등록증'과 '서의용' 텍스트가 명확히 적혀 있음. 전경에 남성의 오른쪽 어깨와 오른손이 구현됨.",
    "hard_violations": [],
    "physics": "신분증은 전경 인물의 오른손에 의해 안정적으로 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "신경희의 시선이 전경의 신분증을 향하고 있음.",
    "built_space": "배경의 다방 입구와 스쿠터의 위치가 레퍼런스와 다르게 좌우가 반전되거나 왜곡되어 배치됨.",
    "entities": "신경희의 캐릭터 레퍼런스는 일치하나, 서의용의 신분증에 신경희 본인의 증명사진이 들어간 치명적 오류가 있음. 전경에는 오른쪽 어깨가 아닌 왼쪽 어깨가 보임.",
    "hard_violations": [
     "지정된 인물의 신분증에 다른 인물의 사진이 들어간 발명된 객체(invented objects/details)",
     "명시된 카메라 구도(서의용의 오른쪽 어깨 뒤)를 위반하고 왼쪽 어깨를 렌더링함"
    ],
    "physics": "신분증이 전경 인물의 손에 쥐어져 지탱되고 있음."
   }
  ],
  "totals": {
   "A": 15,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "로케이션과 캐릭터 레퍼런스를 완벽하게 구현하고 요구된 텍스트를 정확히 렌더링했으나, 신분증의 앞면이 신경희가 아닌 카메라를 향하고 있는 물리적 방향 오류가 아쉬움."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "프롬프트가 요구한 오른쪽 어깨 대신 왼쪽 어깨가 전경에 배치되어 카메라 구도를 위반했으며, 남성의 신분증에 여성 본인의 사진이 들어가는 결정적인 오류가 있음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_한적한_다방_입구_838434.png"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:843594>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "신분증을 쥐고 있는 오른쪽 전경 인물의 손이 해부학적으로 심하게 꼬여 형태와 구조가 붕괴되었습니다.",
     "fix_en": "Redraw the foreground right hand holding the ID card to correct its anatomical structure, ensuring a natural thumb joint and wrist connection. Strictly preserve the woman's face, her clothing, the ID card text and design, the man's dark jacket, the background buildings, the scooter, and the overall lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "제공된 배경을 정확히 유지하라는 지시와 달리, 주차된 스쿠터의 기종과 색상이 완전히 바뀌었고 커피숍 외벽 타일 등 건축 디테일이 변형되었습니다.",
     "fix_en": "Restore the exact background architecture, tiles, and the specific scooter model and color from the reference image.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "신분증 앞면이 프레임 가장자리 사선이 아니라 화면 안쪽에서 카메라에 거의 정면으로 읽힌다",
     "fix_en": "Reposition the ID card closer to the frame edge and orient it at a more oblique angle.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "전경의 쳐든 손·어깨 가림 대신 남성 뒤통수와 등이 오른쪽을 크게 차지한다",
     "fix_en": "Adjust the foreground framing to match the layout sketch, showing only a narrow shoulder and raised hand obstruction instead of a large back and head.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "신경희가 화면 한가운데가 아니라 왼쪽으로 치우쳐 있다",
     "fix_en": "Reframe the shot to center the woman in the middle of the frame.",
     "severity": "minor",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "신분증을 쥐고 있는 오른쪽 전경 인물의 손이 해부학적으로 심하게 꼬여 형태와 구조가 붕괴되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "제공된 배경을 정확히 유지하라는 지시와 달리, 주차된 스쿠터의 기종과 색상이 완전히 바뀌었고 커피숍 외벽 타일 등 건축 디테일이 변형되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "신분증 앞면이 프레임 가장자리 사선이 아니라 화면 안쪽에서 카메라에 거의 정면으로 읽힌다",
     "severity": "major"
    },
    {
     "issue_ko": "전경의 쳐든 손·어깨 가림 대신 남성 뒤통수와 등이 오른쪽을 크게 차지한다",
     "severity": "major"
    },
    {
     "issue_ko": "레이아웃 스케치에 있던 전경의 크게 쳐든 손이 없다",
     "severity": "major"
    },
    {
     "issue_ko": "신경희가 화면 한가운데가 아니라 왼쪽으로 치우쳐 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "제공된 배경을 정확히 유지하라는 지시와 달리, 주차된 스쿠터의 기종과 색상이 완전히 바뀌었고 커피숍 외벽 타일 등 건축 디테일이 변형되었습니다.",
    "fix_en": "Restore the exact background architecture, tiles, and the specific scooter model and color from the reference image.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "신분증 앞면이 프레임 가장자리 사선이 아니라 화면 안쪽에서 카메라에 거의 정면으로 읽힌다",
    "fix_en": "Reposition the ID card closer to the frame edge and orient it at a more oblique angle.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "전경의 쳐든 손·어깨 가림 대신 남성 뒤통수와 등이 오른쪽을 크게 차지한다",
    "fix_en": "Adjust the foreground framing to match the layout sketch, showing only a narrow shoulder and raised hand obstruction instead of a large back and head.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "신경희가 화면 한가운데가 아니라 왼쪽으로 치우쳐 있다",
    "fix_en": "Reframe the shot to center the woman in the middle of the frame.",
    "severity": "minor",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the foreground right hand holding the ID card to correct its anatomical structure, ensuring a natural thumb joint and wrist connection. Strictly preserve the woman's face, her clothing, the ID card text and design, the man's dark jacket, the background buildings, the scooter, and the overall lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요청된 클로즈업 샷 크기와 스케치를 정확히 따랐으며, 당황한 표정과 신분증의 텍스트 지시사항을 모두 충실히 이행했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍을 무시하고 전신을 보여주었으며, 신분증의 텍스트가 지시와 다르고 방향도 잘못되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선이 화면 우측 하단에서 서의용이 들고 있는 신분증을 정확히 향하고 있습니다.",
      "built_space": "양지다방 입구와 스쿠터가 배경에 올바르게 위치하며 피사계 심도로 인해 적절히 흐리게 표현되었습니다.",
      "entities": "신경희는 참조 이미지의 의상과 외모를 유지하며 당황해 눈을 동그랗게 뜬 표정을 잘 묘사했습니다. 신분증에는 지시된 '주민등록증'과 '서의용'이 정확히 적혀 있습니다.",
      "hard_violations": [],
      "physics": "서의용의 손이 신분증을 자연스럽게 쥐고 있습니다."
     },
     {
      "label": "B",
      "direction": "신경희의 시선이 카메라 쪽을 향하고 있으며, 신분증의 앞면이 신경희가 아닌 카메라(관찰자)를 향하고 있습니다.",
      "built_space": "다방 간판, 출입문, 스쿠터 등 배경 요소가 선명하게 모두 보입니다.",
      "entities": "신경희의 외모는 참조 이미지와 일치하나 표정이 무덤덤합니다. 신분증에는 지시되지 않은 신경희의 사진과 잘못된 텍스트가 적혀 있습니다.",
      "hard_violations": [
       "프레이밍 위반 (클로즈업이 아닌 전신 샷)",
       "텍스트 지시 위반 (지정된 텍스트 대신 임의의 텍스트와 이미지 생성)",
       "소품 방향 위반 (신분증이 신경희가 아닌 카메라를 향함)"
      ],
      "physics": "신경희는 바닥에 서 있고, 서의용의 손이 신분증을 들고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요청된 클로즈업 샷 크기와 스케치를 정확히 따랐으며, 당황한 표정과 신분증의 텍스트 지시사항을 모두 충실히 이행했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍을 무시하고 전신을 보여주었으며, 신분증의 텍스트가 지시와 다르고 방향도 잘못되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선이 화면 우측 하단에서 서의용이 들고 있는 신분증을 정확히 향하고 있습니다.",
      "built_space": "양지다방 입구와 스쿠터가 배경에 올바르게 위치하며 피사계 심도로 인해 적절히 흐리게 표현되었습니다.",
      "entities": "신경희는 참조 이미지의 의상과 외모를 유지하며 당황해 눈을 동그랗게 뜬 표정을 잘 묘사했습니다. 신분증에는 지시된 '주민등록증'과 '서의용'이 정확히 적혀 있습니다.",
      "hard_violations": [],
      "physics": "서의용의 손이 신분증을 자연스럽게 쥐고 있습니다."
     },
     {
      "label": "B",
      "direction": "신경희의 시선이 카메라 쪽을 향하고 있으며, 신분증의 앞면이 신경희가 아닌 카메라(관찰자)를 향하고 있습니다.",
      "built_space": "다방 간판, 출입문, 스쿠터 등 배경 요소가 선명하게 모두 보입니다.",
      "entities": "신경희의 외모는 참조 이미지와 일치하나 표정이 무덤덤합니다. 신분증에는 지시되지 않은 신경희의 사진과 잘못된 텍스트가 적혀 있습니다.",
      "hard_violations": [
       "프레이밍 위반 (클로즈업이 아닌 전신 샷)",
       "텍스트 지시 위반 (지정된 텍스트 대신 임의의 텍스트와 이미지 생성)",
       "소품 방향 위반 (신분증이 신경희가 아닌 카메라를 향함)"
      ],
      "physics": "신경희는 바닥에 서 있고, 서의용의 손이 신분증을 들고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 얼굴 클로즈업 프레이밍을 완벽히 따랐으며, 당황하여 눈을 크게 뜬 표정과 신분증을 향한 시선, 지정된 텍스트('주민등록증', '서의용')를 모두 정확하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 전신 풀샷으로 렌더링했으며, 시선이 카메라를 향하고 신분증에 잘못된 텍스트('신경희' 등)를 임의로 추가하여 연출 의도를 크게 벗어났습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "신경희의 시선이 화면 우측 남자가 들고 있는 신분증을 정확히 향하고 있음.",
      "built_space": "원본 배경의 다방 입구와 스쿠터가 인물 뒤로 적절한 원근감과 심도로 배치됨.",
      "entities": "신경희는 얼굴 클로즈업 상태로 당황한 표정을 짓고 있으며, 신분증 표면에 '주민등록증'과 '서의용' 텍스트가 지시대로 적혀 있음.",
      "hard_violations": [],
      "physics": "신경희는 안정적으로 서 있고, 남자의 손이 신분증을 정상적으로 쥐고 있음."
     },
     {
      "label": "A",
      "direction": "신경희의 시선이 목표물(신분증)이 아닌 정면 렌즈를 향하고 있음.",
      "built_space": "다방 앞 배경이 잘 나타나 있으나, 카메라가 너무 뒤로 물러나 있어 의도된 스케일과 다름.",
      "entities": "신경희가 전신으로 등장하며 표정이 덤덤함. 신분증에는 지시와 다른 사진과 이름('신경희'), 임의의 숫자 텍스트가 포함됨.",
      "hard_violations": [
       "화면 내 지시되지 않은 임의의 텍스트(신경희, 주민등록번호 숫자 등) 생성"
      ],
      "physics": "두 사람 모두 지면에 정상적으로 서 있고, 손으로 물건을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 얼굴 클로즈업 프레이밍을 완벽히 따랐으며, 당황하여 눈을 크게 뜬 표정과 신분증을 향한 시선, 지정된 텍스트('주민등록증', '서의용')를 모두 정확하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 전신 풀샷으로 렌더링했으며, 시선이 카메라를 향하고 신분증에 잘못된 텍스트('신경희' 등)를 임의로 추가하여 연출 의도를 크게 벗어났습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선이 화면 우측 남자가 들고 있는 신분증을 정확히 향하고 있음.",
      "built_space": "원본 배경의 다방 입구와 스쿠터가 인물 뒤로 적절한 원근감과 심도로 배치됨.",
      "entities": "신경희는 얼굴 클로즈업 상태로 당황한 표정을 짓고 있으며, 신분증 표면에 '주민등록증'과 '서의용' 텍스트가 지시대로 적혀 있음.",
      "hard_violations": [],
      "physics": "신경희는 안정적으로 서 있고, 남자의 손이 신분증을 정상적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "신경희의 시선이 목표물(신분증)이 아닌 정면 렌즈를 향하고 있음.",
      "built_space": "다방 앞 배경이 잘 나타나 있으나, 카메라가 너무 뒤로 물러나 있어 의도된 스케일과 다름.",
      "entities": "신경희가 전신으로 등장하며 표정이 덤덤함. 신분증에는 지시와 다른 사진과 이름('신경희'), 임의의 숫자 텍스트가 포함됨.",
      "hard_violations": [
       "화면 내 지시되지 않은 임의의 텍스트(신경희, 주민등록번호 숫자 등) 생성"
      ],
      "physics": "두 사람 모두 지면에 정상적으로 서 있고, 손으로 물건을 지탱하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S56sh4__bgfirst_bg.png",
   "bg_asset_id": "808f3ba8-d525-45e4-8e54-64d9f542b583",
   "bg_record_key": "S56sh4::bgfirst_bg",
   "chain_winner": true,
   "authority": "groupbg",
   "group_key": "한적한 다방 입구",
   "groupbg_asset_id": "b1118328-7462-4819-a2c3-a95353bf31b9"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S56sh4::cine": {
  "applied": true,
  "fingerprint": "53d06d75c3bf1b143aabd9917d828e48f2407aa0140c94d3148f6dcd15a83fdc",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S56sh4_sel.png",
  "source_sha256": "ca45827b22d8d89e34fca875718be5a7210fa53a8de27994868963657fae9b17",
  "file": "S56sh4_cine.png",
  "latency_ms": 12443
 },
 "S56sh6::signage": {
  "fp": "9dfb5bed9ccf750b",
  "inscriptions": [
   {
    "surface_native": "유리문 스티커",
    "text_native": "미시오",
    "reason_ko": "커피숍 입구 유리문에 흔히 붙어 있는 문구로, 인물이 상대방을 밀치는 동작과 어우러져 공간의 사실감을 높입니다."
   }
  ]
 },
 "S56sh6": {
  "input_fingerprint": "1257db7fe7be8e6f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나상혁의 가슴팍을 두 손으로 세게 밀치는 순간의 신경희의 상체\n\nLOCATION (lock): At the coffeehouse's exterior entrance on the quiet street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height on the open side of the group, the tracked three-quarter angle tightens on 신경희's upper body at the instant both palms drive into 나상혁's chest. She occupies the near left-center with her weight committed forward, while 나상혁's torso angles away on the right as he absorbs the push and keeps his attention on her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 신경희 driving both hands forward in the middle-left of the frame, foreground, moves toward 나상혁's chest; 나상혁 receiving the push in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: coffee shop entrance (Beside the group) — The entrance faces the street behind the confrontation; used as Background destination from which 신경희 is about to break away; coffee shop sign (Visible above or near the entrance) — The sign's lettered face is oriented toward the street and remains visible in the background; used as Location marker retained outside the main action area; scooter (Stopped near the coffee shop entrance) — Its side is visible from the confrontation angle; used as Secondary parked object establishing 신경희's arrival without competing with the push.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with subdued neutral-to-cool treatment preserves the force of the contact while remaining grounded and non-sensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유리문 스티커: \"미시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나상혁의 가슴팍을 두 손으로 세게 밀치는 순간의 신경희의 상체\n\nLOCATION (lock): At the coffeehouse's exterior entrance on the quiet street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height on the open side of the group, the tracked three-quarter angle tightens on 신경희's upper body at the instant both palms drive into 나상혁's chest. She occupies the near left-center with her weight committed forward, while 나상혁's torso angles away on the right as he absorbs the push and keeps his attention on her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 신경희 driving both hands forward in the middle-left of the frame, foreground, moves toward 나상혁's chest; 나상혁 receiving the push in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: coffee shop entrance (Beside the group) — The entrance faces the street behind the confrontation; used as Background destination from which 신경희 is about to break away; coffee shop sign (Visible above or near the entrance) — The sign's lettered face is oriented toward the street and remains visible in the background; used as Location marker retained outside the main action area; scooter (Stopped near the coffee shop entrance) — Its side is visible from the confrontation angle; used as Secondary parked object establishing 신경희's arrival without competing with the push.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with subdued neutral-to-cool treatment preserves the force of the contact while remaining grounded and non-sensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유리문 스티커: \"미시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 나상혁의 가슴팍을 두 손으로 세게 밀치는 순간의 신경희의 상체\n\nLOCATION (lock): At the coffeehouse's exterior entrance on the quiet street. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height on the open side of the group, the tracked three-quarter angle tightens on 신경희's upper body at the instant both palms drive into 나상혁's chest. She occupies the near left-center with her weight committed forward, while 나상혁's torso angles away on the right as he absorbs the push and keeps his attention on her.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 신경희 driving both hands forward in the middle-left of the frame, foreground, moves toward 나상혁's chest; 나상혁 receiving the push in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: coffee shop entrance (Beside the group) — The entrance faces the street behind the confrontation; used as Background destination from which 신경희 is about to break away; coffee shop sign (Visible above or near the entrance) — The sign's lettered face is oriented toward the street and remains visible in the background; used as Location marker retained outside the main action area; scooter (Stopped near the coffee shop entrance) — Its side is visible from the confrontation angle; used as Secondary parked object establishing 신경희's arrival without competing with the push.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with subdued neutral-to-cool treatment preserves the force of the contact while remaining grounded and non-sensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유리문 스티커: \"미시오\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "신경희의 시선은 나상혁을 향하고, 두 손은 나상혁의 가슴(재킷)을 향해 뻗어 옷을 쥐고 있음. 나상혁의 시선은 신경희를 향함.",
    "built_space": "배경에 '양지다방' 간판, 유리문에 '미시오' 스티커, 측면이 보이는 스쿠터 등 이전 샷의 환경 요소가 올바른 위치와 방향으로 배치됨.",
    "entities": "신경희(얼굴, 패턴 상의, 스카프)와 나상혁(얼굴, 네이비 재킷, 셔츠) 모두 레퍼런스와 정확히 일치하며, 지시문 외의 인물은 없음.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 서서 체중을 지탱하고 있으며, 체중을 실어 뻗은 신경희의 팔은 나상혁의 상체에 의해 지지됨."
   },
   {
    "label": "B",
    "direction": "신경희의 시선과 두 손이 나상혁의 가슴을 향하며 주먹 쥔 형태로 밀치듯 닿아 있음. 나상혁의 시선은 신경희를 향함.",
    "built_space": "배경의 간판과 유리문 스티커가 정상적으로 배치되었으나, 스쿠터 앞바퀴 및 서스펜션 구조가 불분명하게 렌더링됨.",
    "entities": "신경희와 나상혁의 외모 및 복장이 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "두 인물은 바닥에 서서 지지받고 있으며, 신경희의 손은 나상혁의 가슴팍에 닿아 지탱됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "두 손바닥이 아닌 옷을 쥐는 형태로 표현되었으나, 나상혁이 밀치는 힘을 흡수하며 상체가 뒤로 기울어지는 디테일과 배경 요소의 연속성이 우수함."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "밀치는 동작에 대한 나상혁의 반응(상체가 뒤로 기울어짐)이 덜 묘사되어 서 있는 자세가 다소 꼿꼿하며, 배경 스쿠터의 형태가 뭉개짐."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S56sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:843594>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 신경희가 두 손바닥으로 나상혁의 가슴을 밀치는 동작을 요구했으나, 사진에서는 주먹을 쥐고 나상혁의 재킷 옷깃을 움켜쥐고 있음 (화면 중앙).",
     "fix_en": "Redraw the woman's hands and wrists as two separate hands with open palms pressed flat against the man's chest to push him, replacing the clenched fists grabbing his jacket. Preserve the woman's face, posture, and patterned blouse, the man's face, posture, and jacket, the background coffee shop, scooter, and lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "나상혁을 향해 뻗은 신경희의 두 팔이 교차하는 지점에서 팔뚝과 손목 부위가 해부학적으로 분리되지 않고 한 덩어리로 융합되어 있음 (화면 중앙).",
     "fix_en": "Redraw the woman's forearms and hands to separate them into two distinct, anatomically correct limbs with open palms pushing the man's chest, eliminating the fused mass of flesh. Preserve the woman's face, posture, and patterned blouse, the man's face, posture, and jacket, the background coffee shop, scooter, and lighting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "오른쪽 나상혁이 밀침을 받으며 몸통이 물러나지 않고 그녀를 향해 곧게 서 있다",
     "fix_en": "Redraw the man's torso to angle slightly backward to the right to show him absorbing a push. Preserve his face, the woman's pose and clothing, the background coffee shop, scooter, and lighting.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 신경희가 두 손바닥으로 나상혁의 가슴을 밀치는 동작을 요구했으나, 사진에서는 주먹을 쥐고 나상혁의 재킷 옷깃을 움켜쥐고 있음 (화면 중앙).",
     "severity": "critical"
    },
    {
     "issue_ko": "나상혁을 향해 뻗은 신경희의 두 팔이 교차하는 지점에서 팔뚝과 손목 부위가 해부학적으로 분리되지 않고 한 덩어리로 융합되어 있음 (화면 중앙).",
     "severity": "critical"
    },
    {
     "issue_ko": "신경희가 손바닥으로 나상혁 가슴을 밀치지 않고 화면 중앙에서 재킷 앞섶을 두 손으로 움켜쥐고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 나상혁이 밀침을 받으며 몸통이 물러나지 않고 그녀를 향해 곧게 서 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "오른쪽 나상혁이 밀침을 받으며 몸통이 물러나지 않고 그녀를 향해 곧게 서 있다",
    "fix_en": "Redraw the man's torso to angle slightly backward to the right to show him absorbing a push. Preserve his face, the woman's pose and clothing, the background coffee shop, scooter, and lighting.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the woman's hands and wrists as two separate hands with open palms pressed flat against the man's chest to push him, replacing the clenched fists grabbing his jacket. Preserve the woman's face, posture, and patterned blouse, the man's face, posture, and jacket, the background coffee shop, scooter, and lighting.\n- Redraw the woman's forearms and hands to separate them into two distinct, anatomically correct limbs with open palms pushing the man's chest, eliminating the fused mass of flesh. Preserve the woman's face, posture, and patterned blouse, the man's face, posture, and jacket, the background coffee shop, scooter, and lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "후보 A는 신경희가 두 손으로 나상혁의 가슴을 밀치는 핵심 액션과 요구된 미디엄 샷 프레이밍을 역동적으로 잘 구현했으며, 지정된 '미시오' 텍스트까지 완벽하게 반영하여 프롬프트에 매우 충실합니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "후보 B는 두 손으로 상대방을 밀치는 프롬프트의 지시를 완전히 무시한 채 이전 샷의 구도와 불필요한 소품(신분증)을 그대로 답습하여 샷의 목적을 달성하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 나상혁을 향하며, 두 손은 나상혁의 앞가슴(재킷)으로 뻗어 물리적으로 강하게 접촉하고 있습니다. 나상혁의 시선은 신경희를 향하고 있습니다.",
      "built_space": "커피숍 외부 입구 앞 거리의 배경이 올바르게 구성되었으며, 이전 샷과 동일한 양지다방 간판, 주차된 스쿠터가 배경에 위치합니다. 유리문에 '미시오' 스티커가 정확히 부착되어 있습니다.",
      "entities": "신경희와 나상혁 모두 제공된 레퍼런스 이미지의 인상, 헤어스타일, 의상과 일치합니다. 프롬프트가 지시한 '미시오' 텍스트가 명확히 보입니다.",
      "hard_violations": [],
      "physics": "신경희는 체중을 앞으로 실어 나상혁의 가슴을 손으로 밀치거나 움켜쥐고 있으며, 두 사람 모두 바닥에 안정적으로 발을 딛고 동작의 상호작용과 힘의 균형이 자연스럽게 연출되었습니다."
     },
     {
      "label": "B",
      "direction": "신경희는 정면(혹은 나상혁)을 놀란 표정으로 응시하며 오른손을 방어적으로 들고 있습니다. 나상혁은 신경희를 향해 신분증을 제시하고 있습니다.",
      "built_space": "커피숍 입구와 간판, 스쿠터 등 지정된 배경 요소들이 위치해 있습니다. 하지만 문에 지시된 '미시오' 텍스트는 보이지 않습니다.",
      "entities": "신경희와 나상혁의 외양은 레퍼런스와 일치하나, 이 샷의 프롬프트에서 요구하지 않은 신분증을 나상혁이 들고 있습니다.",
      "hard_violations": [
       "프롬프트에 지시되지 않은 발명된 사물(신분증)이 샷의 주요 액션을 방해하며 등장함"
      ],
      "physics": "두 사람 모두 바닥을 딛고 서 있으나, 프롬프트에서 요구한 '가슴팍을 두 손으로 세게 밀치는' 물리적 행동이나 체중 이동은 전혀 발생하지 않고 단순히 정지된 자세를 취하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "후보 A는 신경희가 두 손으로 나상혁의 가슴을 밀치는 핵심 액션과 요구된 미디엄 샷 프레이밍을 역동적으로 잘 구현했으며, 지정된 '미시오' 텍스트까지 완벽하게 반영하여 프롬프트에 매우 충실합니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "후보 B는 두 손으로 상대방을 밀치는 프롬프트의 지시를 완전히 무시한 채 이전 샷의 구도와 불필요한 소품(신분증)을 그대로 답습하여 샷의 목적을 달성하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 나상혁을 향하며, 두 손은 나상혁의 앞가슴(재킷)으로 뻗어 물리적으로 강하게 접촉하고 있습니다. 나상혁의 시선은 신경희를 향하고 있습니다.",
      "built_space": "커피숍 외부 입구 앞 거리의 배경이 올바르게 구성되었으며, 이전 샷과 동일한 양지다방 간판, 주차된 스쿠터가 배경에 위치합니다. 유리문에 '미시오' 스티커가 정확히 부착되어 있습니다.",
      "entities": "신경희와 나상혁 모두 제공된 레퍼런스 이미지의 인상, 헤어스타일, 의상과 일치합니다. 프롬프트가 지시한 '미시오' 텍스트가 명확히 보입니다.",
      "hard_violations": [],
      "physics": "신경희는 체중을 앞으로 실어 나상혁의 가슴을 손으로 밀치거나 움켜쥐고 있으며, 두 사람 모두 바닥에 안정적으로 발을 딛고 동작의 상호작용과 힘의 균형이 자연스럽게 연출되었습니다."
     },
     {
      "label": "B",
      "direction": "신경희는 정면(혹은 나상혁)을 놀란 표정으로 응시하며 오른손을 방어적으로 들고 있습니다. 나상혁은 신경희를 향해 신분증을 제시하고 있습니다.",
      "built_space": "커피숍 입구와 간판, 스쿠터 등 지정된 배경 요소들이 위치해 있습니다. 하지만 문에 지시된 '미시오' 텍스트는 보이지 않습니다.",
      "entities": "신경희와 나상혁의 외양은 레퍼런스와 일치하나, 이 샷의 프롬프트에서 요구하지 않은 신분증을 나상혁이 들고 있습니다.",
      "hard_violations": [
       "프롬프트에 지시되지 않은 발명된 사물(신분증)이 샷의 주요 액션을 방해하며 등장함"
      ],
      "physics": "두 사람 모두 바닥을 딛고 서 있으나, 프롬프트에서 요구한 '가슴팍을 두 손으로 세게 밀치는' 물리적 행동이나 체중 이동은 전혀 발생하지 않고 단순히 정지된 자세를 취하고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "프롬프트가 요구한 두 손으로 가슴을 밀치는 동작과 지정된 카메라 앵글을 완벽하게 구현했으며, 배경의 '미시오' 스티커와 캐릭터들의 외양까지 정확하게 일치하는 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "이전 샷의 구도와 소품(신분증)을 그대로 답습하여 가슴을 밀치는 핵심 동작과 카메라 변경 지시를 완전히 무시했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "신경희의 시선은 나상혁을 향하고 있으며, 두 손은 나상혁의 가슴팍(재킷)을 향해 뻗어 밀치고 있습니다. 나상혁의 시선 역시 신경희를 향합니다.",
      "built_space": "배경에 커피숍 입구가 있으며, 유리문에는 '미시오' 스티커가 붙어 있고 그 위로 간판이 보입니다. 옆에는 스쿠터가 주차되어 있습니다.",
      "entities": "신경희와 나상혁 모두 제공된 캐릭터 레퍼런스의 얼굴, 헤어스타일, 의상과 정확히 일치합니다. 커피숍, 간판, 스쿠터 등 배경 요소가 프롬프트대로 배치되어 있습니다.",
      "hard_violations": [],
      "physics": "신경희의 두 손이 나상혁의 재킷을 물리적으로 움켜쥐고 밀고 있으며, 두 사람 모두 지면에 안정적으로 서 있는 자세를 취하고 있습니다."
     },
     {
      "label": "A",
      "direction": "신경희의 시선은 나상혁을 향하지만, 손은 가슴을 밀치지 않고 방어하듯 허공에 떠 있습니다. 나상혁은 카메라를 등진 채 신분증을 내밀고 있습니다.",
      "built_space": "배경에 커피숍 입구와 스쿠터, 간판이 보이나 유리문의 '미시오' 스티커가 누락되었습니다.",
      "entities": "신경희의 외형은 레퍼런스와 일치하나 나상혁은 뒷모습만 보이며, 프롬프트에 없는 이전 샷의 신분증이 그대로 남아 있습니다.",
      "hard_violations": [
       "신경희가 나상혁의 가슴팍을 밀치는 동작을 전혀 구현하지 않음",
       "지정된 카메라 구도(측면 쓰리쿼터 앵글)를 무시하고 이전 샷의 구도를 그대로 복사함"
      ],
      "physics": "나상혁의 손이 신분증을 들고 있으며, 신경희의 두 손은 허공에 올려진 채 대상을 밀치고 있지 않습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트가 요구한 두 손으로 가슴을 밀치는 동작과 지정된 카메라 앵글을 완벽하게 구현했으며, 배경의 '미시오' 스티커와 캐릭터들의 외양까지 정확하게 일치하는 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "이전 샷의 구도와 소품(신분증)을 그대로 답습하여 가슴을 밀치는 핵심 동작과 카메라 변경 지시를 완전히 무시했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 나상혁을 향하고 있으며, 두 손은 나상혁의 가슴팍(재킷)을 향해 뻗어 밀치고 있습니다. 나상혁의 시선 역시 신경희를 향합니다.",
      "built_space": "배경에 커피숍 입구가 있으며, 유리문에는 '미시오' 스티커가 붙어 있고 그 위로 간판이 보입니다. 옆에는 스쿠터가 주차되어 있습니다.",
      "entities": "신경희와 나상혁 모두 제공된 캐릭터 레퍼런스의 얼굴, 헤어스타일, 의상과 정확히 일치합니다. 커피숍, 간판, 스쿠터 등 배경 요소가 프롬프트대로 배치되어 있습니다.",
      "hard_violations": [],
      "physics": "신경희의 두 손이 나상혁의 재킷을 물리적으로 움켜쥐고 밀고 있으며, 두 사람 모두 지면에 안정적으로 서 있는 자세를 취하고 있습니다."
     },
     {
      "label": "B",
      "direction": "신경희의 시선은 나상혁을 향하지만, 손은 가슴을 밀치지 않고 방어하듯 허공에 떠 있습니다. 나상혁은 카메라를 등진 채 신분증을 내밀고 있습니다.",
      "built_space": "배경에 커피숍 입구와 스쿠터, 간판이 보이나 유리문의 '미시오' 스티커가 누락되었습니다.",
      "entities": "신경희의 외형은 레퍼런스와 일치하나 나상혁은 뒷모습만 보이며, 프롬프트에 없는 이전 샷의 신분증이 그대로 남아 있습니다.",
      "hard_violations": [
       "신경희가 나상혁의 가슴팍을 밀치는 동작을 전혀 구현하지 않음",
       "지정된 카메라 구도(측면 쓰리쿼터 앵글)를 무시하고 이전 샷의 구도를 그대로 복사함"
      ],
      "physics": "나상혁의 손이 신분증을 들고 있으며, 신경희의 두 손은 허공에 올려진 채 대상을 밀치고 있지 않습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh4"
  }
 },
 "S56sh6::cine": {
  "applied": true,
  "fingerprint": "d5f94cd04aa1413233b01ac1d1b61c56d07d8a39cf0cd048dc0a91b4f4c4f7bf",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S56sh6_sel.png",
  "source_sha256": "56a3a07ec3b175db41c1fc4f233f34a4d289c0037548918f170a3c38e37e68dc",
  "file": "S56sh6_cine.png",
  "latency_ms": 12095
 },
 "S56sh8::signage": {
  "fp": "ceb2c1890f0f0f5f",
  "inscriptions": [
   {
    "surface_native": "다방 간판",
    "text_native": "은하다방",
    "reason_ko": "인물들이 서 있는 다방 입구의 분위기를 살리고 공간의 정체성을 명확하게 전달하기 위해 간판의 상호 표기가 필요합니다."
   }
  ]
 },
 "S56sh8": {
  "input_fingerprint": "a1bca404c31f97ab",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다방 입구 반대편으로 한쪽 발을 내디딘 mid-stride 자세인 신경희의 뒷모습을 씁쓸하게 바라보는 서의용과 나상혁의 전경\n\nLOCATION (lock): On the street outside the coffeehouse, looking away from its entrance after the departing woman. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From shoulder height a few steps behind the two men, hold a static medium-wide frame with 서의용 at the left foreground and 나상혁 at the right foreground, both looking past the entrance toward 신경희. Their shoulders create an open central corridor that preserves 신경희's full rear-facing figure in the background, one foot planted ahead as she leaves toward the opposite side of the entrance.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, foreground, looks toward departing woman; 나상혁 in the middle-right of the frame, foreground, looks toward departing woman; 신경희 in the upper-center of the frame, background, moves toward opposite side of the entrance.\n- KEY BACKGROUND ELEMENTS: 다방 입구 (Visible; opening state is not specified) — The entrance is seen obliquely beyond the two foreground men; used as Marks the threshold behind the departing figure and clarifies her direction of travel; 다방 간판 (Visible) — Its identifying face is visible to the camera; used as Provides location identification above or beside the entrance without competing with the figures; 스쿠터 (Stopped at the entrance) — Its side is seen from the men's side of the street; used as A restrained background scale reference near the entrance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나상혁, 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same quiet street, cafe entrance, daylight, and the three people's clothing. Exclude the pushing impact and show the woman walking away while the two investigators remain.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다방 간판: \"은하다방\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다방 입구 반대편으로 한쪽 발을 내디딘 mid-stride 자세인 신경희의 뒷모습을 씁쓸하게 바라보는 서의용과 나상혁의 전경\n\nLOCATION (lock): On the street outside the coffeehouse, looking away from its entrance after the departing woman. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From shoulder height a few steps behind the two men, hold a static medium-wide frame with 서의용 at the left foreground and 나상혁 at the right foreground, both looking past the entrance toward 신경희. Their shoulders create an open central corridor that preserves 신경희's full rear-facing figure in the background, one foot planted ahead as she leaves toward the opposite side of the entrance.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, foreground, looks toward departing woman; 나상혁 in the middle-right of the frame, foreground, looks toward departing woman; 신경희 in the upper-center of the frame, background, moves toward opposite side of the entrance.\n- KEY BACKGROUND ELEMENTS: 다방 입구 (Visible; opening state is not specified) — The entrance is seen obliquely beyond the two foreground men; used as Marks the threshold behind the departing figure and clarifies her direction of travel; 다방 간판 (Visible) — Its identifying face is visible to the camera; used as Provides location identification above or beside the entrance without competing with the figures; 스쿠터 (Stopped at the entrance) — Its side is seen from the men's side of the street; used as A restrained background scale reference near the entrance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나상혁, 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same quiet street, cafe entrance, daylight, and the three people's clothing. Exclude the pushing impact and show the woman walking away while the two investigators remain.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다방 간판: \"은하다방\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 다방 입구 반대편으로 한쪽 발을 내디딘 mid-stride 자세인 신경희의 뒷모습을 씁쓸하게 바라보는 서의용과 나상혁의 전경\n\nLOCATION (lock): On the street outside the coffeehouse, looking away from its entrance after the departing woman. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From shoulder height a few steps behind the two men, hold a static medium-wide frame with 서의용 at the left foreground and 나상혁 at the right foreground, both looking past the entrance toward 신경희. Their shoulders create an open central corridor that preserves 신경희's full rear-facing figure in the background, one foot planted ahead as she leaves toward the opposite side of the entrance.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 서의용 in the middle-left of the frame, foreground, looks toward departing woman; 나상혁 in the middle-right of the frame, foreground, looks toward departing woman; 신경희 in the upper-center of the frame, background, moves toward opposite side of the entrance.\n- KEY BACKGROUND ELEMENTS: 다방 입구 (Visible; opening state is not specified) — The entrance is seen obliquely beyond the two foreground men; used as Marks the threshold behind the departing figure and clarifies her direction of travel; 다방 간판 (Visible) — Its identifying face is visible to the camera; used as Provides location identification above or beside the entrance without competing with the figures; 스쿠터 (Stopped at the entrance) — Its side is seen from the men's side of the street; used as A restrained background scale reference near the entrance.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime illumination with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나상혁, 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same quiet street, cafe entrance, daylight, and the three people's clothing. Exclude the pushing impact and show the woman walking away while the two investigators remain.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 다방 간판: \"은하다방\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전경의 두 남자(서의용, 나상혁)는 카메라를 등지고 서서 배경에서 멀어지는 여성(신경희)을 똑바로 바라보고 있음. 여성은 카메라 반대편을 향해 걷고 있음.",
    "built_space": "우측 중경에 다방 입구가 위치해 있으며, 그 위에 '은하다방'이라는 텍스트가 적힌 간판이 분명하게 보임. 입구 앞 도로변에는 스쿠터 한 대가 주차되어 있음.",
    "entities": "왼쪽 전경의 서의용(비니와 가죽 재킷), 오른쪽 전경의 나상혁(네이비 재킷), 배경의 신경희(패턴 셔츠와 어두운 바지) 모두 레퍼런스의 복장과 신체적 특징을 정확하게 반영하고 있음.",
    "hard_violations": [],
    "physics": "여성은 한쪽 발을 내딛고 다른 쪽 발로 지면을 지지하는 걷는 자세(mid-stride)를 안정적으로 취하고 있으며, 두 남자 역시 지면에 자연스럽게 서 있음. 공중에 뜬 물체나 부자연스러운 물리적 오류는 없음."
   },
   {
    "label": "B",
    "direction": "여성은 카메라 반대편을 향해 잘 걸어가고 있으나, 전경의 두 남자가 배경의 여성을 향하지 않고 서로를 향해(측면으로) 고개를 돌려 시선을 교환하는 듯한 엉뚱한 방향성을 보임.",
    "built_space": "우측에 다방 입구와 '은하다방' 간판이 위치하며, 입구 앞쪽으로 스쿠터가 정차되어 있어 프롬프트가 요구한 배경 요소를 포함함.",
    "entities": "세 인물 모두 프롬프트와 레퍼런스에서 요구한 의상, 헤어스타일, 외형을 올바르게 갖추고 있음.",
    "hard_violations": [],
    "physics": "걷고 있는 여성의 다리 지지 상태와 서 있는 남자들의 무게 중심이 지면에 안정적으로 고정되어 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 9,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "두 남자가 멀어지는 여성을 바라보는 완벽한 후면 구도와 시선 처리, 정확한 간판 텍스트를 통해 프롬프트의 요구사항을 가장 충실하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "배경의 여성을 바라봐야 할 두 남자가 서로를 향해 고개를 돌리고 있어 명시된 시선(gaze) 및 연출 지시를 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 나상혁, 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S56sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:843594>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "배경 중앙에 있는 신경희의 두 다리 사이에 세 번째 발(부츠)이 공중에 떠 있어 해부학적으로 불가능한 모습입니다.",
     "fix_en": "Erase the extra floating boot between the woman's legs, replacing it with the empty asphalt street surface. Preserve the two foreground men and their positions, the woman's figure and clothing, the set, the buildings, signage, parked scooter, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 샷(Previous Shot Still)의 연속성 지시와 달리 다방 외벽 재질, 창틀 디자인, 주차된 스쿠터의 기종과 색상이 이전 장면과 완전히 다르게 변형되었습니다.",
     "fix_en": "Change the background coffeehouse facade and the scooter to match the white tile, wooden doorframe, and white scooter from the previous still.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "신경희가 프레임 중앙에서 다방 입구 반대편으로 떠나는 게 아니라 정면 먼 다방을 향해 걸어가고 있다.",
     "fix_en": "Replace the background coffeehouse with an open street continuing into the distance to make the woman appear to be walking away from the entrance. Preserve the two foreground men and their positions, the woman's rear-facing figure and clothing, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "다방 입구가 두 남자 너머 비스듬한 문턱이 아니라 화면 상단 정면 상점 전면으로 놓여 있다.",
     "fix_en": "Relocate the coffeehouse entrance from the direct background to an oblique angle in the midground.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "배경 중앙에 있는 신경희의 두 다리 사이에 세 번째 발(부츠)이 공중에 떠 있어 해부학적으로 불가능한 모습입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 샷(Previous Shot Still)의 연속성 지시와 달리 다방 외벽 재질, 창틀 디자인, 주차된 스쿠터의 기종과 색상이 이전 장면과 완전히 다르게 변형되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "신경희가 프레임 중앙에서 다방 입구 반대편으로 떠나는 게 아니라 정면 먼 다방을 향해 걸어가고 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경 거리와 다방 건물이 이전 스틸의 그 장소·입구가 아닌 전혀 다른 골목으로 바뀌어 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "다방 입구가 두 남자 너머 비스듬한 문턱이 아니라 화면 상단 정면 상점 전면으로 놓여 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "입구 옆 스쿠터가 이전 스틸의 흰·검정 스쿠터가 아니고 먼 곳의 짙은색 스쿠터이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 샷(Previous Shot Still)의 연속성 지시와 달리 다방 외벽 재질, 창틀 디자인, 주차된 스쿠터의 기종과 색상이 이전 장면과 완전히 다르게 변형되었습니다.",
    "fix_en": "Change the background coffeehouse facade and the scooter to match the white tile, wooden doorframe, and white scooter from the previous still.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "다방 입구가 두 남자 너머 비스듬한 문턱이 아니라 화면 상단 정면 상점 전면으로 놓여 있다.",
    "fix_en": "Relocate the coffeehouse entrance from the direct background to an oblique angle in the midground.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Erase the extra floating boot between the woman's legs, replacing it with the empty asphalt street surface. Preserve the two foreground men and their positions, the woman's figure and clothing, the set, the buildings, signage, parked scooter, lighting, and framing.\n- Replace the background coffeehouse with an open street continuing into the distance to make the woman appear to be walking away from the entrance. Preserve the two foreground men and their positions, the woman's rear-facing figure and clothing, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 후면 구도, 장소의 연속성, 간판 텍스트(은하다방)를 훌륭하게 구현했으나, 신경희의 다리에 여분의 세 번째 발이 생성되는 치명적인 해부학적 오류가 있어 아쉽습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "카메라 뷰포인트가 반전되어 두 남자가 정면을 응시하고 있으며, 지정된 다방 배경 대신 전혀 다른 현대식 거리를 생성하여 핵심 지시 사항을 모두 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "서의용과 나상혁은 카메라를 등진 채 앞쪽 다방 입구를 지나 멀어지는 신경희를 향해 시선을 두고 있습니다. 카메라의 시선과 인물들의 방향이 지시문과 정확히 일치합니다.",
      "built_space": "이전 샷과 동일한 낡은 다방 외관, 미닫이문, 멈춰 있는 스쿠터가 배경에 정확히 배치되어 있으며, 간판의 위치도 자연스럽습니다.",
      "entities": "서의용(비니, 가죽 재킷), 나상혁(네이비 재킷), 신경희(패턴 상의, 청바지) 모두 레퍼런스의 특징과 복장을 잘 반영했습니다. 다방 간판에 '은하다방'이라는 텍스트가 정확히 적혀 있습니다.",
      "hard_violations": [
       "physically impossible anatomy (신경희의 왼쪽 다리 아래에 바닥을 딛고 있는 여분의 세 번째 발/부츠가 생성됨)"
      ],
      "physics": "두 남자는 바닥에 안정적으로 서 있으나, 신경희의 걷는 자세에서 뒤로 뻗은 왼쪽 다리에 들려 있는 부츠 외에도 그 아래 바닥에 닿아 있는 또 다른 부츠가 존재하여 해부학적으로 불가능한 모습입니다."
     },
     {
      "label": "B",
      "direction": "서의용과 나상혁이 신경희를 바라보지 않고 카메라 정면을 응시하고 있습니다. 카메라 역시 두 남자의 뒤가 아닌 앞에 위치하여 프롬프트의 시점과 방향 지시를 완전히 위반했습니다.",
      "built_space": "이전 샷의 낡은 다방 거리가 아닌 현대식 도로와 건물들이 배치되었으며, 지정된 장소의 연속성이 전혀 지켜지지 않았습니다.",
      "entities": "세 인물의 의상과 외형은 레퍼런스를 따르고 있으나, 지정된 한글 '은하다방' 간판이 없고 영문 'CAFE' 간판이 배치되었습니다. 신경희는 화면 좌측으로 지나치게 치우쳐 있습니다.",
      "hard_violations": [
       "staging violation (카메라 위치와 인물들의 시선/방향이 프롬프트의 명시적 지시와 정반대로 구현됨)",
       "location lock completely failed (이전 샷과 일치해야 할 다방 배경이 엉뚱한 건물로 대체됨)"
      ],
      "physics": "세 인물 모두 두 발로 바닥을 딛고 있거나 걷는 자세를 자연스럽게 유지하며 올바르게 지지받고 있습니다."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 후면 구도, 장소의 연속성, 간판 텍스트(은하다방)를 훌륭하게 구현했으나, 신경희의 다리에 여분의 세 번째 발이 생성되는 치명적인 해부학적 오류가 있어 아쉽습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "카메라 뷰포인트가 반전되어 두 남자가 정면을 응시하고 있으며, 지정된 다방 배경 대신 전혀 다른 현대식 거리를 생성하여 핵심 지시 사항을 모두 위반했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "서의용과 나상혁은 카메라를 등진 채 앞쪽 다방 입구를 지나 멀어지는 신경희를 향해 시선을 두고 있습니다. 카메라의 시선과 인물들의 방향이 지시문과 정확히 일치합니다.",
      "built_space": "이전 샷과 동일한 낡은 다방 외관, 미닫이문, 멈춰 있는 스쿠터가 배경에 정확히 배치되어 있으며, 간판의 위치도 자연스럽습니다.",
      "entities": "서의용(비니, 가죽 재킷), 나상혁(네이비 재킷), 신경희(패턴 상의, 청바지) 모두 레퍼런스의 특징과 복장을 잘 반영했습니다. 다방 간판에 '은하다방'이라는 텍스트가 정확히 적혀 있습니다.",
      "hard_violations": [
       "physically impossible anatomy (신경희의 왼쪽 다리 아래에 바닥을 딛고 있는 여분의 세 번째 발/부츠가 생성됨)"
      ],
      "physics": "두 남자는 바닥에 안정적으로 서 있으나, 신경희의 걷는 자세에서 뒤로 뻗은 왼쪽 다리에 들려 있는 부츠 외에도 그 아래 바닥에 닿아 있는 또 다른 부츠가 존재하여 해부학적으로 불가능한 모습입니다."
     },
     {
      "label": "B",
      "direction": "서의용과 나상혁이 신경희를 바라보지 않고 카메라 정면을 응시하고 있습니다. 카메라 역시 두 남자의 뒤가 아닌 앞에 위치하여 프롬프트의 시점과 방향 지시를 완전히 위반했습니다.",
      "built_space": "이전 샷의 낡은 다방 거리가 아닌 현대식 도로와 건물들이 배치되었으며, 지정된 장소의 연속성이 전혀 지켜지지 않았습니다.",
      "entities": "세 인물의 의상과 외형은 레퍼런스를 따르고 있으나, 지정된 한글 '은하다방' 간판이 없고 영문 'CAFE' 간판이 배치되었습니다. 신경희는 화면 좌측으로 지나치게 치우쳐 있습니다.",
      "hard_violations": [
       "staging violation (카메라 위치와 인물들의 시선/방향이 프롬프트의 명시적 지시와 정반대로 구현됨)",
       "location lock completely failed (이전 샷과 일치해야 할 다방 배경이 엉뚱한 건물로 대체됨)"
      ],
      "physics": "세 인물 모두 두 발로 바닥을 딛고 있거나 걷는 자세를 자연스럽게 유지하며 올바르게 지지받고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "두 남자의 어깨너머로 멀어지는 여자를 포착한 완벽한 샷 구도를 보여주며, 이전 샷의 배경을 정확히 유지하고 요구된 '은하다방' 간판까지 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "카메라가 두 남자의 뒤에 위치해야 한다는 명확한 구도 지시를 위반하고 정면에서 촬영했으며, 배경 장소와 간판 텍스트(CAFE)도 지정된 내용과 다릅니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 남자는 정면으로 카메라 렌즈를 응시하고 있습니다. 신경희는 왼쪽 배경에서 화면 밖을 향해 걸어가고 있습니다.",
      "built_space": "이전 샷의 장소와 일치하지 않는 새로운 거리 형태입니다. 왼쪽에 입구가 있으며, 간판에는 지정되지 않은 영문 'CAFE'와 'OPEN'이 적혀 있습니다.",
      "entities": "세 인물의 복장은 캐릭터 레퍼런스와 일치합니다. 하지만 요구된 '은하다방' 간판은 존재하지 않습니다.",
      "hard_violations": [
       "카메라 시점 위반 (두 남자의 뒤에서 앞을 바라보는 오버더숄더 형태의 구도가 아닌 정면 구도)",
       "장소 불일치 (이전 샷 스틸컷의 다방 외관 및 거리 배경과 전혀 다름)"
      ],
      "physics": "두 남자는 움직임 없이 양발을 땅에 대고 정자세로 서 있습니다. 신경희는 한쪽 발을 내디딘 걷는 자세를 취하고 있으며 물리적으로 안정적입니다."
     },
     {
      "label": "B",
      "direction": "두 남자는 카메라를 등지고 서서 프레임 중앙 멀리 걸어가는 신경희의 뒷모습을 향해 시선을 두고 있습니다.",
      "built_space": "이전 샷 스틸컷과 정확히 일치하는 거리와 다방 입구 외관입니다. 지정된 스쿠터가 입구 앞 원래 위치에 세워져 있습니다.",
      "entities": "세 인물 모두 레퍼런스의 복장을 착용한 상태로 뒷모습이 정확하게 표현되었습니다. 다방 간판에는 지시된 '은하다방' 텍스트가 한국어로 올바르게 적혀 있습니다.",
      "hard_violations": [],
      "physics": "서의용과 나상혁은 체중을 지탱하며 자연스럽게 서 있습니다. 신경희는 한쪽 발을 들고 앞으로 내딛는 역동적인 mid-stride 자세를 지지점의 어색함 없이 잘 보여줍니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "두 남자의 어깨너머로 멀어지는 여자를 포착한 완벽한 샷 구도를 보여주며, 이전 샷의 배경을 정확히 유지하고 요구된 '은하다방' 간판까지 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "카메라가 두 남자의 뒤에 위치해야 한다는 명확한 구도 지시를 위반하고 정면에서 촬영했으며, 배경 장소와 간판 텍스트(CAFE)도 지정된 내용과 다릅니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "두 남자는 정면으로 카메라 렌즈를 응시하고 있습니다. 신경희는 왼쪽 배경에서 화면 밖을 향해 걸어가고 있습니다.",
      "built_space": "이전 샷의 장소와 일치하지 않는 새로운 거리 형태입니다. 왼쪽에 입구가 있으며, 간판에는 지정되지 않은 영문 'CAFE'와 'OPEN'이 적혀 있습니다.",
      "entities": "세 인물의 복장은 캐릭터 레퍼런스와 일치합니다. 하지만 요구된 '은하다방' 간판은 존재하지 않습니다.",
      "hard_violations": [
       "카메라 시점 위반 (두 남자의 뒤에서 앞을 바라보는 오버더숄더 형태의 구도가 아닌 정면 구도)",
       "장소 불일치 (이전 샷 스틸컷의 다방 외관 및 거리 배경과 전혀 다름)"
      ],
      "physics": "두 남자는 움직임 없이 양발을 땅에 대고 정자세로 서 있습니다. 신경희는 한쪽 발을 내디딘 걷는 자세를 취하고 있으며 물리적으로 안정적입니다."
     },
     {
      "label": "A",
      "direction": "두 남자는 카메라를 등지고 서서 프레임 중앙 멀리 걸어가는 신경희의 뒷모습을 향해 시선을 두고 있습니다.",
      "built_space": "이전 샷 스틸컷과 정확히 일치하는 거리와 다방 입구 외관입니다. 지정된 스쿠터가 입구 앞 원래 위치에 세워져 있습니다.",
      "entities": "세 인물 모두 레퍼런스의 복장을 착용한 상태로 뒷모습이 정확하게 표현되었습니다. 다방 간판에는 지시된 '은하다방' 텍스트가 한국어로 올바르게 적혀 있습니다.",
      "hard_violations": [],
      "physics": "서의용과 나상혁은 체중을 지탱하며 자연스럽게 서 있습니다. 신경희는 한쪽 발을 들고 앞으로 내딛는 역동적인 mid-stride 자세를 지지점의 어색함 없이 잘 보여줍니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 3
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S56sh6"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S56sh8::cine": {
  "applied": true,
  "fingerprint": "fcd1a65ee2925f319bf2afb8d4b343f9801d2c0caffa383cc23a95053f0549d2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S56sh8_sel.png",
  "source_sha256": "8aa6fc143b5083d74becd2a998b6b758de951600aac99c4008dbca1b7e2ec187",
  "file": "S56sh8_cine.png",
  "latency_ms": 11175
 },
 "S57sh3::signage": {
  "fp": "0b951e2a11b4a8aa",
  "inscriptions": []
 },
 "S57sh3": {
  "input_fingerprint": "eb8bd6818f45769e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽으로 상체를 바짝 내민 채 자신만만한 표정으로 손가락을 뻗은 고원효의 측면\n\nLOCATION (lock): Inside a small prosecution interview room, across the table from the prosecutor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height beside the table midpoint, dolly inward on a slightly widened medium side view of 고원효, positioned on the right and leaning toward 장원섭 at frame left. Keep his extended finger, forward upper body, and confident expression together in focus while 장원섭 remains across the table as the recipient of the argument.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 고원효 in the middle-right of the frame, midground, points to 장원섭 across the table; 장원섭 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 테이블 (Positioned between the two men) — Its long edge recedes between 고원효 and 장원섭; used as Maintains the interrogation axis and separates the two seated men; 커피잔 (Present on the table); used as Connects 고원효's prior relaxed behavior to his present confidence without becoming the focal point.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to a daytime investigation room, rendered in a subdued neutral-to-cool range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 고원효 (Korean 남성, 40대 중반, 둥근 얼굴, 짧은 검은색 머리) — wearing: 교도소 재소자가 착용하는 푸른색 계열의 수의. 가슴에 죄수번호표가 부착되어 있으며 거친 면 소재이다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽으로 상체를 바짝 내민 채 자신만만한 표정으로 손가락을 뻗은 고원효의 측면\n\nLOCATION (lock): Inside a small prosecution interview room, across the table from the prosecutor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height beside the table midpoint, dolly inward on a slightly widened medium side view of 고원효, positioned on the right and leaning toward 장원섭 at frame left. Keep his extended finger, forward upper body, and confident expression together in focus while 장원섭 remains across the table as the recipient of the argument.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 고원효 in the middle-right of the frame, midground, points to 장원섭 across the table; 장원섭 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 테이블 (Positioned between the two men) — Its long edge recedes between 고원효 and 장원섭; used as Maintains the interrogation axis and separates the two seated men; 커피잔 (Present on the table); used as Connects 고원효's prior relaxed behavior to his present confidence without becoming the focal point.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to a daytime investigation room, rendered in a subdued neutral-to-cool range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 고원효 (Korean 남성, 40대 중반, 둥근 얼굴, 짧은 검은색 머리) — wearing: 교도소 재소자가 착용하는 푸른색 계열의 수의. 가슴에 죄수번호표가 부착되어 있으며 거친 면 소재이다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽으로 상체를 바짝 내민 채 자신만만한 표정으로 손가락을 뻗은 고원효의 측면\n\nLOCATION (lock): Inside a small prosecution interview room, across the table from the prosecutor. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated chest height beside the table midpoint, dolly inward on a slightly widened medium side view of 고원효, positioned on the right and leaning toward 장원섭 at frame left. Keep his extended finger, forward upper body, and confident expression together in focus while 장원섭 remains across the table as the recipient of the argument.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 고원효 in the middle-right of the frame, midground, points to 장원섭 across the table; 장원섭 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 테이블 (Positioned between the two men) — Its long edge recedes between 고원효 and 장원섭; used as Maintains the interrogation axis and separates the two seated men; 커피잔 (Present on the table); used as Connects 고원효's prior relaxed behavior to his present confidence without becoming the focal point.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to a daytime investigation room, rendered in a subdued neutral-to-cool range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 고원효 (Korean 남성, 40대 중반, 둥근 얼굴, 짧은 검은색 머리) — wearing: 교도소 재소자가 착용하는 푸른색 계열의 수의. 가슴에 죄수번호표가 부착되어 있으며 거친 면 소재이다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "고원효의 시선과 가리키는 손가락이 맞은편의 장원섭을 정확히 향하고 있음.",
    "built_space": "취조실 내부의 테이블을 사이에 두고 두 사람이 마주 앉아 지정된 위치를 지키고 있음.",
    "entities": "고원효, 장원섭, 서류, 커피잔 등이 모두 존재하나, 고원효의 명찰에 '수사기록'이라는 잘못된 텍스트가 표시됨.",
    "hard_violations": [
     "고원효의 죄수번호표에 테이블 위 서류의 텍스트인 '수사기록'이 유출되어 표기됨 (leaked text)"
    ],
    "physics": "두 인물 모두 의자에 안정적으로 앉아 있으며, 고원효가 뻗은 팔과 자세가 중력에 맞게 자연스럽게 지지됨."
   },
   {
    "label": "B",
    "direction": "고원효의 시선과 가리키는 손가락이 맞은편의 장원섭을 정확히 향하고 있음.",
    "built_space": "취조실 내부의 테이블을 사이에 두고 두 사람이 마주 앉아 지정된 위치를 지키고 있음.",
    "entities": "고원효, 장원섭, 서류, 커피잔 등이 모두 존재하나, 고원효의 명찰에 '수사기록'이라는 잘못된 텍스트가 표시됨.",
    "hard_violations": [
     "고원효의 죄수번호표에 테이블 위 서류의 텍스트인 '수사기록'이 유출되어 표기됨 (leaked text)"
    ],
    "physics": "두 인물 모두 의자에 안정적으로 앉아 있으며, 고원효가 뻗은 팔과 자세가 중력에 맞게 자연스럽게 지지됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 4,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "구도와 인물의 자세는 프롬프트를 잘 따랐으나, 죄수번호표에 서류의 텍스트('수사기록')가 유출되는 치명적 오류가 발생함."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "A와 동일하게 텍스트 유출 오류가 있으며, 커피잔의 위치가 덜 자연스러워 A보다 약간 낮은 평가를 받음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S53sh15_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "오른쪽 인물(고원효)의 가슴에 부착된 죄수번호표에 숫자가 아닌 '수사기록'이라는 텍스트가 잘못 들어가 있습니다.",
     "fix_en": "Replace the Korean text on the white rectangular tag on the right man's chest with indistinct, blurred numerals to represent a prisoner number. Preserve both men's faces, poses, clothing, the table, all props, and the background unchanged.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "오른쪽 고원효가 상체를 장원섭 쪽으로 바짝 내밀지 않고 허리를 세운 채 앉아 손가락만 뻗고 있다",
     "fix_en": "Adjust the right man's posture so his torso leans far forward across the table. Preserve his pointing hand, facial expression, blue uniform, the left man, all props, and the background.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "오른쪽 인물(고원효)의 가슴에 부착된 죄수번호표에 숫자가 아닌 '수사기록'이라는 텍스트가 잘못 들어가 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "오른쪽 고원효가 상체를 장원섭 쪽으로 바짝 내밀지 않고 허리를 세운 채 앉아 손가락만 뻗고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 고원효 가슴의 죄수번호표에 번호가 아니라 '수사기록'이 적혀 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "오른쪽 고원효가 상체를 장원섭 쪽으로 바짝 내밀지 않고 허리를 세운 채 앉아 손가락만 뻗고 있다",
    "fix_en": "Adjust the right man's posture so his torso leans far forward across the table. Preserve his pointing hand, facial expression, blue uniform, the left man, all props, and the background.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the Korean text on the white rectangular tag on the right man's chest with indistinct, blurred numerals to represent a prisoner number. Preserve both men's faces, poses, clothing, the table, all props, and the background unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 구도와 인물의 의상 및 소품을 충실히 구현하였으며, 죄수번호표의 텍스트도 자연스럽게 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물과 구도는 좋으나, 고원효의 죄수번호표에 서류의 텍스트('수사기록')가 잘못 유출되어 표기된 치명적인 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "고원효가 맞은편에 앉은 장원섭을 향해 손가락을 가리키고 있음.",
      "built_space": "취조실 내부. 두 사람이 마주 앉은 테이블 위에 명패, 서류, 종이봉투, 커피잔이 올바른 위치에 놓여 있음.",
      "entities": "좌측에는 정장 차림의 장원섭, 우측에는 푸른색 수의와 가슴에 숫자 명찰(48291)을 단 고원효가 프롬프트에 맞게 묘사됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 안정적으로 앉아 있으며, 팔과 손이 테이블과 자연스럽게 접촉되어 지지됨."
     },
     {
      "label": "A",
      "direction": "고원효가 맞은편에 앉은 장원섭을 향해 손가락을 가리키고 있음.",
      "built_space": "취조실 내부. 테이블 위에 명패, 서류, 종이봉투, 커피잔이 배치되어 있음.",
      "entities": "좌측의 장원섭과 우측의 고원효가 나타나 있으나, 고원효의 수의 가슴 명찰에 숫자가 아닌 '수사기록'이라는 주변 사물의 텍스트가 잘못 기재됨.",
      "hard_violations": [
       "수의 명찰에 서류의 텍스트('수사기록')가 유출되어 잘못 표기됨"
      ],
      "physics": "두 인물 모두 의자에 안정적으로 앉아 있으며 상체와 팔이 적절히 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "요구된 구도와 인물의 의상 및 소품을 충실히 구현하였으며, 죄수번호표의 텍스트도 자연스럽게 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "인물과 구도는 좋으나, 고원효의 죄수번호표에 서류의 텍스트('수사기록')가 잘못 유출되어 표기된 치명적인 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "고원효가 맞은편에 앉은 장원섭을 향해 손가락을 가리키고 있음.",
      "built_space": "취조실 내부. 두 사람이 마주 앉은 테이블 위에 명패, 서류, 종이봉투, 커피잔이 올바른 위치에 놓여 있음.",
      "entities": "좌측에는 정장 차림의 장원섭, 우측에는 푸른색 수의와 가슴에 숫자 명찰(48291)을 단 고원효가 프롬프트에 맞게 묘사됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 안정적으로 앉아 있으며, 팔과 손이 테이블과 자연스럽게 접촉되어 지지됨."
     },
     {
      "label": "A",
      "direction": "고원효가 맞은편에 앉은 장원섭을 향해 손가락을 가리키고 있음.",
      "built_space": "취조실 내부. 테이블 위에 명패, 서류, 종이봉투, 커피잔이 배치되어 있음.",
      "entities": "좌측의 장원섭과 우측의 고원효가 나타나 있으나, 고원효의 수의 가슴 명찰에 숫자가 아닌 '수사기록'이라는 주변 사물의 텍스트가 잘못 기재됨.",
      "hard_violations": [
       "수의 명찰에 서류의 텍스트('수사기록')가 유출되어 잘못 표기됨"
      ],
      "physics": "두 인물 모두 의자에 안정적으로 앉아 있으며 상체와 팔이 적절히 지지됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트가 요구한 고원효의 인상착의(푸른색 수의, 죄수번호표), 자세(손가락으로 가리키며 상체를 내민 모습), 자신만만한 표정을 정확하게 구현했으며, 레퍼런스의 환경과 인물도 잘 유지했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "고원효의 가슴에 있는 죄수번호표에 책상 위 서류의 텍스트인 '수사기록'이 잘못 유출되어 텍스트 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "오른쪽의 고원효가 왼쪽의 장원섭을 향해 오른손 집게손가락을 뻗어 가리키고 있음",
      "built_space": "조사실 내부, 테이블을 사이에 두고 두 남자가 마주보고 앉아 있음",
      "entities": "장원섭(검사)의 외모, 복장, 명패, 서류, 빵 봉투 등은 레퍼런스와 일치. 고원효는 40대 중반 남성으로 푸른색 수의를 입고 가슴에 '48291' 번호표를 달고 있음. 테이블 위에 커피잔이 있음",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 앉아 체중이 지지되고 있으며, 고원효의 뻗은 팔은 어깨와 몸통에 의해 자연스럽게 지지됨"
     },
     {
      "label": "B",
      "direction": "오른쪽의 고원효가 왼쪽의 장원섭을 향해 오른손 집게손가락을 뻗어 가리키고 있음",
      "built_space": "조사실 내부, 테이블을 사이에 두고 두 남자가 마주보고 앉아 있음",
      "entities": "장원섭과 주변 사물은 레퍼런스와 일치. 고원효는 푸른색 수의를 입고 있으나 가슴의 명찰에 숫자 대신 '수사기록'이라는 글자가 적혀 있음. 커피잔이 테이블 위에 있음",
      "hard_violations": [
       "텍스트 유출 (죄수번호표에 '수사기록' 텍스트가 잘못 들어감)"
      ],
      "physics": "두 인물 모두 의자에 앉아 체중이 지지되고 있으며, 뻗은 팔과 자세도 물리적으로 자연스러움"
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프롬프트가 요구한 고원효의 인상착의(푸른색 수의, 죄수번호표), 자세(손가락으로 가리키며 상체를 내민 모습), 자신만만한 표정을 정확하게 구현했으며, 레퍼런스의 환경과 인물도 잘 유지했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "고원효의 가슴에 있는 죄수번호표에 책상 위 서류의 텍스트인 '수사기록'이 잘못 유출되어 텍스트 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "오른쪽의 고원효가 왼쪽의 장원섭을 향해 오른손 집게손가락을 뻗어 가리키고 있음",
      "built_space": "조사실 내부, 테이블을 사이에 두고 두 남자가 마주보고 앉아 있음",
      "entities": "장원섭(검사)의 외모, 복장, 명패, 서류, 빵 봉투 등은 레퍼런스와 일치. 고원효는 40대 중반 남성으로 푸른색 수의를 입고 가슴에 '48291' 번호표를 달고 있음. 테이블 위에 커피잔이 있음",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 앉아 체중이 지지되고 있으며, 고원효의 뻗은 팔은 어깨와 몸통에 의해 자연스럽게 지지됨"
     },
     {
      "label": "A",
      "direction": "오른쪽의 고원효가 왼쪽의 장원섭을 향해 오른손 집게손가락을 뻗어 가리키고 있음",
      "built_space": "조사실 내부, 테이블을 사이에 두고 두 남자가 마주보고 앉아 있음",
      "entities": "장원섭과 주변 사물은 레퍼런스와 일치. 고원효는 푸른색 수의를 입고 있으나 가슴의 명찰에 숫자 대신 '수사기록'이라는 글자가 적혀 있음. 커피잔이 테이블 위에 있음",
      "hard_violations": [
       "텍스트 유출 (죄수번호표에 '수사기록' 텍스트가 잘못 들어감)"
      ],
      "physics": "두 인물 모두 의자에 앉아 체중이 지지되고 있으며, 뻗은 팔과 자세도 물리적으로 자연스러움"
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S53sh15"
  }
 },
 "S57sh3::cine": {
  "applied": true,
  "fingerprint": "470fc36f7d813fd0c2e12566463684081b26e2784230e9a1e69d9311758f0a40",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S57sh3_sel.png",
  "source_sha256": "d5e5860ecb12a614ba1edd107656edafa0376ea7542971fd99a0680efbc8a9df",
  "file": "S57sh3_cine.png",
  "latency_ms": 11800
 },
 "S57sh6::signage": {
  "fp": "ca689b171bc54a7c",
  "inscriptions": [
   {
    "surface_native": "수사기록 표지",
    "text_native": "수사기록",
    "reason_ko": "검사 측 테이블 위에 놓인 서류 봉투에 '수사기록'이라는 문구를 표시하여 검찰 조사실이라는 공간적 맥락과 사실성을 더합니다."
   }
  ]
 },
 "S57sh6": {
  "input_fingerprint": "a4bd9c89b4058526",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 쪽으로 상체를 바짝 기울여 앉은 채 고원효를 향해 입을 벌린 장원섭의 상체\n\nLOCATION (lock): At the prosecutor's side of the table in the small interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Continue the push from just behind 고원효's near shoulder at slightly below 장원섭's seated eye line, tightening to 장원섭's upper body across the table. 장원섭 occupies the center and leans into his question with his mouth open, while only a narrow, soft-edged portion of 고원효's shoulder remains at the near right edge as the target of his scrutiny.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 테이블 (Between the two men) — The near edge crosses the lower frame and leads toward 장원섭; used as Forms a compressed horizontal boundary beneath 장원섭's forward-leaning torso.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime room, with controlled contrast and restrained neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same narrow interview room, table, institutional light, and prisoner's uniform. Exclude the prisoner's pointing gesture and frame the prosecutor leaning forward to question him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사기록 표지: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 쪽으로 상체를 바짝 기울여 앉은 채 고원효를 향해 입을 벌린 장원섭의 상체\n\nLOCATION (lock): At the prosecutor's side of the table in the small interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Continue the push from just behind 고원효's near shoulder at slightly below 장원섭's seated eye line, tightening to 장원섭's upper body across the table. 장원섭 occupies the center and leans into his question with his mouth open, while only a narrow, soft-edged portion of 고원효's shoulder remains at the near right edge as the target of his scrutiny.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 테이블 (Between the two men) — The near edge crosses the lower frame and leads toward 장원섭; used as Forms a compressed horizontal boundary beneath 장원섭's forward-leaning torso.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime room, with controlled contrast and restrained neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same narrow interview room, table, institutional light, and prisoner's uniform. Exclude the prisoner's pointing gesture and frame the prosecutor leaning forward to question him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사기록 표지: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 쪽으로 상체를 바짝 기울여 앉은 채 고원효를 향해 입을 벌린 장원섭의 상체\n\nLOCATION (lock): At the prosecutor's side of the table in the small interview room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Continue the push from just behind 고원효's near shoulder at slightly below 장원섭's seated eye line, tightening to 장원섭's upper body across the table. 장원섭 occupies the center and leans into his question with his mouth open, while only a narrow, soft-edged portion of 고원효's shoulder remains at the near right edge as the target of his scrutiny.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 테이블 (Between the two men) — The near edge crosses the lower frame and leads toward 장원섭; used as Forms a compressed horizontal boundary beneath 장원섭's forward-leaning torso.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient light appropriate to the daytime room, with controlled contrast and restrained neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same narrow interview room, table, institutional light, and prisoner's uniform. Exclude the prisoner's pointing gesture and frame the prosecutor leaning forward to question him.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사기록 표지: \"수사기록\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.429,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.571
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1571,
   "A": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "요청된 '상체를 바짝 기울이고 입을 벌린' 핵심 동작과 긴장감을 훌륭하게 구현했으나, 죄수복 앞면의 수감번호 패치가 어깨 뒤편에 잘못 나타나는 의상 오류가 있음."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "화면 구성과 소품, 텍스트 배치는 매우 깔끔하고 정확하나, 상체를 바짝 기울이거나 입을 벌리는 극적인 동작 지시가 너무 밋밋하게 표현되어 프롬프트의 의도를 살리지 못함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S57sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "테이블 위 '수사기록' 서류가 사용자(검사)가 아닌 카메라(죄수 시점)를 향해 정방향으로 놓여 있어, 소품이 사용자의 시선을 향해야 한다는 지시와 이전 컷의 배치 연속성을 위반했습니다.",
     "fix_en": "Rotate the document 180 degrees so it faces the seated man. Preserve all people, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "서류 표지의 '수사기록' 텍스트 글꼴이 이전 컷(명조체 계열)과 달리 고딕체 계열로 변경되었습니다.",
     "fix_en": "Change the document text font to a serif style. Preserve all people, clothing, set, lighting, and framing.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "배경 우측 문의 디자인에 이전 컷에는 없던 가로선 패널 장식이 생겼습니다.",
     "fix_en": "Remove the horizontal panel lines from the right background door. Preserve all people, clothing, set, lighting, and framing.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "오른쪽 가장자리에 고원효의 어깨 좁은 일부만 있어야 하는데 머리·귀·등까지 크게 들어와 있다",
     "fix_en": "Erase the head and ear on the right edge, replacing them with the background wall and leaving only a blurred shoulder. Preserve all people, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "장원섭이 테이블 쪽으로 상체를 바짝 기울이지 않고 비교적 곧게 앉아 있다",
     "fix_en": "Angle the man's torso forward to lean closely over the table. Preserve all people, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    },
    {
     "issue_ko": "카메라가 장원섭 앉은 눈높이보다 약간 아래가 아니라 테이블 윗면을 내려다보는 높이다",
     "fix_en": "Lower the camera to look up slightly from below the subject's eye level. Preserve all people, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 6,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "테이블 위 '수사기록' 서류가 사용자(검사)가 아닌 카메라(죄수 시점)를 향해 정방향으로 놓여 있어, 소품이 사용자의 시선을 향해야 한다는 지시와 이전 컷의 배치 연속성을 위반했습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 우측 근경에 고원효의 어깨만 좁게(narrow) 보여야 한다는 프레이밍 지시와 달리, 뒷머리와 귀까지 상대적으로 넓은 면적이 포함되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "서류 표지의 '수사기록' 텍스트 글꼴이 이전 컷(명조체 계열)과 달리 고딕체 계열로 변경되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "배경 우측 문의 디자인에 이전 컷에는 없던 가로선 패널 장식이 생겼습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "오른쪽 가장자리에 고원효의 어깨 좁은 일부만 있어야 하는데 머리·귀·등까지 크게 들어와 있다",
     "severity": "major"
    },
    {
     "issue_ko": "장원섭이 테이블 쪽으로 상체를 바짝 기울이지 않고 비교적 곧게 앉아 있다",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 장원섭 앉은 눈높이보다 약간 아래가 아니라 테이블 윗면을 내려다보는 높이다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 6,
  "fix_severity_skipped": [
   {
    "issue_ko": "테이블 위 '수사기록' 서류가 사용자(검사)가 아닌 카메라(죄수 시점)를 향해 정방향으로 놓여 있어, 소품이 사용자의 시선을 향해야 한다는 지시와 이전 컷의 배치 연속성을 위반했습니다.",
    "fix_en": "Rotate the document 180 degrees so it faces the seated man. Preserve all people, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "서류 표지의 '수사기록' 텍스트 글꼴이 이전 컷(명조체 계열)과 달리 고딕체 계열로 변경되었습니다.",
    "fix_en": "Change the document text font to a serif style. Preserve all people, clothing, set, lighting, and framing.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "배경 우측 문의 디자인에 이전 컷에는 없던 가로선 패널 장식이 생겼습니다.",
    "fix_en": "Remove the horizontal panel lines from the right background door. Preserve all people, clothing, set, lighting, and framing.",
    "severity": "minor",
    "observation_index": 3
   },
   {
    "issue_ko": "오른쪽 가장자리에 고원효의 어깨 좁은 일부만 있어야 하는데 머리·귀·등까지 크게 들어와 있다",
    "fix_en": "Erase the head and ear on the right edge, replacing them with the background wall and leaving only a blurred shoulder. Preserve all people, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "장원섭이 테이블 쪽으로 상체를 바짝 기울이지 않고 비교적 곧게 앉아 있다",
    "fix_en": "Angle the man's torso forward to lean closely over the table. Preserve all people, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   },
   {
    "issue_ko": "카메라가 장원섭 앉은 눈높이보다 약간 아래가 아니라 테이블 윗면을 내려다보는 높이다",
    "fix_en": "Lower the camera to look up slightly from below the subject's eye level. Preserve all people, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 6,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S57sh3"
  }
 },
 "S57sh6::cine": {
  "applied": true,
  "fingerprint": "953fe43952bc4b1d15da882bff6575e42b8c80241df1918c68c41a050d9e6bda",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S57sh6_sel.png",
  "source_sha256": "eaf97de1fd9922f6fed54ac57e5e3a99c4754526011d207d4817162a71d1661a",
  "file": "S57sh6_cine.png",
  "latency_ms": 12575
 },
 "S58sh1::signage": {
  "fp": "2aecbce9072d319e",
  "inscriptions": [
   {
    "surface_native": "수사관의 조끼",
    "text_native": "검찰",
    "reason_ko": "압수수색을 진행하는 검찰 수사관의 신분을 명확히 나타내어 상황의 법적 긴장감을 높이기 위해 조끼에 '검찰' 표기가 필요합니다."
   }
  ]
 },
 "era_assess::c3dd63c3f87d120f": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 가정집의 안방과 장롱",
    "search_terms_native": [
     "2000년대 한국 안방",
     "가정집 장롱 2000년대",
     "한국 빌라 방 인테리어 장판",
     "옛날 주택 방 장롱"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가 단어를 영어로 작성하지 마십시오.",
    "reason_ko": "2000년대 한국 서민 가정집 특유의 노란색 장판 바닥, 벽지 패턴, 그리고 안방을 차지하는 특유의 나무 장롱 스타일은 일반적인 서구식 침실이나 현대식 미니멀 인테리어와 확연히 다릅니다."
   }
  ]
 },
 "era_ref::6b6693ca31a97228": {
  "subject": "2000년대 한국 가정집의 안방과 장롱",
  "terms": [
   "2000년대 한국 안방",
   "가정집 장롱 2000년대",
   "한국 빌라 방 인테리어 장판",
   "옛날 주택 방 장롱"
  ],
  "queries": [
   [
    "2000년대 한국 가정집 안방 장롱 빌라 방 인테리어 장판",
    "옛날 한국 주택 방 장롱 장판 2000년대"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://www.filmmakers.co.kr/files/attach/images/2021/11/30/55f126f7ab84cbeb2b995106163dc378.jpg",
  "picked_reason_ko": "2000년대 한국 가정집 안방의 전형적인 장롱과 침대, 창호, 바닥 마감이 선명하게 보여 형태와 재료를 가장 잘 파악할 수 있다.",
  "sha256": "2a60a6d441f809013364939ecbacfcfa5253a335f0587698a1480f08d9fa5f64",
  "file": "eraref_6b6693ca31a97228.png"
 },
 "S58sh1::bgfirst_bg": {
  "input_fingerprint": "5b2708eb8dbd13a3",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어수선한 방 안, 악을 쓰며 삿대질을 하는 유경자의 앞을 온몸으로 막아선 검찰 수사관의 상체\n\nLOCATION (lock): Inside the cramped bedroom of the suspect's family home, amid a bed, wardrobe, and boxes being searched.\n\nTIME OF DAY (lock): day into night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At chest height behind and slightly beside the 검찰 수사관, use a restrained handheld medium frame across his oblique shoulder line as his upper body blocks 유경자. The investigator braces across the center-right while 유경자 presses forward from the left background with one arm raised in accusation; partial search activity remains deeper in the narrow room without displacing their confrontation.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 검찰 수사관 in the middle-right of the frame, foreground; 유경자 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 침대 (Being searched) — Only part of its side and top remain visible behind the figures; used as Establishes the cramped search area behind the confrontation; 옷장 (Being searched) — Its front faces into the room and is seen obliquely by the camera; used as Creates a vertical boundary in the deeper room and indicates another search area; 잡동사니 상자들 (Being searched) — Different sides are visible where the boxes have been opened for inspection; used as Provides irregular foreground and background layers around the searching figures; 증거물 상자 (Receiving letters, notebooks, and photographs) — Its open upper side is visible from the oblique camera position; used as Procedural evidence anchor kept in the deeper portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with moderate contrast and subdued neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 가정집의 안방과 장롱: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 어수선한 방 안, 악을 쓰며 삿대질을 하는 유경자의 앞을 온몸으로 막아선 검찰 수사관의 상체\n\nLOCATION (lock): Inside the cramped bedroom of the suspect's family home, amid a bed, wardrobe, and boxes being searched.\n\nTIME OF DAY (lock): day into night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At chest height behind and slightly beside the 검찰 수사관, use a restrained handheld medium frame across his oblique shoulder line as his upper body blocks 유경자. The investigator braces across the center-right while 유경자 presses forward from the left background with one arm raised in accusation; partial search activity remains deeper in the narrow room without displacing their confrontation.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 검찰 수사관 in the middle-right of the frame, foreground; 유경자 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 침대 (Being searched) — Only part of its side and top remain visible behind the figures; used as Establishes the cramped search area behind the confrontation; 옷장 (Being searched) — Its front faces into the room and is seen obliquely by the camera; used as Creates a vertical boundary in the deeper room and indicates another search area; 잡동사니 상자들 (Being searched) — Different sides are visible where the boxes have been opened for inspection; used as Provides irregular foreground and background layers around the searching figures; 증거물 상자 (Receiving letters, notebooks, and photographs) — Its open upper side is visible from the oblique camera position; used as Procedural evidence anchor kept in the deeper portion of the frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with moderate contrast and subdued neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 가정집의 안방과 장롱: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S58sh1__bgfirst_bg.png",
  "asset_id": "b939e106-558d-407c-8d99-9fef88379129",
  "input_asset_ids": [
   "c61025b6-e634-495e-9bb7-11dbeb53b187",
   "f5a33ae3-d58c-4ed6-b591-a001a102d09a"
  ],
  "era_research": {
   "subject": "2000년대 한국 가정집의 안방과 장롱",
   "queries": [
    [
     "2000년대 한국 가정집 안방 장롱 빌라 방 인테리어 장판",
     "옛날 한국 주택 방 장롱 장판 2000년대"
    ]
   ],
   "picked_url": "https://www.filmmakers.co.kr/files/attach/images/2021/11/30/55f126f7ab84cbeb2b995106163dc378.jpg",
   "sha256": "2a60a6d441f809013364939ecbacfcfa5253a335f0587698a1480f08d9fa5f64",
   "file": "eraref_6b6693ca31a97228.png"
  }
 },
 "S58sh1": {
  "input_fingerprint": "4dcc09ec619ba906",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 어수선한 방 안, 악을 쓰며 삿대질을 하는 유경자의 앞을 온몸으로 막아선 검찰 수사관의 상체\n\nLOCATION (lock): Inside the cramped bedroom of the suspect's family home, amid a bed, wardrobe, and boxes being searched. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height behind and slightly beside the 검찰 수사관, use a restrained handheld medium frame across his oblique shoulder line as his upper body blocks 유경자. The investigator braces across the center-right while 유경자 presses forward from the left background with one arm raised in accusation; partial search activity remains deeper in the narrow room without displacing their confrontation.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 검찰 수사관 in the middle-right of the frame, foreground; 유경자 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 침대 (Being searched) — Only part of its side and top remain visible behind the figures; used as Establishes the cramped search area behind the confrontation; 옷장 (Being searched) — Its front faces into the room and is seen obliquely by the camera; used as Creates a vertical boundary in the deeper room and indicates another search area; 잡동사니 상자들 (Being searched) — Different sides are visible where the boxes have been opened for inspection; used as Provides irregular foreground and background layers around the searching figures; 증거물 상자 (Receiving letters, notebooks, and photographs) — Its open upper side is visible from the oblique camera position; used as Procedural evidence anchor kept in the deeper portion of the frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with moderate contrast and subdued neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The searched room is left disturbed, while seized letters, notebooks, and photographs are collected in evidence boxes.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사관의 조끼: \"검찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 어수선한 방 안, 악을 쓰며 삿대질을 하는 유경자의 앞을 온몸으로 막아선 검찰 수사관의 상체\n\nLOCATION (lock): Inside the cramped bedroom of the suspect's family home, amid a bed, wardrobe, and boxes being searched. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height behind and slightly beside the 검찰 수사관, use a restrained handheld medium frame across his oblique shoulder line as his upper body blocks 유경자. The investigator braces across the center-right while 유경자 presses forward from the left background with one arm raised in accusation; partial search activity remains deeper in the narrow room without displacing their confrontation.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 검찰 수사관 in the middle-right of the frame, foreground; 유경자 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 침대 (Being searched) — Only part of its side and top remain visible behind the figures; used as Establishes the cramped search area behind the confrontation; 옷장 (Being searched) — Its front faces into the room and is seen obliquely by the camera; used as Creates a vertical boundary in the deeper room and indicates another search area; 잡동사니 상자들 (Being searched) — Different sides are visible where the boxes have been opened for inspection; used as Provides irregular foreground and background layers around the searching figures; 증거물 상자 (Receiving letters, notebooks, and photographs) — Its open upper side is visible from the oblique camera position; used as Procedural evidence anchor kept in the deeper portion of the frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with moderate contrast and subdued neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The searched room is left disturbed, while seized letters, notebooks, and photographs are collected in evidence boxes.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사관의 조끼: \"검찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 어수선한 방 안, 악을 쓰며 삿대질을 하는 유경자의 앞을 온몸으로 막아선 검찰 수사관의 상체\n\nLOCATION (lock): Inside the cramped bedroom of the suspect's family home, amid a bed, wardrobe, and boxes being searched. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height behind and slightly beside the 검찰 수사관, use a restrained handheld medium frame across his oblique shoulder line as his upper body blocks 유경자. The investigator braces across the center-right while 유경자 presses forward from the left background with one arm raised in accusation; partial search activity remains deeper in the narrow room without displacing their confrontation.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 검찰 수사관 in the middle-right of the frame, foreground; 유경자 in the middle-left of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 침대 (Being searched) — Only part of its side and top remain visible behind the figures; used as Establishes the cramped search area behind the confrontation; 옷장 (Being searched) — Its front faces into the room and is seen obliquely by the camera; used as Creates a vertical boundary in the deeper room and indicates another search area; 잡동사니 상자들 (Being searched) — Different sides are visible where the boxes have been opened for inspection; used as Provides irregular foreground and background layers around the searching figures; 증거물 상자 (Receiving letters, notebooks, and photographs) — Its open upper side is visible from the oblique camera position; used as Procedural evidence anchor kept in the deeper portion of the frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with moderate contrast and subdued neutral-to-cool tonality.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The searched room is left disturbed, while seized letters, notebooks, and photographs are collected in evidence boxes.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리); 검찰 수사관 (Korean 남성, 성인, 타원형 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 수사관의 조끼: \"검찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S58sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S58sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 유경자: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1007208>"
    },
    {
     "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:831079>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L55B02.png"
    },
    {
     "label": "CHARACTER REFERENCE — 유경자: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1007208>"
    },
    {
     "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:831079>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "수사관이 양팔을 벌려 온몸으로 막아서는 핵심 동작과 지정된 의상(가디건)을 정확히 연출하여 더 높은 우선순위 지시를 충족함 (시간대 묘사는 다소 미흡)."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "로케이션 배경은 완벽히 복제했으나, 수사관이 가만히 서 있어 '막아선' 동작 지시를 어겼고 지정된 의상도 누락됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "유경자의 시선과 삿대질이 수사관을 향함. 수사관은 유경자를 마주보고 양팔을 벌려 앞을 막고 있음.",
      "built_space": "방 안의 왼쪽 문과 옷장, 안쪽 벽에 배치된 침대가 보임. 수사관의 어깨 너머로 잡은 미디엄 샷 구도.",
      "entities": "유경자(얼굴 및 복장 일치), 수사관(남색 가디건, '검찰' 조끼 등 의상 일치). 증거물 상자와 어질러진 방의 모습이 잘 나타남.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥에 안정적으로 서 있으며, 뻗은 팔과 신체의 무게 중심이 자연스러움."
     },
     {
      "label": "B",
      "direction": "유경자의 삿대질은 수사관을 향하나, 수사관은 팔을 내린 채 서 있기만 함.",
      "built_space": "로케이션 레퍼런스의 방 구조와 완벽히 동일한 배경. 수사관 등 뒤에서의 미디엄 샷.",
      "entities": "유경자의 외형과 복장은 일치함. 수사관은 '검찰' 조끼를 입었으나 지정된 남색 가디건이 누락됨. 배경의 상자 및 소품은 레퍼런스와 일치.",
      "hard_violations": [],
      "physics": "두 사람 모두 바닥에 정상적으로 두 발을 딛고 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "수사관이 양팔을 벌려 온몸으로 막아서는 핵심 동작과 지정된 의상(가디건)을 정확히 연출하여 더 높은 우선순위 지시를 충족함 (시간대 묘사는 다소 미흡)."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "로케이션 배경은 완벽히 복제했으나, 수사관이 가만히 서 있어 '막아선' 동작 지시를 어겼고 지정된 의상도 누락됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "유경자의 시선과 삿대질이 수사관을 향함. 수사관은 유경자를 마주보고 양팔을 벌려 앞을 막고 있음.",
      "built_space": "방 안의 왼쪽 문과 옷장, 안쪽 벽에 배치된 침대가 보임. 수사관의 어깨 너머로 잡은 미디엄 샷 구도.",
      "entities": "유경자(얼굴 및 복장 일치), 수사관(남색 가디건, '검찰' 조끼 등 의상 일치). 증거물 상자와 어질러진 방의 모습이 잘 나타남.",
      "hard_violations": [],
      "physics": "두 인물 모두 바닥에 안정적으로 서 있으며, 뻗은 팔과 신체의 무게 중심이 자연스러움."
     },
     {
      "label": "B",
      "direction": "유경자의 삿대질은 수사관을 향하나, 수사관은 팔을 내린 채 서 있기만 함.",
      "built_space": "로케이션 레퍼런스의 방 구조와 완벽히 동일한 배경. 수사관 등 뒤에서의 미디엄 샷.",
      "entities": "유경자의 외형과 복장은 일치함. 수사관은 '검찰' 조끼를 입었으나 지정된 남색 가디건이 누락됨. 배경의 상자 및 소품은 레퍼런스와 일치.",
      "hard_violations": [],
      "physics": "두 사람 모두 바닥에 정상적으로 두 발을 딛고 서 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "지정된 장소 레퍼런스의 방 구조와 가구(투톤 옷장, 창문, 문손잡이 등)를 완벽하게 재현했으며, 카메라 구도와 '검찰' 텍스트 모두 지시사항을 정확히 따랐습니다."
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "인물의 자세와 텍스트는 지시를 따랐으나, 옷장, 창문, 침대 등 방의 구조와 가구가 위치 레퍼런스와 완전히 달라 장소 고정 조건을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.571
     },
     "adjusted": {
      "A": 1.571,
      "B": 1.571
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "지정된 장소 레퍼런스의 방 구조와 가구(투톤 옷장, 창문, 문손잡이 등)를 완벽하게 재현했으며, 카메라 구도와 '검찰' 텍스트 모두 지시사항을 정확히 따랐습니다."
     },
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "인물의 자세와 텍스트는 지시를 따랐으나, 옷장, 창문, 침대 등 방의 구조와 가구가 위치 레퍼런스와 완전히 달라 장소 고정 조건을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1578,
     "B": 1575
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "유경자의 시선과 삿대질이 수사관을 향함. 수사관은 유경자를 마주보고 양팔을 벌려 앞을 막고 있음.",
    "built_space": "방 안의 왼쪽 문과 옷장, 안쪽 벽에 배치된 침대가 보임. 수사관의 어깨 너머로 잡은 미디엄 샷 구도.",
    "entities": "유경자(얼굴 및 복장 일치), 수사관(남색 가디건, '검찰' 조끼 등 의상 일치). 증거물 상자와 어질러진 방의 모습이 잘 나타남.",
    "hard_violations": [],
    "physics": "두 인물 모두 바닥에 안정적으로 서 있으며, 뻗은 팔과 신체의 무게 중심이 자연스러움."
   },
   {
    "label": "B",
    "direction": "유경자의 삿대질은 수사관을 향하나, 수사관은 팔을 내린 채 서 있기만 함.",
    "built_space": "로케이션 레퍼런스의 방 구조와 완벽히 동일한 배경. 수사관 등 뒤에서의 미디엄 샷.",
    "entities": "유경자의 외형과 복장은 일치함. 수사관은 '검찰' 조끼를 입었으나 지정된 남색 가디건이 누락됨. 배경의 상자 및 소품은 레퍼런스와 일치.",
    "hard_violations": [],
    "physics": "두 사람 모두 바닥에 정상적으로 두 발을 딛고 서 있음."
   }
  ],
  "totals": {
   "A": 1578,
   "B": 1575
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "수사관이 양팔을 벌려 온몸으로 막아서는 핵심 동작과 지정된 의상(가디건)을 정확히 연출하여 더 높은 우선순위 지시를 충족함 (시간대 묘사는 다소 미흡)."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "로케이션 배경은 완벽히 복제했으나, 수사관이 가만히 서 있어 '막아선' 동작 지시를 어겼고 지정된 의상도 누락됨."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L55B02.png"
   },
   {
    "label": "CHARACTER REFERENCE — 유경자: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1007208>"
   },
   {
    "label": "CHARACTER REFERENCE — 검찰 수사관: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:831079>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 우측 수사관의 오른손이 해부학적으로 불가능한 구조입니다(오른팔임에도 손바닥이 카메라를 향하고 엄지가 위를 향해 왼손처럼 보임).",
     "fix_en": "Redraw the investigator's extended right hand to be anatomically correct for a right arm seen from behind: the back of the hand must face the camera, with the thumb pointing up. Preserve both figures' poses, clothing, the '검찰' text, and the surrounding room and boxes.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "배경 원본의 어두운 갈색 수납형 침대 헤드보드가 밝은 나무색의 전혀 다른 디자인으로 변경되었습니다.",
     "fix_en": "Restore the bed's headboard to the dark brown shelving design seen in the background reference.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "우측 창문의 격자 디자인과 유리 재질(전체 불투명)이 원본과 달라졌으며, 창문 우측에 원본에 없던 기둥/벽 구조물이 추가되었습니다.",
     "fix_en": "Restore the original sliding window design and remove the newly invented wall pillar on the right.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "벽걸이 에어컨의 모델 디자인이 원본과 다르게 변경되었고, 우측으로 늘어져 있던 전선이 사라졌습니다.",
     "fix_en": "Restore the original air conditioner design and its hanging wire.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "삿대질을 하는 유경자의 왼손 엄지손가락 부위가 부자연스럽게 뭉개져 있습니다.",
     "fix_en": "Refine the pointing left hand so the thumb resolves clearly.",
     "severity": "minor",
     "observation_index": 4
    },
    {
     "issue_ko": "배경 사진의 짙은 옷장·침구·유로핀스 상자 등이 바뀌거나 사라져 고정 배경이 그대로 유지되지 않았다.",
     "fix_en": "Restore the original dark wardrobe doors, bedding colors, and background boxes.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 우측 수사관의 오른손이 해부학적으로 불가능한 구조입니다(오른팔임에도 손바닥이 카메라를 향하고 엄지가 위를 향해 왼손처럼 보임).",
     "severity": "critical"
    },
    {
     "issue_ko": "배경 원본의 어두운 갈색 수납형 침대 헤드보드가 밝은 나무색의 전혀 다른 디자인으로 변경되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 창문의 격자 디자인과 유리 재질(전체 불투명)이 원본과 달라졌으며, 창문 우측에 원본에 없던 기둥/벽 구조물이 추가되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "벽걸이 에어컨의 모델 디자인이 원본과 다르게 변경되었고, 우측으로 늘어져 있던 전선이 사라졌습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "삿대질을 하는 유경자의 왼손 엄지손가락 부위가 부자연스럽게 뭉개져 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "샷 텍스트가 허용하지 않은 세 번째 인물이 화면 오른쪽 침대 위에 앉아 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경 사진의 짙은 옷장·침구·유로핀스 상자 등이 바뀌거나 사라져 고정 배경이 그대로 유지되지 않았다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 5,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "배경 원본의 어두운 갈색 수납형 침대 헤드보드가 밝은 나무색의 전혀 다른 디자인으로 변경되었습니다.",
    "fix_en": "Restore the bed's headboard to the dark brown shelving design seen in the background reference.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "우측 창문의 격자 디자인과 유리 재질(전체 불투명)이 원본과 달라졌으며, 창문 우측에 원본에 없던 기둥/벽 구조물이 추가되었습니다.",
    "fix_en": "Restore the original sliding window design and remove the newly invented wall pillar on the right.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "벽걸이 에어컨의 모델 디자인이 원본과 다르게 변경되었고, 우측으로 늘어져 있던 전선이 사라졌습니다.",
    "fix_en": "Restore the original air conditioner design and its hanging wire.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "삿대질을 하는 유경자의 왼손 엄지손가락 부위가 부자연스럽게 뭉개져 있습니다.",
    "fix_en": "Refine the pointing left hand so the thumb resolves clearly.",
    "severity": "minor",
    "observation_index": 4
   },
   {
    "issue_ko": "배경 사진의 짙은 옷장·침구·유로핀스 상자 등이 바뀌거나 사라져 고정 배경이 그대로 유지되지 않았다.",
    "fix_en": "Restore the original dark wardrobe doors, bedding colors, and background boxes.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the investigator's extended right hand to be anatomically correct for a right arm seen from behind: the back of the hand must face the camera, with the thumb pointing up. Preserve both figures' poses, clothing, the '검찰' text, and the surrounding room and boxes.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "스케치의 제3의 인물을 규칙에 따라 정확히 배제하고, 악을 쓰는 유경자의 표정과 '검찰' 조끼 등 텍스트 지시를 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 엄격히 금지한 제3의 인물을 프레임 안에 임의로 추가하여 치명적인 지시 위반을 범했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "유경자가 수사관의 얼굴을 향해 삿대질을 하고 있으며, 수사관은 몸을 돌려 유경자를 주시하며 막아서고 있습니다.",
      "built_space": "배경 레퍼런스와 동일한 구조의 좁은 방으로, 옷장, 침대, 그리고 바닥에 흩어진 잡동사니 상자들이 카메라 시점에 맞게 적절히 배치되어 있습니다.",
      "entities": "유경자는 캐릭터 레퍼런스와 완벽히 일치하는 복장과 외모를 갖추고 지시된 대로 소리치는 표정입니다. 수사관 역시 텍스트 지시대로 '검찰'이 적힌 수사관 조끼를 입고 있습니다. 명시되지 않은 인물은 없습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 방바닥에 단단히 발을 딛고 있으며, 수사관이 양팔을 뻗어 길을 막는 동작이 자연스럽게 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "유경자가 수사관을 향해 손가락질을 하고 있으며, 침대에 앉은 남자는 손에 든 서류를 내려다보고 있습니다.",
      "built_space": "방의 가구 배치와 공간감은 배경 이미지의 구도를 잘 따르고 있습니다.",
      "entities": "유경자의 표정이 소리치는 모습이 아닌 무표정에 가깝습니다. 수사관은 조끼를 입지 않고 가디건 등판에 직접 글씨가 인쇄되어 있습니다. 프롬프트에 없는 젊은 남성이 침대 위에 존재합니다.",
      "hard_violations": [
       "invented people (프롬프트에 명시되지 않은 제3의 인물이 침대에 앉아 있음)"
      ],
      "physics": "서 있는 두 사람과 침대에 앉은 사람 모두 중력에 맞게 자연스럽게 몸을 가누고 있으며 사물을 쥐는 손의 형태도 정상적입니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "스케치의 제3의 인물을 규칙에 따라 정확히 배제하고, 악을 쓰는 유경자의 표정과 '검찰' 조끼 등 텍스트 지시를 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트가 엄격히 금지한 제3의 인물을 프레임 안에 임의로 추가하여 치명적인 지시 위반을 범했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "유경자가 수사관의 얼굴을 향해 삿대질을 하고 있으며, 수사관은 몸을 돌려 유경자를 주시하며 막아서고 있습니다.",
      "built_space": "배경 레퍼런스와 동일한 구조의 좁은 방으로, 옷장, 침대, 그리고 바닥에 흩어진 잡동사니 상자들이 카메라 시점에 맞게 적절히 배치되어 있습니다.",
      "entities": "유경자는 캐릭터 레퍼런스와 완벽히 일치하는 복장과 외모를 갖추고 지시된 대로 소리치는 표정입니다. 수사관 역시 텍스트 지시대로 '검찰'이 적힌 수사관 조끼를 입고 있습니다. 명시되지 않은 인물은 없습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 방바닥에 단단히 발을 딛고 있으며, 수사관이 양팔을 뻗어 길을 막는 동작이 자연스럽게 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "유경자가 수사관을 향해 손가락질을 하고 있으며, 침대에 앉은 남자는 손에 든 서류를 내려다보고 있습니다.",
      "built_space": "방의 가구 배치와 공간감은 배경 이미지의 구도를 잘 따르고 있습니다.",
      "entities": "유경자의 표정이 소리치는 모습이 아닌 무표정에 가깝습니다. 수사관은 조끼를 입지 않고 가디건 등판에 직접 글씨가 인쇄되어 있습니다. 프롬프트에 없는 젊은 남성이 침대 위에 존재합니다.",
      "hard_violations": [
       "invented people (프롬프트에 명시되지 않은 제3의 인물이 침대에 앉아 있음)"
      ],
      "physics": "서 있는 두 사람과 침대에 앉은 사람 모두 중력에 맞게 자연스럽게 몸을 가누고 있으며 사물을 쥐는 손의 형태도 정상적입니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프롬프트에 명시된 두 인물만 정확히 배치하였으며, 유경자의 악을 쓰는 표정과 수사관의 조끼 텍스트를 완벽하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트에 없는 제3의 인물을 추가한 치명적 위반이 있으며, 유경자가 소리치는 표정이 아닙니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "유경자는 수사관을 향해 손가락질을 하고 있으며, 수사관은 팔을 벌려 막고 있음.",
      "built_space": "방 안의 옷장, 침대, 문 등 구조물의 배치와 시점이 레퍼런스와 일치함.",
      "entities": "유경자의 의상은 일치하나 표정이 평온함. 수사관은 조끼가 아닌 스웨터를 입고 있음. 텍스트에 없는 제3의 남성이 침대에 앉아 있음.",
      "hard_violations": [
       "발명된 인물: 샷 텍스트에 명시되지 않은 제3의 인물이 침대에 앉아 있음"
      ],
      "physics": "인물들의 발이 바닥에 닿아 있으며, 상자들은 바닥에 안정적으로 놓여 있음."
     },
     {
      "label": "B",
      "direction": "유경자는 수사관을 향해 손가락질을 하고, 수사관은 양팔을 벌려 유경자를 막아서고 있음.",
      "built_space": "배경의 옷장, 침대, 방의 구조가 레퍼런스와 정확히 일치함.",
      "entities": "유경자는 악을 쓰는 표정과 의상이 모두 일치함. 수사관은 지시된 대로 조끼를 입고 '검찰' 텍스트가 적혀 있음. 불필요한 인물이 없음.",
      "hard_violations": [],
      "physics": "모든 인물이 바닥에 서서 체중을 지탱하고 있으며, 바닥의 상자와 사진들도 자연스럽게 놓여 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트에 명시된 두 인물만 정확히 배치하였으며, 유경자의 악을 쓰는 표정과 수사관의 조끼 텍스트를 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프롬프트에 없는 제3의 인물을 추가한 치명적 위반이 있으며, 유경자가 소리치는 표정이 아닙니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "유경자는 수사관을 향해 손가락질을 하고 있으며, 수사관은 팔을 벌려 막고 있음.",
      "built_space": "방 안의 옷장, 침대, 문 등 구조물의 배치와 시점이 레퍼런스와 일치함.",
      "entities": "유경자의 의상은 일치하나 표정이 평온함. 수사관은 조끼가 아닌 스웨터를 입고 있음. 텍스트에 없는 제3의 남성이 침대에 앉아 있음.",
      "hard_violations": [
       "발명된 인물: 샷 텍스트에 명시되지 않은 제3의 인물이 침대에 앉아 있음"
      ],
      "physics": "인물들의 발이 바닥에 닿아 있으며, 상자들은 바닥에 안정적으로 놓여 있음."
     },
     {
      "label": "A",
      "direction": "유경자는 수사관을 향해 손가락질을 하고, 수사관은 양팔을 벌려 유경자를 막아서고 있음.",
      "built_space": "배경의 옷장, 침대, 방의 구조가 레퍼런스와 정확히 일치함.",
      "entities": "유경자는 악을 쓰는 표정과 의상이 모두 일치함. 수사관은 지시된 대로 조끼를 입고 '검찰' 텍스트가 적혀 있음. 불필요한 인물이 없음.",
      "hard_violations": [],
      "physics": "모든 인물이 바닥에 서서 체중을 지탱하고 있으며, 바닥의 상자와 사진들도 자연스럽게 놓여 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S58sh1__bgfirst_bg.png",
   "bg_asset_id": "b939e106-558d-407c-8d99-9fef88379129",
   "bg_record_key": "S58sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S58sh1::cine": {
  "applied": true,
  "fingerprint": "3192bccd2d6eff4a0eba290f440b183fdfcfda5e1e15ea604f9dcb7c683dfda9",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S58sh1_sel.png",
  "source_sha256": "c85379bd587845de7640d63aa16a4ff1dbaabe67425412b505551ed2180f1e39",
  "file": "S58sh1_cine.png",
  "latency_ms": 10505
 },
 "S58sh4::signage": {
  "fp": "4a44da8c508bfd47",
  "inscriptions": [
   {
    "surface_native": "죄수복 가슴 표찰",
    "text_native": "4021",
    "reason_ko": "2000년대 한국 교도소 독방에 수감된 주인공의 신분을 사실적으로 묘사하기 위해 죄수복 가슴에 부착된 수인번호 표찰이 필요합니다."
   }
  ]
 },
 "S58sh4": {
  "input_fingerprint": "ce5a727e20eca946",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 독방 안 수납장 앞에 서서 낡은 사진첩을 양손으로 펼쳐 든 장원섭의 상체\n\nLOCATION (lock): Inside the sparse prison solitary cell, directly before its storage cabinet during the search. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Inside the cell at chest height, dolly toward 장원섭 from a three-quarter side angle as he holds the open album at mid-torso in front of the storage unit. His upper body occupies one side of the medium frame, his lowered attention rests on the pages, and the doorway and corridor remain visible along the outward side of his eyeline for the coming exchange.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 수납장 (Holding books and the discovered album) — Its front and shelves are seen obliquely beside 장원섭; used as The immediate discovery point directly behind 장원섭; 낡은 사진첩 (Open and being examined) — Both open page faces are visible at an oblique angle, showing photographs of 지국현 in his twenties with family members; used as Carries the family evidence within 장원섭's hands while remaining proportionate to his torso; 자격증 서적들 (Slotted in the storage unit) — The spines facing outward identify 제과제빵, 자동차 정비, and 목공기능사 subjects; used as Frames the album's original location and gives procedural context to the search; 독방 출입구와 복도 (Visible from inside the cell) — The doorway is seen on the outward side of 장원섭, with the corridor extending beyond it; used as Preserves the visual route from the cell interior toward the watched search.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light appropriate to the cell, kept restrained and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Wonseop holds open the photograph album removed from among the certificate-study books in Ji Gukhyeon's cell cabinet.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 죄수복 가슴 표찰: \"4021\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 독방 안 수납장 앞에 서서 낡은 사진첩을 양손으로 펼쳐 든 장원섭의 상체\n\nLOCATION (lock): Inside the sparse prison solitary cell, directly before its storage cabinet during the search. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Inside the cell at chest height, dolly toward 장원섭 from a three-quarter side angle as he holds the open album at mid-torso in front of the storage unit. His upper body occupies one side of the medium frame, his lowered attention rests on the pages, and the doorway and corridor remain visible along the outward side of his eyeline for the coming exchange.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 수납장 (Holding books and the discovered album) — Its front and shelves are seen obliquely beside 장원섭; used as The immediate discovery point directly behind 장원섭; 낡은 사진첩 (Open and being examined) — Both open page faces are visible at an oblique angle, showing photographs of 지국현 in his twenties with family members; used as Carries the family evidence within 장원섭's hands while remaining proportionate to his torso; 자격증 서적들 (Slotted in the storage unit) — The spines facing outward identify 제과제빵, 자동차 정비, and 목공기능사 subjects; used as Frames the album's original location and gives procedural context to the search; 독방 출입구와 복도 (Visible from inside the cell) — The doorway is seen on the outward side of 장원섭, with the corridor extending beyond it; used as Preserves the visual route from the cell interior toward the watched search.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light appropriate to the cell, kept restrained and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Wonseop holds open the photograph album removed from among the certificate-study books in Ji Gukhyeon's cell cabinet.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 죄수복 가슴 표찰: \"4021\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 독방 안 수납장 앞에 서서 낡은 사진첩을 양손으로 펼쳐 든 장원섭의 상체\n\nLOCATION (lock): Inside the sparse prison solitary cell, directly before its storage cabinet during the search. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Inside the cell at chest height, dolly toward 장원섭 from a three-quarter side angle as he holds the open album at mid-torso in front of the storage unit. His upper body occupies one side of the medium frame, his lowered attention rests on the pages, and the doorway and corridor remain visible along the outward side of his eyeline for the coming exchange.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 수납장 (Holding books and the discovered album) — Its front and shelves are seen obliquely beside 장원섭; used as The immediate discovery point directly behind 장원섭; 낡은 사진첩 (Open and being examined) — Both open page faces are visible at an oblique angle, showing photographs of 지국현 in his twenties with family members; used as Carries the family evidence within 장원섭's hands while remaining proportionate to his torso; 자격증 서적들 (Slotted in the storage unit) — The spines facing outward identify 제과제빵, 자동차 정비, and 목공기능사 subjects; used as Frames the album's original location and gives procedural context to the search; 독방 출입구와 복도 (Visible from inside the cell) — The doorway is seen on the outward side of 장원섭, with the corridor extending beyond it; used as Preserves the visual route from the cell interior toward the watched search.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light appropriate to the cell, kept restrained and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Wonseop holds open the photograph album removed from among the certificate-study books in Ji Gukhyeon's cell cabinet.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 죄수복 가슴 표찰: \"4021\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "장원섭의 시선은 두 손에 든 사진첩의 페이지를 향해 아래로 고정되어 있다.",
    "built_space": "독방 내부의 철제 수납장이 인물 뒤에 있으며, 화면 우측으로 열린 문과 복도가 보인다.",
    "entities": "장원섭(레퍼런스와 일치), 4021 표찰의 죄수복, 수납장에 세로로 꽂힌 제과제빵, 자동차 정비, 목공기능사 책, 펼쳐진 가족 사진첩이 정확히 나타난다.",
    "hard_violations": [],
    "physics": "양손으로 사진첩의 양 하단부를 안정적으로 쥐고 들고 있으며, 자세와 지지 구조가 자연스럽다."
   },
   {
    "label": "B",
    "direction": "장원섭의 시선이 사진첩을 향하고 있으며, 시선 앞쪽(화면 좌측)으로 출입구가 위치한다.",
    "built_space": "수납장이 인물 뒤편에 있고, 왼쪽으로 열린 출입구와 복도가 이어진다.",
    "entities": "장원섭, 4021 표찰, 가로로 쌓인 책(자동차 정비가 중복되고 알 수 없는 텍스트 포함), 펼쳐진 사진첩이 존재한다.",
    "hard_violations": [
     "사진첩 우측 페이지를 잡고 있는 손이 물리적으로 불가능한 해부학적 구조(추가되거나 융합된 손가락)를 보임"
    ],
    "physics": "우측 하단을 받치는 오른손과 별개로 우측 페이지 중앙을 쥐고 있는 비정상적인 손가락들이 있어 물리적 형태가 성립하지 않는다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "요구된 인물의 외모, 서적의 정확한 한글 텍스트, 수납장과 사진첩 등 주요 소품들을 지시문대로 훌륭하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "출입구 배치는 지시문에 부합하나, 책의 텍스트가 중복 및 왜곡되었고 사진첩을 잡고 있는 손의 해부학적 오류가 치명적입니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S38sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "지정된 40대 캐릭터 레퍼런스 대신, 사용이 금지된 이전 컷 레퍼런스의 젊은 남성 얼굴이 그대로 렌더링됨.",
     "fix_en": "Replace the young man's head and face with the face and neat short black hair of the 40-year-old man from the character reference, maintaining the exact same head angle, downward gaze, and lighting. Preserve the character's body, pose, prison uniform, hands, the held open album, the storage cabinet with books, the cell walls, and the background doorway and corridor.",
     "severity": "critical",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "지정된 40대 캐릭터 레퍼런스 대신, 사용이 금지된 이전 컷 레퍼런스의 젊은 남성 얼굴이 그대로 렌더링됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 컷에서 설정된 투톤 벽면(상단 베이지, 하단 청록색)이 단색 벽으로 처리되어 배경 연속성이 어긋남.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 상체 인물이 캐릭터 레퍼런스의 장원섭(40대 초반)이 아니라 이전 샷의 20대 얼굴이다",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the young man's head and face with the face and neat short black hair of the 40-year-old man from the character reference, maintaining the exact same head angle, downward gaze, and lighting. Preserve the character's body, pose, prison uniform, hands, the held open album, the storage cabinet with books, the cell walls, and the background doorway and corridor.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 2000,
      "verdict_ko": "요구된 모든 요소와 텍스트를 정확히 구현했으며, 사진첩 내부 가족사진의 일그러진 얼굴들을 손상 없이 성공적으로 복구하여 완성도가 더 높습니다."
     },
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "지시사항을 잘 따랐으나, 소품인 사진첩 내부의 인물 얼굴들이 심하게 일그러져 있어 디테일에서 아쉬움이 남습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.857,
      "B": 2.0
     },
     "adjusted": {
      "A": 1.857,
      "B": 2.0
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 2000,
      "verdict_ko": "요구된 모든 요소와 텍스트를 정확히 구현했으며, 사진첩 내부 가족사진의 일그러진 얼굴들을 손상 없이 성공적으로 복구하여 완성도가 더 높습니다."
     },
     {
      "label": "A",
      "score": 1857,
      "verdict_ko": "지시사항을 잘 따랐으나, 소품인 사진첩 내부의 인물 얼굴들이 심하게 일그러져 있어 디테일에서 아쉬움이 남습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트의 모든 세부 요소와 텍스트를 정확히 반영했으며, 양손으로 사진첩을 드는 물리적 지지 상태가 자연스러움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "오른손이 삭제되어 사진첩의 오른쪽이 허공에 떠 있는 심각한 물리적 오류와 지시사항 위반이 발생함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선이 펼쳐진 낡은 사진첩을 정확히 향함.",
      "built_space": "독방 안 수납장 앞이며, 우측으로 출입구와 복도가 보임.",
      "entities": "장원섭의 외모, 수납장, 자격증 서적(제과제빵 등), 사진첩, 가슴 표찰(4021) 모두 지시와 일치함.",
      "hard_violations": [],
      "physics": "왼손이 사진첩 왼쪽을 잡고, 오른손 손가락이 오른쪽 아래를 안정적으로 지지함."
     },
     {
      "label": "B",
      "direction": "시선이 펼쳐진 낡은 사진첩을 향함.",
      "built_space": "독방 안 수납장 앞이며, 우측으로 출입구와 복도가 보임.",
      "entities": "장원섭의 외모, 수납장, 자격증 서적, 사진첩, 가슴 표찰 모두 일치함.",
      "hard_violations": [
       "사진첩의 오른쪽을 지지하는 손이 없어 허공에 떠 있음"
      ],
      "physics": "왼손만 사진첩을 잡고 있으며, 오른쪽은 지지대 없이 허공에 떠 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "프롬프트의 모든 세부 요소와 텍스트를 정확히 반영했으며, 양손으로 사진첩을 드는 물리적 지지 상태가 자연스러움."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "오른손이 삭제되어 사진첩의 오른쪽이 허공에 떠 있는 심각한 물리적 오류와 지시사항 위반이 발생함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선이 펼쳐진 낡은 사진첩을 정확히 향함.",
      "built_space": "독방 안 수납장 앞이며, 우측으로 출입구와 복도가 보임.",
      "entities": "장원섭의 외모, 수납장, 자격증 서적(제과제빵 등), 사진첩, 가슴 표찰(4021) 모두 지시와 일치함.",
      "hard_violations": [],
      "physics": "왼손이 사진첩 왼쪽을 잡고, 오른손 손가락이 오른쪽 아래를 안정적으로 지지함."
     },
     {
      "label": "A",
      "direction": "시선이 펼쳐진 낡은 사진첩을 향함.",
      "built_space": "독방 안 수납장 앞이며, 우측으로 출입구와 복도가 보임.",
      "entities": "장원섭의 외모, 수납장, 자격증 서적, 사진첩, 가슴 표찰 모두 일치함.",
      "hard_violations": [
       "사진첩의 오른쪽을 지지하는 손이 없어 허공에 떠 있음"
      ],
      "physics": "왼손만 사진첩을 잡고 있으며, 오른쪽은 지지대 없이 허공에 떠 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1860,
     "B": 2008
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S38sh4"
  }
 },
 "S58sh4::cine": {
  "applied": true,
  "fingerprint": "171b26a0dbc564d2e76736d6d7f2b2bf2e73ac0b539c2b0c16f3c6648d7c8849",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S58sh4_sel.png",
  "source_sha256": "9844c357755c03e305a94dfa9d034078373e721d706659e1888ed7f04e739976",
  "file": "S58sh4_cine.png",
  "latency_ms": 11203
 },
 "S58sh8::signage": {
  "fp": "c710c5bfa5113a7b",
  "inscriptions": [
   {
    "surface_native": "화이트보드 상단",
    "text_native": "수사 상황판",
    "reason_ko": "검사실 내부의 사건 조사용 화이트보드임을 명확히 나타내기 위해 상단에 적절한 제목 표기가 필요합니다."
   },
   {
    "surface_native": "화이트보드 좌측 영역",
    "text_native": "피해자 시간대별 행적",
    "reason_ko": "피해자의 동선 기록 자료들이 부착된 화이트보드 구역을 나타내는 표기로 극의 사실성을 높입니다."
   }
  ]
 },
 "S58sh8": {
  "input_fingerprint": "0ae5f441e2d3e55a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 피해자의 동선 기록과 부검 사진들이 화이트보드 위에 빼곡하게 붙어 있는 사무실 전경\n\nLOCATION (lock): Inside the prosecutor's office at night, facing the evidence-covered whiteboard and piles of seized items. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near the rear of the office above standing eye level, complete the crane rise in a wide, slightly downward view that places the evidence board prominently across the upper central background and the work area below it. The movement records, statements, and forensic photographs remain readable as a dense procedural arrangement but occupy only part of the full office composition, with seized materials providing scale in the lower frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: evidence board in the upper-center of the frame, background; office work area with seized items in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 화이트보드 (Covered with investigation records) — Its front face is visible to camera and densely carries the victim's three-day movement record, witness statements, and photographs of defensive wounds; used as Primary evidentiary focus within the broader office reveal; 압수품 더미와 작업 공간 (Filled with seized belongings under review) — Boxes and reviewed items are seen from above across the office work area; used as Forms the lower procedural layer and demonstrates the volume of reviewed evidence.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime office illumination is restrained in a subdued neutral-to-cool palette, with enough tonal range to preserve the board's procedural detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon's seized belongings remain piled in the office, and the whiteboard stays densely covered with the victim's timeline, witness statements, and defense-wound photographs.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 화이트보드 상단: \"수사 상황판\"\n- 화이트보드 좌측 영역: \"피해자 시간대별 행적\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 피해자의 동선 기록과 부검 사진들이 화이트보드 위에 빼곡하게 붙어 있는 사무실 전경\n\nLOCATION (lock): Inside the prosecutor's office at night, facing the evidence-covered whiteboard and piles of seized items. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near the rear of the office above standing eye level, complete the crane rise in a wide, slightly downward view that places the evidence board prominently across the upper central background and the work area below it. The movement records, statements, and forensic photographs remain readable as a dense procedural arrangement but occupy only part of the full office composition, with seized materials providing scale in the lower frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: evidence board in the upper-center of the frame, background; office work area with seized items in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 화이트보드 (Covered with investigation records) — Its front face is visible to camera and densely carries the victim's three-day movement record, witness statements, and photographs of defensive wounds; used as Primary evidentiary focus within the broader office reveal; 압수품 더미와 작업 공간 (Filled with seized belongings under review) — Boxes and reviewed items are seen from above across the office work area; used as Forms the lower procedural layer and demonstrates the volume of reviewed evidence.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime office illumination is restrained in a subdued neutral-to-cool palette, with enough tonal range to preserve the board's procedural detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon's seized belongings remain piled in the office, and the whiteboard stays densely covered with the victim's timeline, witness statements, and defense-wound photographs.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 화이트보드 상단: \"수사 상황판\"\n- 화이트보드 좌측 영역: \"피해자 시간대별 행적\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day into night.\n\nSHOT TEXT (authoritative, Korean): 피해자의 동선 기록과 부검 사진들이 화이트보드 위에 빼곡하게 붙어 있는 사무실 전경\n\nLOCATION (lock): Inside the prosecutor's office at night, facing the evidence-covered whiteboard and piles of seized items. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near the rear of the office above standing eye level, complete the crane rise in a wide, slightly downward view that places the evidence board prominently across the upper central background and the work area below it. The movement records, statements, and forensic photographs remain readable as a dense procedural arrangement but occupy only part of the full office composition, with seized materials providing scale in the lower frame.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: evidence board in the upper-center of the frame, background; office work area with seized items in the lower-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 화이트보드 (Covered with investigation records) — Its front face is visible to camera and densely carries the victim's three-day movement record, witness statements, and photographs of defensive wounds; used as Primary evidentiary focus within the broader office reveal; 압수품 더미와 작업 공간 (Filled with seized belongings under review) — Boxes and reviewed items are seen from above across the office work area; used as Forms the lower procedural layer and demonstrates the volume of reviewed evidence.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Nighttime office illumination is restrained in a subdued neutral-to-cool palette, with enough tonal range to preserve the board's procedural detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon's seized belongings remain piled in the office, and the whiteboard stays densely covered with the victim's timeline, witness statements, and defense-wound photographs.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 화이트보드 상단: \"수사 상황판\"\n- 화이트보드 좌측 영역: \"피해자 시간대별 행적\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.375,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.5,
    "B": 1.625
   },
   "adjusted": {
    "A": 1.25,
    "B": 1.625
   },
   "violations": {
    "A": [
     "[gemini-pro] 레퍼런스에서 설정된 큰 임원용 책상이 명패와 봉투 크기에 비해 비현실적으로 작게 왜곡되어 사물 간의 비례가 완전히 어긋남."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1625,
   "A": 1250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "와이드 샷의 구도와 조명 분위기를 잘 살렸으며, 지정된 텍스트와 레퍼런스의 소품들을 자연스러운 비율로 정확하게 배치했습니다."
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "지정된 텍스트는 정확하게 구현되었으나, 레퍼런스의 책상 크기를 심하게 왜곡하여 물리적 비례가 어긋납니다.  ★위반: [gemini-pro] 레퍼런스에서 설정된 큰 임원용 책상이 명패와 봉투 크기에 비해 비현실적으로 작게 왜곡되어 사물 간의 비례가 완전히 어긋남."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S54sh18_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화이트보드 상단 글자가 '수사 상 황판'으로 띄어쓰기가 깨져 있다",
     "fix_en": "Remove the space between the second and third characters on the top whiteboard header so it correctly reads '수사 상황판'. Maintain the whiteboard's layout, all surrounding documents, desk clutter, boxes, lighting, and framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "보드와 책상 위 사진들이 부검·방어흔이 아니라 혈흔 비산 이미지로 보인다",
     "fix_en": "Replace the identical blood splatter photos on the board and desk with distinct forensic photographs showing varied defensive wounds on human skin. Maintain the whiteboard text, the arrangement of other documents, the desk, boxes, nameplates, lighting, and framing.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화이트보드와 책상 위에 놓인 수많은 법의학 사진들이 각기 다른 부검 및 방어흔 사진으로 묘사되지 않고, 완전히 동일한 추상적인 핏자국 이미지가 반복적으로 복사되어 배치되어 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "화이트보드 상단 글자가 '수사 상 황판'으로 띄어쓰기가 깨져 있다",
     "severity": "major"
    },
    {
     "issue_ko": "보드와 책상 위 사진들이 부검·방어흔이 아니라 혈흔 비산 이미지로 보인다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "화이트보드 상단 글자가 '수사 상 황판'으로 띄어쓰기가 깨져 있다",
    "fix_en": "Remove the space between the second and third characters on the top whiteboard header so it correctly reads '수사 상황판'. Maintain the whiteboard's layout, all surrounding documents, desk clutter, boxes, lighting, and framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "보드와 책상 위 사진들이 부검·방어흔이 아니라 혈흔 비산 이미지로 보인다",
    "fix_en": "Replace the identical blood splatter photos on the board and desk with distinct forensic photographs showing varied defensive wounds on human skin. Maintain the whiteboard text, the arrangement of other documents, the desk, boxes, nameplates, lighting, and framing.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S54sh18"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S58sh8::cine": {
  "applied": true,
  "fingerprint": "8dffb25790a35ce1c85dcd723ccff7a0b9cf44d91a5225240796a964b68a3872",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S58sh8_sel.png",
  "source_sha256": "98e9fb0c99920c6edf4790ea50fb4c9d76bc417f2912c70c39bcf7d7041329d1",
  "file": "S58sh8_cine.png",
  "latency_ms": 12342
 },
 "S59sh1::signage": {
  "fp": "0c4539bcc96f54c8",
  "inscriptions": [
   {
    "surface_native": "책상 위 삼각 명패",
    "text_native": "차장검사",
    "reason_ko": "검찰청 차장검사실 내부의 책상 앞이라는 공간적 배경과 권위를 직관적으로 보여주기 위해 책상 위 명패에 직책명을 표시합니다."
   }
  ]
 },
 "S59sh1": {
  "input_fingerprint": "99e70ee4bc08374b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차장검사 책상 앞에 서서 눈을 크게 뜨고 목에 핏대를 세운 채 입을 벌린 장원섭의 상체\n\nLOCATION (lock): Directly before the desk inside the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From standing torso height near a front corner of the desk, hold a tight upper-body view of 장원섭 at an off-center three-quarter angle as he argues toward the seated 차장검사 beyond his eyeline. 장원섭 leans over the desk edge with widened eyes, a taut neck, and mouth open mid-appeal, while the 차장검사 remains farther back signing documents without matching his urgency.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, looks toward seated deputy chief prosecutor; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 차장검사 책상 (Positioned between them) — A front corner and the top plane recede from 장원섭 toward the seated 차장검사; used as Separates 장원섭's urgent appeal from the 차장검사's controlled response and anchors their hierarchy; 결재서류 (Being signed) — The written face lies upward toward the 차장검사 and is seen obliquely by the camera; used as Supports the 차장검사's detached administrative action.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office illumination with moderate contrast and restrained neutral-to-cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 삼각 명패: \"차장검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차장검사 책상 앞에 서서 눈을 크게 뜨고 목에 핏대를 세운 채 입을 벌린 장원섭의 상체\n\nLOCATION (lock): Directly before the desk inside the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From standing torso height near a front corner of the desk, hold a tight upper-body view of 장원섭 at an off-center three-quarter angle as he argues toward the seated 차장검사 beyond his eyeline. 장원섭 leans over the desk edge with widened eyes, a taut neck, and mouth open mid-appeal, while the 차장검사 remains farther back signing documents without matching his urgency.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, looks toward seated deputy chief prosecutor; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 차장검사 책상 (Positioned between them) — A front corner and the top plane recede from 장원섭 toward the seated 차장검사; used as Separates 장원섭's urgent appeal from the 차장검사's controlled response and anchors their hierarchy; 결재서류 (Being signed) — The written face lies upward toward the 차장검사 and is seen obliquely by the camera; used as Supports the 차장검사's detached administrative action.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office illumination with moderate contrast and restrained neutral-to-cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 삼각 명패: \"차장검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 차장검사 책상 앞에 서서 눈을 크게 뜨고 목에 핏대를 세운 채 입을 벌린 장원섭의 상체\n\nLOCATION (lock): Directly before the desk inside the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From standing torso height near a front corner of the desk, hold a tight upper-body view of 장원섭 at an off-center three-quarter angle as he argues toward the seated 차장검사 beyond his eyeline. 장원섭 leans over the desk edge with widened eyes, a taut neck, and mouth open mid-appeal, while the 차장검사 remains farther back signing documents without matching his urgency.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, looks toward seated deputy chief prosecutor; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 차장검사 책상 (Positioned between them) — A front corner and the top plane recede from 장원섭 toward the seated 차장검사; used as Separates 장원섭's urgent appeal from the 차장검사's controlled response and anchors their hierarchy; 결재서류 (Being signed) — The written face lies upward toward the 차장검사 and is seen obliquely by the camera; used as Supports the 차장검사's detached administrative action.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office illumination with moderate contrast and restrained neutral-to-cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 삼각 명패: \"차장검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "장원섭은 배경의 차장검사를 향해 시선을 두며 소리치고, 차장검사는 책상 위 서류를 내려다봄.",
    "built_space": "두 인물 사이에 책상이 있으며, 명패와 결재서류가 책상 위에 올바르게 배치됨.",
    "entities": "장원섭이 레퍼런스의 회색 정장과 푸른 셔츠를 입고 지시된 표정을 지음. 명패의 '차장검사' 텍스트가 정확함.",
    "hard_violations": [],
    "physics": "몸을 앞쪽으로 기울인 장원섭의 자세와 펜을 쥔 차장검사의 손 모두 지탱점이 확실하고 자연스러움."
   },
   {
    "label": "B",
    "direction": "전경 인물은 차장검사를 향해 소리치고, 차장검사는 서류를 내려다봄.",
    "built_space": "책상을 사이에 둔 두 인물의 공간 배치가 지시와 일치함.",
    "entities": "전경 인물이 프롬프트가 금지한 이전 컷 인물의 얼굴과 의상을 착용함. 명패 텍스트는 정확함.",
    "hard_violations": [
     "이전 컷에서 배제하도록 지시된 인물의 얼굴과 의상(네이비 핀스트라이프 정장)을 전경의 장원섭에게 그대로 적용함"
    ],
    "physics": "몸을 기울인 자세와 펜을 쥐고 있는 손 모두 물리적으로 정상적인 지탱을 보임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 8,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 8,
    "verdict_ko": "지시된 장원섭의 의상(회색 정장)을 올바르게 반영하고, 긴장감 있는 상체 동작과 구도를 정확히 구현함."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "이전 컷에서 배제해야 할 인물의 얼굴과 핀스트라이프 정장을 장원섭에게 잘못 적용하여 지시를 엄중히 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S32sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "왼쪽 배경의 파란색 바인더 개수가 이전 샷(2개)과 달리 3개로 바뀜.",
     "fix_en": "Reduce the number of blue binders on the left background shelf from three to two.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "차장검사가 오른손에 쥔 펜의 구조가 왜곡되어 클립 부분이 손가락을 관통한 것처럼 보임.",
     "fix_en": "Redraw the pen in the seated man's right hand to remove the distortion, placing the clip on the barrel instead of intersecting the finger.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "이전 샷 레퍼런스의 TV 화면 내부에 있던 방송 자막 텍스트가 누락됨.",
     "fix_en": "Restore the broadcast subtitle text inside the TV screen.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "전경 왼쪽 장원섭의 얼굴·머리·체격이 지정 캐릭터 레퍼런스가 아니라 이전 샷 속 차장검사의 얼굴이다.",
     "fix_en": "Replace the face and hair of the standing man in the left foreground with the exact face and neatly parted short black hair of the Jang Won-seop character reference. Retain his angry expression, wide eyes, open mouth, grey suit, and leaning pose. Preserve the seated man, the desk, the TV screen, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "왼쪽 배경의 파란색 바인더 개수가 이전 샷(2개)과 달리 3개로 바뀜.",
     "severity": "minor"
    },
    {
     "issue_ko": "차장검사가 오른손에 쥔 펜의 구조가 왜곡되어 클립 부분이 손가락을 관통한 것처럼 보임.",
     "severity": "minor"
    },
    {
     "issue_ko": "이전 샷 레퍼런스의 TV 화면 내부에 있던 방송 자막 텍스트가 누락됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "전경 왼쪽 장원섭의 얼굴·머리·체격이 지정 캐릭터 레퍼런스가 아니라 이전 샷 속 차장검사의 얼굴이다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 배경의 파란색 바인더 개수가 이전 샷(2개)과 달리 3개로 바뀜.",
    "fix_en": "Reduce the number of blue binders on the left background shelf from three to two.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "차장검사가 오른손에 쥔 펜의 구조가 왜곡되어 클립 부분이 손가락을 관통한 것처럼 보임.",
    "fix_en": "Redraw the pen in the seated man's right hand to remove the distortion, placing the clip on the barrel instead of intersecting the finger.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "이전 샷 레퍼런스의 TV 화면 내부에 있던 방송 자막 텍스트가 누락됨.",
    "fix_en": "Restore the broadcast subtitle text inside the TV screen.",
    "severity": "minor",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the face and hair of the standing man in the left foreground with the exact face and neatly parted short black hair of the Jang Won-seop character reference. Retain his angry expression, wide eyes, open mouth, grey suit, and leaning pose. Preserve the seated man, the desk, the TV screen, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "장원섭의 외형이 캐릭터 레퍼런스와 완벽히 일치하며, 이전 샷 인물의 외형을 배제하라는 지시와 구도, 명패 텍스트 모두 훌륭하게 구현함 (단, TV 화면 자막 누락)."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "A와 거의 동일한 수준으로 프롬프트의 구도, 인물 외형, 표정 및 네거티브 제약을 매우 잘 따랐으나 동일하게 TV 화면의 자막이 누락됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭은 우측에 앉은 차장검사를 향해 시선을 고정하고 항의하고 있으며, 차장검사는 책상 위 서류를 내려다보고 있음.",
      "built_space": "차장검사실 내부. 두 사람 사이에 책상이 위치하며 명패와 결재서류가 놓여 있음. 배경에 TV, 벽면 액자(正道), 바인더가 올바른 위치에 있음.",
      "entities": "장원섭은 캐릭터 레퍼런스와 일치함(회색 정장, 넥타이, 얼굴). 책상 위 명패에 '차장검사'가 정확히 적혀 있음. 이전 샷의 인물 외형 배제 지시를 준수하여 차장검사는 이전 샷과 다른 외형임. 단, TV 화면의 자막 누락됨.",
      "hard_violations": [],
      "physics": "장원섭은 책상 쪽으로 몸을 기울여 지탱하고 있으며, 차장검사는 의자에 앉아 오른손으로 펜을 쥐고 있음. 물리적 오류 없음."
     },
     {
      "label": "B",
      "direction": "장원섭은 우측에 앉은 차장검사를 향해 시선을 고정하고 항의하고 있으며, 차장검사는 책상 위 서류를 내려다보고 있음.",
      "built_space": "차장검사실 내부. 두 사람 사이에 책상이 위치하며 명패와 결재서류가 놓여 있음. 배경에 TV, 벽면 액자(正道), 바인더가 올바른 위치에 있음.",
      "entities": "장원섭은 캐릭터 레퍼런스와 일치함(회색 정장, 넥타이, 얼굴). 책상 위 명패에 '차장검사'가 정확히 적혀 있음. 이전 샷의 인물 외형 배제 지시를 준수하여 차장검사는 이전 샷과 다른 외형임. 단, TV 화면의 자막 누락됨.",
      "hard_violations": [],
      "physics": "장원섭은 책상 쪽으로 몸을 기울여 지탱하고 있으며, 차장검사는 의자에 앉아 오른손으로 펜을 쥐고 있음. 물리적 오류 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "장원섭의 외형이 캐릭터 레퍼런스와 완벽히 일치하며, 이전 샷 인물의 외형을 배제하라는 지시와 구도, 명패 텍스트 모두 훌륭하게 구현함 (단, TV 화면 자막 누락)."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "A와 거의 동일한 수준으로 프롬프트의 구도, 인물 외형, 표정 및 네거티브 제약을 매우 잘 따랐으나 동일하게 TV 화면의 자막이 누락됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭은 우측에 앉은 차장검사를 향해 시선을 고정하고 항의하고 있으며, 차장검사는 책상 위 서류를 내려다보고 있음.",
      "built_space": "차장검사실 내부. 두 사람 사이에 책상이 위치하며 명패와 결재서류가 놓여 있음. 배경에 TV, 벽면 액자(正道), 바인더가 올바른 위치에 있음.",
      "entities": "장원섭은 캐릭터 레퍼런스와 일치함(회색 정장, 넥타이, 얼굴). 책상 위 명패에 '차장검사'가 정확히 적혀 있음. 이전 샷의 인물 외형 배제 지시를 준수하여 차장검사는 이전 샷과 다른 외형임. 단, TV 화면의 자막 누락됨.",
      "hard_violations": [],
      "physics": "장원섭은 책상 쪽으로 몸을 기울여 지탱하고 있으며, 차장검사는 의자에 앉아 오른손으로 펜을 쥐고 있음. 물리적 오류 없음."
     },
     {
      "label": "B",
      "direction": "장원섭은 우측에 앉은 차장검사를 향해 시선을 고정하고 항의하고 있으며, 차장검사는 책상 위 서류를 내려다보고 있음.",
      "built_space": "차장검사실 내부. 두 사람 사이에 책상이 위치하며 명패와 결재서류가 놓여 있음. 배경에 TV, 벽면 액자(正道), 바인더가 올바른 위치에 있음.",
      "entities": "장원섭은 캐릭터 레퍼런스와 일치함(회색 정장, 넥타이, 얼굴). 책상 위 명패에 '차장검사'가 정확히 적혀 있음. 이전 샷의 인물 외형 배제 지시를 준수하여 차장검사는 이전 샷과 다른 외형임. 단, TV 화면의 자막 누락됨.",
      "hard_violations": [],
      "physics": "장원섭은 책상 쪽으로 몸을 기울여 지탱하고 있으며, 차장검사는 의자에 앉아 오른손으로 펜을 쥐고 있음. 물리적 오류 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "명패 텍스트를 프롬프트에 제시된 '차장검사' 그대로 띄어쓰기 없이 정확하게 구현하여 가장 우수함.  ★위반: [openrouter:x-ai/grok-4.6] 장원섭이 지정 인물이 아님(이전 스틸 차장검사 얼굴을 이식) / [openrouter:x-ai/grok-4.6] 이전 샷의 다른 인물 얼굴을 이 샷 인물에게 옮김"
     },
     {
      "label": "A",
      "score": 1375,
      "verdict_ko": "대부분의 프롬프트 요구사항을 잘 충족했으나, 명패 텍스트에 요구하지 않은 띄어쓰기('차장 검사')가 포함되어 아쉬움.  ★위반: [openrouter:x-ai/grok-4.6] 장원섭이 지정 인물이 아님(이전 스틸 차장검사 얼굴을 이식) / [openrouter:x-ai/grok-4.6] 이전 샷의 다른 인물 얼굴을 이 샷 인물에게 옮김"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.875,
      "B": 2.0
     },
     "adjusted": {
      "A": 1.375,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 장원섭이 지정 인물이 아님(이전 스틸 차장검사 얼굴을 이식)",
       "[openrouter:x-ai/grok-4.6] 이전 샷의 다른 인물 얼굴을 이 샷 인물에게 옮김"
      ],
      "B": [
       "[openrouter:x-ai/grok-4.6] 장원섭이 지정 인물이 아님(이전 스틸 차장검사 얼굴을 이식)",
       "[openrouter:x-ai/grok-4.6] 이전 샷의 다른 인물 얼굴을 이 샷 인물에게 옮김"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "명패 텍스트를 프롬프트에 제시된 '차장검사' 그대로 띄어쓰기 없이 정확하게 구현하여 가장 우수함.  ★위반: [openrouter:x-ai/grok-4.6] 장원섭이 지정 인물이 아님(이전 스틸 차장검사 얼굴을 이식) / [openrouter:x-ai/grok-4.6] 이전 샷의 다른 인물 얼굴을 이 샷 인물에게 옮김"
     },
     {
      "label": "B",
      "score": 1375,
      "verdict_ko": "대부분의 프롬프트 요구사항을 잘 충족했으나, 명패 텍스트에 요구하지 않은 띄어쓰기('차장 검사')가 포함되어 아쉬움.  ★위반: [openrouter:x-ai/grok-4.6] 장원섭이 지정 인물이 아님(이전 스틸 차장검사 얼굴을 이식) / [openrouter:x-ai/grok-4.6] 이전 샷의 다른 인물 얼굴을 이 샷 인물에게 옮김"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1509,
     "B": 1384
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S32sh2"
  }
 },
 "S59sh1::cine": {
  "applied": true,
  "fingerprint": "fe2b0e2272c1e6ee207c043aa33034631959f40572ba1bea3396e437480e3930",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S59sh1_sel.png",
  "source_sha256": "1bb245af59ed549a74fc329b23f7d65f47c0290e165b652a4cee3285cc559791",
  "file": "S59sh1_cine.png",
  "latency_ms": 11373
 },
 "S59sh4::signage": {
  "fp": "c48af0c180337e0d",
  "inscriptions": [
   {
    "surface_native": "책상 위 명패",
    "text_native": "차장검사",
    "reason_ko": "검사실 내부의 현장감을 살리고 인물의 직책과 권위를 시각적으로 나타내기 위해 책상 위 명패가 필요합니다."
   }
  ]
 },
 "S59sh4": {
  "input_fingerprint": "b8281426ad4c487b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 치켜든 채 장원섭을 향해 매서운 눈빛으로 입을 벌린 차장검사의 얼굴\n\nLOCATION (lock): Behind the desk inside the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the dolly-in close behind and just outside 장원섭's shoulder line, slightly below the seated 차장검사's eye level and tilted upward. The 차장검사's lifted chin and severe gaze hold the facial close-up near center, while a minimal soft edge of 장원섭's shoulder identifies the person receiving the rebuke without obscuring the face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 책상 가장자리 (Between the two men) — Its upper edge appears below the seated 차장검사; used as A narrow lower-frame indication of the official setting rather than a competing focal point.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office illumination remains controlled and neutral-to-cool, preserving detail in the 차장검사's severe expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same executive office, desk, paperwork, daylight, and both men's formal clothing. Exclude the prosecutor's standing outburst from the close frame and isolate the chief prosecutor's stern response.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"차장검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 치켜든 채 장원섭을 향해 매서운 눈빛으로 입을 벌린 차장검사의 얼굴\n\nLOCATION (lock): Behind the desk inside the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the dolly-in close behind and just outside 장원섭's shoulder line, slightly below the seated 차장검사's eye level and tilted upward. The 차장검사's lifted chin and severe gaze hold the facial close-up near center, while a minimal soft edge of 장원섭's shoulder identifies the person receiving the rebuke without obscuring the face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 책상 가장자리 (Between the two men) — Its upper edge appears below the seated 차장검사; used as A narrow lower-frame indication of the official setting rather than a competing focal point.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office illumination remains controlled and neutral-to-cool, preserving detail in the 차장검사's severe expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same executive office, desk, paperwork, daylight, and both men's formal clothing. Exclude the prosecutor's standing outburst from the close frame and isolate the chief prosecutor's stern response.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"차장검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 치켜든 채 장원섭을 향해 매서운 눈빛으로 입을 벌린 차장검사의 얼굴\n\nLOCATION (lock): Behind the desk inside the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: End the dolly-in close behind and just outside 장원섭's shoulder line, slightly below the seated 차장검사's eye level and tilted upward. The 차장검사's lifted chin and severe gaze hold the facial close-up near center, while a minimal soft edge of 장원섭's shoulder identifies the person receiving the rebuke without obscuring the face.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 책상 가장자리 (Between the two men) — Its upper edge appears below the seated 차장검사; used as A narrow lower-frame indication of the official setting rather than a competing focal point.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime office illumination remains controlled and neutral-to-cool, preserving detail in the 차장검사's severe expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same executive office, desk, paperwork, daylight, and both men's formal clothing. Exclude the prosecutor's standing outburst from the close frame and isolate the chief prosecutor's stern response.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 책상 위 명패: \"차장검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.429,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.625
   },
   "adjusted": {
    "A": 1.571,
    "B": 1.625
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1571,
   "B": 1625
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "캐릭터 레퍼런스의 의상(더블 브레스티드 수트, 넥타이핀, 행거치프)과 이전 컷의 명패 디테일을 매우 정확하게 재현했으며, 요구된 로우 앵글과 표정 연기를 훌륭하게 소화했습니다."
   },
   {
    "label": "B",
    "score": 1625,
    "verdict_ko": "인물의 표정과 구도는 지시사항을 따랐으나, 캐릭터 레퍼런스의 핵심 의상 디테일(넥타이핀, 행거치프, 수트 형태)을 누락했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S59sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869633>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 샷의 책상 위에 명확히 존재하던 서류철과 문서가 화면 하단 책상 위에서 완전히 누락됨.",
     "fix_en": "Add the official documents and pen to the desk surface right behind the nameplate.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 왼쪽 전경의 어깨가 '최소한의 가장자리(minimal soft edge)'라는 지시와 달리 화면의 상당 부분을 차지할 정도로 크게 렌더링됨.",
     "fix_en": "Reduce the out-of-focus shoulder on the left edge so it only occupies a narrow sliver of the frame.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "배경 TV 모니터 속 사진 인물들의 얼굴에 적용되어 있던 모자이크(픽셀화) 처리가 누락되어 연속성이 떨어짐.",
     "fix_en": "Apply a mosaic blur to the faces of the people shown on the background TV monitor.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "화면 중앙의 입을 벌린 남자가 차장검사 참조 인물이 아니라 이전 컷의 서 있는 남자(장원섭) 얼굴이다.",
     "fix_en": "Replace the seated man's face and hair with the exact features from the provided CHARACTER REFERENCE; remove the face of the standing man from the previous shot. Preserve the seated posture, the severe shouting expression, the framing, the pinstripe suit and tie, the foreground shoulder, the nameplate, and the background office with the monitor.",
     "severity": "critical",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이전 샷의 책상 위에 명확히 존재하던 서류철과 문서가 화면 하단 책상 위에서 완전히 누락됨.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 왼쪽 전경의 어깨가 '최소한의 가장자리(minimal soft edge)'라는 지시와 달리 화면의 상당 부분을 차지할 정도로 크게 렌더링됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "배경 TV 모니터 속 사진 인물들의 얼굴에 적용되어 있던 모자이크(픽셀화) 처리가 누락되어 연속성이 떨어짐.",
     "severity": "minor"
    },
    {
     "issue_ko": "화면 중앙의 입을 벌린 남자가 차장검사 참조 인물이 아니라 이전 컷의 서 있는 남자(장원섭) 얼굴이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "차장검사 자리에 앉은 인물이 참조 차장검사가 아닌 다른 얼굴·머리·체형이다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 샷의 책상 위에 명확히 존재하던 서류철과 문서가 화면 하단 책상 위에서 완전히 누락됨.",
    "fix_en": "Add the official documents and pen to the desk surface right behind the nameplate.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "화면 왼쪽 전경의 어깨가 '최소한의 가장자리(minimal soft edge)'라는 지시와 달리 화면의 상당 부분을 차지할 정도로 크게 렌더링됨.",
    "fix_en": "Reduce the out-of-focus shoulder on the left edge so it only occupies a narrow sliver of the frame.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "배경 TV 모니터 속 사진 인물들의 얼굴에 적용되어 있던 모자이크(픽셀화) 처리가 누락되어 연속성이 떨어짐.",
    "fix_en": "Apply a mosaic blur to the faces of the people shown on the background TV monitor.",
    "severity": "minor",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the seated man's face and hair with the exact features from the provided CHARACTER REFERENCE; remove the face of the standing man from the previous shot. Preserve the seated posture, the severe shouting expression, the framing, the pinstripe suit and tie, the foreground shoulder, the nameplate, and the background office with the monitor.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "모든 프롬프트 지시사항을 완벽히 충족하며, 후보 A에 비해 치아의 세부 묘사가 훨씬 사실적으로 개선되어 시네마틱한 완성도를 높임."
     },
     {
      "label": "A",
      "score": 1900,
      "verdict_ko": "구도와 묘사는 훌륭하나, 벌린 입 안의 치아 형태가 다소 뭉개져 인위적인 느낌을 줌."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.9,
      "B": 1.667
     },
     "adjusted": {
      "A": 1.9,
      "B": 1.667
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.333,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "모든 프롬프트 지시사항을 완벽히 충족하며, 후보 A에 비해 치아의 세부 묘사가 훨씬 사실적으로 개선되어 시네마틱한 완성도를 높임."
     },
     {
      "label": "A",
      "score": 1900,
      "verdict_ko": "구도와 묘사는 훌륭하나, 벌린 입 안의 치아 형태가 다소 뭉개져 인위적인 느낌을 줌."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "후보 B는 차장검사의 매서운 표정과 시선을 훌륭하게 포착했으며, 이전 컷 및 레퍼런스와의 일치도를 완벽히 유지하면서 후보 A에서 발생한 치아의 형태 왜곡을 자연스럽게 수정하여 가장 우수한 결과물입니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도와 앵글, 레퍼런스의 연속성 등 전반적인 지시사항을 잘 따랐으나, 벌린 입 안의 치아가 부자연스럽고 뾰족하게 왜곡되어 렌더링된 형태적 결함이 큽니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "차장검사의 매서운 시선과 고개는 화면 좌측 전경에 부드럽게 걸쳐진 장원섭(어깨 실루엣)을 정확히 향하고 있음.",
      "built_space": "이전 컷과 동일한 사무실 책상 뒤. 명패, 배경의 모니터(화면 내용 일치), '正道' 액자 등 주변 환경이 정확히 일치함.",
      "entities": "차장검사의 얼굴, 헤어스타일, 의상(핀스트라이프 정장, 넥타이)이 레퍼런스와 일치함. 명패의 '차장검사' 텍스트가 정확함. 전경의 장원섭은 이전 컷의 회색 정장을 입은 어깨 부분만 올바르게 노출됨.",
      "hard_violations": [],
      "physics": "화면 밖 의자에 앉아 상체를 약간 기울인 자세가 자연스러우며, 의상의 주름과 질감이 물리적으로 타당하게 안착됨."
     },
     {
      "label": "B",
      "direction": "차장검사의 시선이 좌측 전경의 장원섭(어깨 부분)을 정확히 타겟팅하여 내려다보는 듯한 앵글의 지시를 잘 따름.",
      "built_space": "이전 컷의 책상 가장자리, 명패 위치, 배경 모니터 화면과 액자 등 고정된 실내 환경 요소가 완벽하게 동일함.",
      "entities": "차장검사의 인상착의가 레퍼런스와 완벽히 일치하며, 치아를 포함한 구강 구조가 자연스러운 인간의 형태로 잘 묘사됨. 명패 텍스트 정상. 장원섭의 어깨 쪽에 흰 셔츠 깃이 살짝 보여 인물 간 관계가 더 명확함.",
      "hard_violations": [],
      "physics": "좌석에 앉아 고개를 치켜든 자세가 하체와 허리에 의해 안정적으로 지탱되며, 모든 객체와 인물이 바닥과 중력에 맞게 위치함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "후보 B는 차장검사의 매서운 표정과 시선을 훌륭하게 포착했으며, 이전 컷 및 레퍼런스와의 일치도를 완벽히 유지하면서 후보 A에서 발생한 치아의 형태 왜곡을 자연스럽게 수정하여 가장 우수한 결과물입니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "구도와 앵글, 레퍼런스의 연속성 등 전반적인 지시사항을 잘 따랐으나, 벌린 입 안의 치아가 부자연스럽고 뾰족하게 왜곡되어 렌더링된 형태적 결함이 큽니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "차장검사의 매서운 시선과 고개는 화면 좌측 전경에 부드럽게 걸쳐진 장원섭(어깨 실루엣)을 정확히 향하고 있음.",
      "built_space": "이전 컷과 동일한 사무실 책상 뒤. 명패, 배경의 모니터(화면 내용 일치), '正道' 액자 등 주변 환경이 정확히 일치함.",
      "entities": "차장검사의 얼굴, 헤어스타일, 의상(핀스트라이프 정장, 넥타이)이 레퍼런스와 일치함. 명패의 '차장검사' 텍스트가 정확함. 전경의 장원섭은 이전 컷의 회색 정장을 입은 어깨 부분만 올바르게 노출됨.",
      "hard_violations": [],
      "physics": "화면 밖 의자에 앉아 상체를 약간 기울인 자세가 자연스러우며, 의상의 주름과 질감이 물리적으로 타당하게 안착됨."
     },
     {
      "label": "A",
      "direction": "차장검사의 시선이 좌측 전경의 장원섭(어깨 부분)을 정확히 타겟팅하여 내려다보는 듯한 앵글의 지시를 잘 따름.",
      "built_space": "이전 컷의 책상 가장자리, 명패 위치, 배경 모니터 화면과 액자 등 고정된 실내 환경 요소가 완벽하게 동일함.",
      "entities": "차장검사의 인상착의가 레퍼런스와 완벽히 일치하며, 치아를 포함한 구강 구조가 자연스러운 인간의 형태로 잘 묘사됨. 명패 텍스트 정상. 장원섭의 어깨 쪽에 흰 셔츠 깃이 살짝 보여 인물 간 관계가 더 명확함.",
      "hard_violations": [],
      "physics": "좌석에 앉아 고개를 치켜든 자세가 하체와 허리에 의해 안정적으로 지탱되며, 모든 객체와 인물이 바닥과 중력에 맞게 위치함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1910,
     "B": 1674
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh1"
  }
 },
 "S59sh4::cine": {
  "applied": true,
  "fingerprint": "df521c2eafd1dd82fd63b9f48941da83380d9097b9ca5d392959593b4d0eb9f5",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S59sh4_sel.png",
  "source_sha256": "9ce4525f8252da65967a4dc829e33896c9d0d67b323378d67fc045824013162a",
  "file": "S59sh4_cine.png",
  "latency_ms": 11099
 },
 "S60sh2::signage": {
  "fp": "0d49b227f9af19fc",
  "inscriptions": []
 },
 "era_assess::3f84a6acd892b971": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 한국의 클래식 바 및 모던 바 인테리어",
    "search_terms_native": [
     "한국 클래식바 인테리어",
     "청담동 바 인테리어",
     "모던바 카운터",
     "위스키바 내부"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 작성되어야 하며, 영어나 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "2000년대에서 2010년대 사이 한국의 클래식 바나 모던 바는 특유의 어두운 우드 톤 인테리어, 정갈한 조명, 바텐더의 정장 차림 등 서구식 펍이나 일반적인 바와 구분되는 독특한 로컬 분위기가 있습니다."
   }
  ]
 },
 "era_ref::d9fd318da3aa3fdc": {
  "subject": "2000년대-2010년대 한국의 클래식 바 및 모던 바 인테리어",
  "terms": [
   "한국 클래식바 인테리어",
   "청담동 바 인테리어",
   "모던바 카운터",
   "위스키바 내부"
  ],
  "queries": [
   [
    "한국 클래식 바 인테리어 청담동 위스키바 내부 2000년대 2010년대",
    "모던바 카운터 인테리어 한국 위스키바 내부"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://www.menupan.com/restaurant/restimg/009/zzmenuimg/d1032091_z.jpg",
  "picked_reason_ko": "한국 2000~2010년대 클래식 바의 전형적인 목재 카운터, 가죽 좌석, 주류 진열장과 조명을 가장 명료하고 일상적인 형태로 보여준다.",
  "sha256": "25419d8494b35645e1fe47e02c08482ba88887d7d26077d6ddadc35813694a39",
  "file": "eraref_d9fd318da3aa3fdc.png"
 },
 "S60sh2::bgfirst_bg": {
  "input_fingerprint": "63e6663e2012468c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 전택수의 시선을 피한 채 씁쓸한 기색으로 자신의 술잔을 내려다보는 장원섭의 측면\n\nLOCATION (lock): At the bar counter inside a quiet cocktail lounge, opposite the bartender.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Dolly inward at seated shoulder height from 장원섭's outer side, tightening from the two-person axis until his profile occupies most of the left-center frame. He folds slightly toward the counter and looks down at his own glass with a bitter expression, while only a restrained trace of 전택수 remains farther along the bar axis, watching him rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Both men are seated along it) — Its long edge recedes from 장원섭 toward 전택수 and the bartender's side; used as Carries the line of the dolly move and holds both men within the same conversational space; 장원섭의 술잔 (Resting before 장원섭) — The transparent sides reveal the drink remaining inside; used as Immediate focus for 장원섭's averted gaze, kept small in the lower frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to a cocktail bar at night, with subdued neutral-to-cool color and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국의 클래식 바 및 모던 바 인테리어: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 전택수의 시선을 피한 채 씁쓸한 기색으로 자신의 술잔을 내려다보는 장원섭의 측면\n\nLOCATION (lock): At the bar counter inside a quiet cocktail lounge, opposite the bartender.\n\nTIME OF DAY (lock): night.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Dolly inward at seated shoulder height from 장원섭's outer side, tightening from the two-person axis until his profile occupies most of the left-center frame. He folds slightly toward the counter and looks down at his own glass with a bitter expression, while only a restrained trace of 전택수 remains farther along the bar axis, watching him rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Both men are seated along it) — Its long edge recedes from 장원섭 toward 전택수 and the bartender's side; used as Carries the line of the dolly move and holds both men within the same conversational space; 장원섭의 술잔 (Resting before 장원섭) — The transparent sides reveal the drink remaining inside; used as Immediate focus for 장원섭's averted gaze, kept small in the lower frame.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to a cocktail bar at night, with subdued neutral-to-cool color and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국의 클래식 바 및 모던 바 인테리어: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S60sh2__bgfirst_bg.png",
  "asset_id": "9558a3fd-908f-4eeb-85a4-24bf8018dd00",
  "input_asset_ids": [
   "fd581c2d-a0c4-4de5-abb3-b789adb7c76e",
   "7cd6314b-d3d9-4217-9daa-582e76665db5"
  ],
  "era_research": {
   "subject": "2000년대-2010년대 한국의 클래식 바 및 모던 바 인테리어",
   "queries": [
    [
     "한국 클래식 바 인테리어 청담동 위스키바 내부 2000년대 2010년대",
     "모던바 카운터 인테리어 한국 위스키바 내부"
    ]
   ],
   "picked_url": "https://www.menupan.com/restaurant/restimg/009/zzmenuimg/d1032091_z.jpg",
   "sha256": "25419d8494b35645e1fe47e02c08482ba88887d7d26077d6ddadc35813694a39",
   "file": "eraref_d9fd318da3aa3fdc.png"
  }
 },
 "S60sh2": {
  "input_fingerprint": "541b5c14df371529",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수의 시선을 피한 채 씁쓸한 기색으로 자신의 술잔을 내려다보는 장원섭의 측면\n\nLOCATION (lock): At the bar counter inside a quiet cocktail lounge, opposite the bartender. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Dolly inward at seated shoulder height from 장원섭's outer side, tightening from the two-person axis until his profile occupies most of the left-center frame. He folds slightly toward the counter and looks down at his own glass with a bitter expression, while only a restrained trace of 전택수 remains farther along the bar axis, watching him rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Both men are seated along it) — Its long edge recedes from 장원섭 toward 전택수 and the bartender's side; used as Carries the line of the dolly move and holds both men within the same conversational space; 장원섭의 술잔 (Resting before 장원섭) — The transparent sides reveal the drink remaining inside; used as Immediate focus for 장원섭's averted gaze, kept small in the lower frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to a cocktail bar at night, with subdued neutral-to-cool color and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수의 시선을 피한 채 씁쓸한 기색으로 자신의 술잔을 내려다보는 장원섭의 측면\n\nLOCATION (lock): At the bar counter inside a quiet cocktail lounge, opposite the bartender. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Dolly inward at seated shoulder height from 장원섭's outer side, tightening from the two-person axis until his profile occupies most of the left-center frame. He folds slightly toward the counter and looks down at his own glass with a bitter expression, while only a restrained trace of 전택수 remains farther along the bar axis, watching him rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Both men are seated along it) — Its long edge recedes from 장원섭 toward 전택수 and the bartender's side; used as Carries the line of the dolly move and holds both men within the same conversational space; 장원섭의 술잔 (Resting before 장원섭) — The transparent sides reveal the drink remaining inside; used as Immediate focus for 장원섭's averted gaze, kept small in the lower frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to a cocktail bar at night, with subdued neutral-to-cool color and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수의 시선을 피한 채 씁쓸한 기색으로 자신의 술잔을 내려다보는 장원섭의 측면\n\nLOCATION (lock): At the bar counter inside a quiet cocktail lounge, opposite the bartender. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Dolly inward at seated shoulder height from 장원섭's outer side, tightening from the two-person axis until his profile occupies most of the left-center frame. He folds slightly toward the counter and looks down at his own glass with a bitter expression, while only a restrained trace of 전택수 remains farther along the bar axis, watching him rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Both men are seated along it) — Its long edge recedes from 장원섭 toward 전택수 and the bartender's side; used as Carries the line of the dolly move and holds both men within the same conversational space; 장원섭의 술잔 (Resting before 장원섭) — The transparent sides reveal the drink remaining inside; used as Immediate focus for 장원섭's averted gaze, kept small in the lower frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained ambient illumination appropriate to a cocktail bar at night, with subdued neutral-to-cool color and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S60sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S60sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L52B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "화면 좌측 중앙에 장원섭을 배치하고 바 카운터가 전택수 쪽으로 깊어지는 프레이밍과 구도를 완벽히 구현했으며, 바에 놓인 술잔과 인물의 시선 처리 등 세부 지시사항을 훌륭하게 따랐습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "바 카운터가 지시된 방향과 반대로 연장되어 프레이밍 구도를 어겼으며, 술잔을 손에 쥐고 있고 배경 인물의 얼굴이 주인공과 동일하게 복제되어 출력된 점이 감점 요인입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭은 양손으로 쥔 술잔을 내려다보고 있으며, 배경 좌측의 전택수는 장원섭의 등 쪽을 바라보고 있습니다.",
      "built_space": "바 카운터는 참조 이미지의 형태와 일치하나 좌측으로 깊어지는 구조이며, 두 인물 모두 바 스툴에 앉아 있습니다.",
      "entities": "장원섭의 외모와 정장, 넥타이는 참조 이미지와 일치하지만, 배경의 전택수가 장원섭과 완전히 동일한 얼굴로 생성되었습니다. 술잔은 바닥에 놓여 있지 않고 손에 들려 있습니다.",
      "hard_violations": [],
      "physics": "인물이 스툴에 앉아 술잔을 쥐고 있는 자세는 물리적으로 안정적이며 어색함이 없습니다."
     },
     {
      "label": "B",
      "direction": "장원섭은 카운터 위에 놓인 자신의 술잔을 향해 시선을 아래로 향하고 있으며, 배경 우측의 전택수는 장원섭을 응시하고 있습니다.",
      "built_space": "바 카운터가 우측으로 깊어지며 카메라 축을 형성하고, 우측 배경에 밤거리가 보이는 창(또는 문)이 배치되어 있습니다. 인물들은 바를 따라 정상적으로 착석해 있습니다.",
      "entities": "장원섭의 얼굴, 체형, 의상(도트 무늬 넥타이 포함)이 참조 이미지와 정확히 일치하며, 지시된 대로 술잔이 카운터 위에 놓여 있고 배경의 전택수는 절제된 모습으로 다른 외모를 가집니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 정상적으로 체중을 싣고 앉아 있으며, 술잔은 카운터 표면에 안정적으로 놓여 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "화면 좌측 중앙에 장원섭을 배치하고 바 카운터가 전택수 쪽으로 깊어지는 프레이밍과 구도를 완벽히 구현했으며, 바에 놓인 술잔과 인물의 시선 처리 등 세부 지시사항을 훌륭하게 따랐습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "바 카운터가 지시된 방향과 반대로 연장되어 프레이밍 구도를 어겼으며, 술잔을 손에 쥐고 있고 배경 인물의 얼굴이 주인공과 동일하게 복제되어 출력된 점이 감점 요인입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭은 양손으로 쥔 술잔을 내려다보고 있으며, 배경 좌측의 전택수는 장원섭의 등 쪽을 바라보고 있습니다.",
      "built_space": "바 카운터는 참조 이미지의 형태와 일치하나 좌측으로 깊어지는 구조이며, 두 인물 모두 바 스툴에 앉아 있습니다.",
      "entities": "장원섭의 외모와 정장, 넥타이는 참조 이미지와 일치하지만, 배경의 전택수가 장원섭과 완전히 동일한 얼굴로 생성되었습니다. 술잔은 바닥에 놓여 있지 않고 손에 들려 있습니다.",
      "hard_violations": [],
      "physics": "인물이 스툴에 앉아 술잔을 쥐고 있는 자세는 물리적으로 안정적이며 어색함이 없습니다."
     },
     {
      "label": "B",
      "direction": "장원섭은 카운터 위에 놓인 자신의 술잔을 향해 시선을 아래로 향하고 있으며, 배경 우측의 전택수는 장원섭을 응시하고 있습니다.",
      "built_space": "바 카운터가 우측으로 깊어지며 카메라 축을 형성하고, 우측 배경에 밤거리가 보이는 창(또는 문)이 배치되어 있습니다. 인물들은 바를 따라 정상적으로 착석해 있습니다.",
      "entities": "장원섭의 얼굴, 체형, 의상(도트 무늬 넥타이 포함)이 참조 이미지와 정확히 일치하며, 지시된 대로 술잔이 카운터 위에 놓여 있고 배경의 전택수는 절제된 모습으로 다른 외모를 가집니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 정상적으로 체중을 싣고 앉아 있으며, 술잔은 카운터 표면에 안정적으로 놓여 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "지시된 화면 좌측 중앙 프레이밍을 정확히 구현했으며, 인물의 외양과 장소 레퍼런스의 세부 요소(바 카운터, 선반, 냉장고 등)를 완벽하게 재현하여 프롬프트 충실도가 매우 높습니다."
     },
     {
      "label": "A",
      "score": 875,
      "verdict_ko": "인물을 화면 좌측 중앙에 배치하라는 지시를 어기고 우측에 배치했으며, 지정된 장소에 존재하지 않는 창문을 임의로 추가하여 장소 고정(Location lock) 조건을 위반했습니다.  ★위반: [gemini-pro] 지시된 프레이밍(left-center)과 반대로 인물이 화면 우측에 배치됨. / [gemini-pro] Location lock 위반: 레퍼런스 장소에 없는 창문이 배경에 등장함."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.375,
      "B": 1.571
     },
     "adjusted": {
      "A": 0.875,
      "B": 1.571
     },
     "violations": {
      "A": [
       "[gemini-pro] 지시된 프레이밍(left-center)과 반대로 인물이 화면 우측에 배치됨.",
       "[gemini-pro] Location lock 위반: 레퍼런스 장소에 없는 창문이 배경에 등장함."
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "지시된 화면 좌측 중앙 프레이밍을 정확히 구현했으며, 인물의 외양과 장소 레퍼런스의 세부 요소(바 카운터, 선반, 냉장고 등)를 완벽하게 재현하여 프롬프트 충실도가 매우 높습니다."
     },
     {
      "label": "B",
      "score": 875,
      "verdict_ko": "인물을 화면 좌측 중앙에 배치하라는 지시를 어기고 우측에 배치했으며, 지정된 장소에 존재하지 않는 창문을 임의로 추가하여 장소 고정(Location lock) 조건을 위반했습니다.  ★위반: [gemini-pro] 지시된 프레이밍(left-center)과 반대로 인물이 화면 우측에 배치됨. / [gemini-pro] Location lock 위반: 레퍼런스 장소에 없는 창문이 배경에 등장함."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1575,
     "B": 882
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "장원섭은 양손으로 쥔 술잔을 내려다보고 있으며, 배경 좌측의 전택수는 장원섭의 등 쪽을 바라보고 있습니다.",
    "built_space": "바 카운터는 참조 이미지의 형태와 일치하나 좌측으로 깊어지는 구조이며, 두 인물 모두 바 스툴에 앉아 있습니다.",
    "entities": "장원섭의 외모와 정장, 넥타이는 참조 이미지와 일치하지만, 배경의 전택수가 장원섭과 완전히 동일한 얼굴로 생성되었습니다. 술잔은 바닥에 놓여 있지 않고 손에 들려 있습니다.",
    "hard_violations": [],
    "physics": "인물이 스툴에 앉아 술잔을 쥐고 있는 자세는 물리적으로 안정적이며 어색함이 없습니다."
   },
   {
    "label": "B",
    "direction": "장원섭은 카운터 위에 놓인 자신의 술잔을 향해 시선을 아래로 향하고 있으며, 배경 우측의 전택수는 장원섭을 응시하고 있습니다.",
    "built_space": "바 카운터가 우측으로 깊어지며 카메라 축을 형성하고, 우측 배경에 밤거리가 보이는 창(또는 문)이 배치되어 있습니다. 인물들은 바를 따라 정상적으로 착석해 있습니다.",
    "entities": "장원섭의 얼굴, 체형, 의상(도트 무늬 넥타이 포함)이 참조 이미지와 정확히 일치하며, 지시된 대로 술잔이 카운터 위에 놓여 있고 배경의 전택수는 절제된 모습으로 다른 외모를 가집니다.",
    "hard_violations": [],
    "physics": "두 인물 모두 의자에 정상적으로 체중을 싣고 앉아 있으며, 술잔은 카운터 표면에 안정적으로 놓여 있습니다."
   }
  ],
  "totals": {
   "A": 1575,
   "B": 882
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "화면 좌측 중앙에 장원섭을 배치하고 바 카운터가 전택수 쪽으로 깊어지는 프레이밍과 구도를 완벽히 구현했으며, 바에 놓인 술잔과 인물의 시선 처리 등 세부 지시사항을 훌륭하게 따랐습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "바 카운터가 지시된 방향과 반대로 연장되어 프레이밍 구도를 어겼으며, 술잔을 손에 쥐고 있고 배경 인물의 얼굴이 주인공과 동일하게 복제되어 출력된 점이 감점 요인입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L52B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "배경에 앉아 있는 인물(전택수)이 전경의 기준 인물(장원섭)과 완전히 동일한 얼굴과 복장으로 복제되어 묘사됨.",
     "fix_en": "Redraw the background man with a different face, hairstyle, and suit color to remove the cloning effect. Preserve the foreground man's identity, pose, suit, the bar counter, background, and lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "전경 인물의 왼쪽 어깨에서 이어진 팔 끝에 달린 손이 엄지손가락의 위치(바깥쪽/아래쪽)상 오른손으로 잘못 그려짐.",
     "fix_en": "Redraw the foreground man's hand holding the glass as an anatomically correct left hand with the thumb in the proper position. Preserve the man's pose, face, suit, the glass, bar counter, background, and lighting.",
     "severity": "critical",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "배경에 앉아 있는 인물(전택수)이 전경의 기준 인물(장원섭)과 완전히 동일한 얼굴과 복장으로 복제되어 묘사됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "전경 인물의 왼쪽 어깨에서 이어진 팔 끝에 달린 손이 엄지손가락의 위치(바깥쪽/아래쪽)상 오른손으로 잘못 그려짐.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수가 바 축 먼 곳의 자취가 아니라 왼쪽 중경에 얼굴이 선명하고 크게 앉아 있다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 중경 남자의 얼굴이 장원섭과 동일해 전택수가 다른 인물로 보이지 않는다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the background man with a different face, hairstyle, and suit color to remove the cloning effect. Preserve the foreground man's identity, pose, suit, the bar counter, background, and lighting.\n- Redraw the foreground man's hand holding the glass as an anatomically correct left hand with the thumb in the proper position. Preserve the man's pose, face, suit, the glass, bar counter, background, and lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트의 지시사항과 스케치의 구도, 캐릭터 레퍼런스를 충실히 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "일직선이어야 할 바 카운터의 형태를 임의로 변형하였으며, 기본 설정인 '한국인'을 무시하고 배경 인물을 다른 인종으로 렌더링하여 실격입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전경의 인물(장원섭)은 테이블 위의 술잔을 내려다보고 있으며, 배경의 인물(전택수)은 전경의 인물을 응시하고 있습니다.",
      "built_space": "레퍼런스 이미지의 일직선 바 카운터 구조가 그대로 유지되었으며, 인물들이 카운터 너머로 올바르게 착석해 있습니다.",
      "entities": "전경 인물은 장원섭의 외모, 의상 레퍼런스와 일치합니다. 배경 인물은 지시된 대로 한국인 남성으로 렌더링되었습니다. 장원섭 앞의 술잔도 잘 묘사되었습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 앉아 카운터에 팔을 자연스럽게 지지하고 있습니다."
     },
     {
      "label": "B",
      "direction": "전경의 인물은 술잔을 바라보고, 배경의 인물은 전경 인물을 향해 시선을 두고 있습니다.",
      "built_space": "원본 배경 이미지에서는 일직선으로 뻗어 있던 바 카운터가 배경 인물 쪽에서 갑자기 끊기며 직각 형태의 다른 테이블 구조로 임의 변형되었습니다.",
      "entities": "전경 인물은 장원섭 레퍼런스와 일치합니다. 그러나 배경 인물이 기본 설정인 한국인이 아닌 서양인 외모로 생성되었습니다.",
      "hard_violations": [
       "배경 구조 훼손(카운터 중간이 끊기고 직각 테이블 구조가 임의로 추가됨)",
       "기본 설정 위반(배경 인물이 한국인이 아님)"
      ],
      "physics": "인물들은 의자에 안정적으로 앉아있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "프롬프트의 지시사항과 스케치의 구도, 캐릭터 레퍼런스를 충실히 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "일직선이어야 할 바 카운터의 형태를 임의로 변형하였으며, 기본 설정인 '한국인'을 무시하고 배경 인물을 다른 인종으로 렌더링하여 실격입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전경의 인물(장원섭)은 테이블 위의 술잔을 내려다보고 있으며, 배경의 인물(전택수)은 전경의 인물을 응시하고 있습니다.",
      "built_space": "레퍼런스 이미지의 일직선 바 카운터 구조가 그대로 유지되었으며, 인물들이 카운터 너머로 올바르게 착석해 있습니다.",
      "entities": "전경 인물은 장원섭의 외모, 의상 레퍼런스와 일치합니다. 배경 인물은 지시된 대로 한국인 남성으로 렌더링되었습니다. 장원섭 앞의 술잔도 잘 묘사되었습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 의자에 앉아 카운터에 팔을 자연스럽게 지지하고 있습니다."
     },
     {
      "label": "B",
      "direction": "전경의 인물은 술잔을 바라보고, 배경의 인물은 전경 인물을 향해 시선을 두고 있습니다.",
      "built_space": "원본 배경 이미지에서는 일직선으로 뻗어 있던 바 카운터가 배경 인물 쪽에서 갑자기 끊기며 직각 형태의 다른 테이블 구조로 임의 변형되었습니다.",
      "entities": "전경 인물은 장원섭 레퍼런스와 일치합니다. 그러나 배경 인물이 기본 설정인 한국인이 아닌 서양인 외모로 생성되었습니다.",
      "hard_violations": [
       "배경 구조 훼손(카운터 중간이 끊기고 직각 테이블 구조가 임의로 추가됨)",
       "기본 설정 위반(배경 인물이 한국인이 아님)"
      ],
      "physics": "인물들은 의자에 안정적으로 앉아있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 400,
      "verdict_ko": "전경 인물의 외형과 시선은 지시를 잘 따랐으나, 배경 인물이 백인으로 묘사되고 지정된 바 카운터 구조가 끊어지듯 변형되어 심각한 오류가 있습니다.  ★위반: [gemini-pro] invented people (모든 인물은 한국인이어야 하나 배경에 백인 남성이 등장함) / [gemini-pro] altered built structure (레퍼런스의 일자형 바 카운터 구조가 끊어지고 왜곡됨) / [openrouter:x-ai/grok-4.6] 전택수 자리에 비한국인 백인 남성을 발명해 착석시킴 / [openrouter:x-ai/grok-4.6] 스케치·연출이 정한 손님석 바 축이 아닌 바텐더/서비스 쪽에 착석"
     },
     {
      "label": "B",
      "score": 917,
      "verdict_ko": "전경 인물의 묘사는 좋으나, 배경 인물에 장원섭의 얼굴과 복장이 그대로 복제되었고 스케치의 지정 위치를 완전히 벗어나 화면 좌측에 배치되어 실격입니다.  ★위반: [gemini-pro] duplicated bodies (장원섭의 신원과 복장이 배경 인물에 그대로 복제됨) / [gemini-pro] person seated in a place the staging does not put them (스케치의 지정된 우측 바 위치를 무시하고 좌측 구석에 인물을 배치함) / [openrouter:x-ai/grok-4.6] 전택수를 바가 후퇴하는 손님석 축이 아닌 반대쪽 끝에 착석시킴"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.4,
      "B": 1.667
     },
     "adjusted": {
      "A": 0.4,
      "B": 0.917
     },
     "violations": {
      "A": [
       "[gemini-pro] invented people (모든 인물은 한국인이어야 하나 배경에 백인 남성이 등장함)",
       "[gemini-pro] altered built structure (레퍼런스의 일자형 바 카운터 구조가 끊어지고 왜곡됨)",
       "[openrouter:x-ai/grok-4.6] 전택수 자리에 비한국인 백인 남성을 발명해 착석시킴",
       "[openrouter:x-ai/grok-4.6] 스케치·연출이 정한 손님석 바 축이 아닌 바텐더/서비스 쪽에 착석"
      ],
      "B": [
       "[gemini-pro] duplicated bodies (장원섭의 신원과 복장이 배경 인물에 그대로 복제됨)",
       "[gemini-pro] person seated in a place the staging does not put them (스케치의 지정된 우측 바 위치를 무시하고 좌측 구석에 인물을 배치함)",
       "[openrouter:x-ai/grok-4.6] 전택수를 바가 후퇴하는 손님석 축이 아닌 반대쪽 끝에 착석시킴"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.6,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 400,
      "verdict_ko": "전경 인물의 외형과 시선은 지시를 잘 따랐으나, 배경 인물이 백인으로 묘사되고 지정된 바 카운터 구조가 끊어지듯 변형되어 심각한 오류가 있습니다.  ★위반: [gemini-pro] invented people (모든 인물은 한국인이어야 하나 배경에 백인 남성이 등장함) / [gemini-pro] altered built structure (레퍼런스의 일자형 바 카운터 구조가 끊어지고 왜곡됨) / [openrouter:x-ai/grok-4.6] 전택수 자리에 비한국인 백인 남성을 발명해 착석시킴 / [openrouter:x-ai/grok-4.6] 스케치·연출이 정한 손님석 바 축이 아닌 바텐더/서비스 쪽에 착석"
     },
     {
      "label": "A",
      "score": 917,
      "verdict_ko": "전경 인물의 묘사는 좋으나, 배경 인물에 장원섭의 얼굴과 복장이 그대로 복제되었고 스케치의 지정 위치를 완전히 벗어나 화면 좌측에 배치되어 실격입니다.  ★위반: [gemini-pro] duplicated bodies (장원섭의 신원과 복장이 배경 인물에 그대로 복제됨) / [gemini-pro] person seated in a place the staging does not put them (스케치의 지정된 우측 바 위치를 무시하고 좌측 구석에 인물을 배치함) / [openrouter:x-ai/grok-4.6] 전택수를 바가 후퇴하는 손님석 축이 아닌 반대쪽 끝에 착석시킴"
     }
    ],
    "all_candidates_fail": true
   },
   "combined": {
    "totals": {
     "A": 925,
     "B": 403
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S60sh2__bgfirst_bg.png",
   "bg_asset_id": "9558a3fd-908f-4eeb-85a4-24bf8018dd00",
   "bg_record_key": "S60sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S60sh2::cine": {
  "applied": true,
  "fingerprint": "eca02f287657812c2877f559bb9ba5f4738184691cce053a6c2b4fd415d7c320",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S60sh2_sel.png",
  "source_sha256": "91c774597204d8abf8eaef1a925c29212553fb59675a71cc736599f2ae8f80d0",
  "file": "S60sh2_cine.png",
  "latency_ms": 9319
 },
 "S60sh10::signage": {
  "fp": "70fd2987590c0c51",
  "inscriptions": []
 },
 "S60sh10": {
  "input_fingerprint": "141b75dab6bc1398",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수의 말에 정곡을 찔린 듯 동공이 확장된 채 멈칫한 장원섭의 얼굴 클로즈업\n\nLOCATION (lock): At the bar counter inside the cocktail lounge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from just behind 전택수's near shoulder at slightly below 장원섭's eye line, arriving at a facial close-up without becoming an extreme close-up. 장원섭's face occupies the center-left as he freezes mid-response with widened pupils and fixed attention on 전택수; a narrow, soft shoulder edge at frame right preserves the source of the proposal.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Present beside the seated men) — Only a receding portion remains visible below and behind 장원섭; used as A minimal contextual line beneath the close-up, keeping the exchange grounded in the bar.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime bar ambience holds a subdued neutral-to-cool palette and natural facial detail at the moment of realization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same dim cocktail bar, counter materials, glassware, and warm night lighting. Exclude the senior officer from the close frame and isolate the prosecutor's startled reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both men's emptied glasses have been refilled, and the liquor bottle remains in use at the bar.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수의 말에 정곡을 찔린 듯 동공이 확장된 채 멈칫한 장원섭의 얼굴 클로즈업\n\nLOCATION (lock): At the bar counter inside the cocktail lounge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from just behind 전택수's near shoulder at slightly below 장원섭's eye line, arriving at a facial close-up without becoming an extreme close-up. 장원섭's face occupies the center-left as he freezes mid-response with widened pupils and fixed attention on 전택수; a narrow, soft shoulder edge at frame right preserves the source of the proposal.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Present beside the seated men) — Only a receding portion remains visible below and behind 장원섭; used as A minimal contextual line beneath the close-up, keeping the exchange grounded in the bar.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime bar ambience holds a subdued neutral-to-cool palette and natural facial detail at the moment of realization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same dim cocktail bar, counter materials, glassware, and warm night lighting. Exclude the senior officer from the close frame and isolate the prosecutor's startled reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both men's emptied glasses have been refilled, and the liquor bottle remains in use at the bar.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 전택수의 말에 정곡을 찔린 듯 동공이 확장된 채 멈칫한 장원섭의 얼굴 클로즈업\n\nLOCATION (lock): At the bar counter inside the cocktail lounge. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from just behind 전택수's near shoulder at slightly below 장원섭's eye line, arriving at a facial close-up without becoming an extreme close-up. 장원섭's face occupies the center-left as he freezes mid-response with widened pupils and fixed attention on 전택수; a narrow, soft shoulder edge at frame right preserves the source of the proposal.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 바 카운터 (Present beside the seated men) — Only a receding portion remains visible below and behind 장원섭; used as A minimal contextual line beneath the close-up, keeping the exchange grounded in the bar.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Restrained nighttime bar ambience holds a subdued neutral-to-cool palette and natural facial detail at the moment of realization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same dim cocktail bar, counter materials, glassware, and warm night lighting. Exclude the senior officer from the close frame and isolate the prosecutor's startled reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Both men's emptied glasses have been refilled, and the liquor bottle remains in use at the bar.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "장원섭의 시선은 화면 우측 프레임에 걸친 인물(전택수)의 어깨 쪽을 향하고 있습니다.",
    "built_space": "바 카운터가 인물 앞을 가로질러 우측으로 멀어집니다. 그러나 카메라 앵글상 뒤에 보여야 할 바의 술병 진열장 대신 출처 불명의 램프와 어두운 벽면이 나타납니다.",
    "entities": "장원섭의 얼굴과 복장(정장, 도트 무늬 넥타이)이 레퍼런스와 일치합니다. 프롬프트가 요구한 술병과 잔이 카운터 위에 존재합니다.",
    "hard_violations": [],
    "physics": "카운터 위의 술잔과 술병은 바닥에 안정적으로 놓여 있으며, 인물의 팔 자세 또한 지지면에 맞게 자연스럽습니다."
   },
   {
    "label": "B",
    "direction": "장원섭의 확장된 동공과 시선이 화면 우측의 어깨(전택수)를 정확히 응시하고 있습니다.",
    "built_space": "인물 앞의 카운터와 배경의 술병 진열장이 카메라 앵글에 맞는 올바른 기하학적 구조와 위치에 배치되어 있습니다.",
    "entities": "장원섭의 인상착의가 레퍼런스와 완벽히 일치하며 우측에 어깨가 배치되어 있습니다. 인물이 잔을 들고 있으나 카운터 위 술병은 명확히 보이지 않습니다.",
    "hard_violations": [],
    "physics": "인물이 들고 있는 술잔은 손에 의해 단단히 쥐어져 있으며(지지됨), 자세와 중력 표현에 어색함이 없습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 8,
   "A": 6
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "앵글에 따른 배경 진열장의 구조를 완벽히 구현하여 공간 일관성(Priority 3)을 충족했고, 인물의 클로즈업과 시선 처리도 매우 훌륭합니다."
   },
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "요구된 소품(술병)은 등장하나 카메라 앵글상 보여야 할 백바(진열장) 대신 엉뚱한 벽면과 조명을 배치하여 공간적 일관성이 부족합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S60sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "장원섭의 얼굴(눈매, 얼굴형 등)이 레퍼런스 이미지와 전혀 다른 인물로 렌더링되었습니다.",
     "fix_en": "Redraw the main character's face to exactly match the reference identity's facial structure, eye shape, and proportions. Preserve the people present and their positions, their clothing, the hair, the set, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "장원섭의 머리 가르마 방향이 레퍼런스(인물 기준 왼쪽)와 반대로 뒤집혀 있습니다.",
     "fix_en": "Change the character's hair parting to his left side. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "배경의 바 진열장 구조가 이전 컷 스틸에서 확립된 형태와 다르게 변경되었습니다.",
     "fix_en": "Restore the background bar display to match the metal cabinet and shelving structure of the previous still. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "넥타이의 무늬가 레퍼런스의 일정한 단일 도트가 아닌, 여러 점이 뭉친 패턴으로 변형되었습니다.",
     "fix_en": "Redraw the tie pattern to feature evenly spaced single dots instead of clusters. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신과 바가 넓게 보이는 미디엄 클로즈업이다.",
     "fix_en": "Push the camera in to frame a tight facial close-up. Preserve the people present, their clothing, the set, and the light.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "프레임 오른쪽에 상대의 머리와 넓은 어깨·등이 들어와 좁은 어깨 가장자리가 아니다.",
     "fix_en": "Adjust the staging on the right to show only a narrow shoulder edge instead of a head and wide back. Preserve the main character's position, his clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "장원섭의 얼굴(눈매, 얼굴형 등)이 레퍼런스 이미지와 전혀 다른 인물로 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "장원섭의 머리 가르마 방향이 레퍼런스(인물 기준 왼쪽)와 반대로 뒤집혀 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경의 바 진열장 구조가 이전 컷 스틸에서 확립된 형태와 다르게 변경되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "넥타이의 무늬가 레퍼런스의 일정한 단일 도트가 아닌, 여러 점이 뭉친 패턴으로 변형되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신과 바가 넓게 보이는 미디엄 클로즈업이다.",
     "severity": "major"
    },
    {
     "issue_ko": "프레임 오른쪽에 상대의 머리와 넓은 어깨·등이 들어와 좁은 어깨 가장자리가 아니다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "장원섭의 머리 가르마 방향이 레퍼런스(인물 기준 왼쪽)와 반대로 뒤집혀 있습니다.",
    "fix_en": "Change the character's hair parting to his left side. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "배경의 바 진열장 구조가 이전 컷 스틸에서 확립된 형태와 다르게 변경되었습니다.",
    "fix_en": "Restore the background bar display to match the metal cabinet and shelving structure of the previous still. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "넥타이의 무늬가 레퍼런스의 일정한 단일 도트가 아닌, 여러 점이 뭉친 패턴으로 변형되었습니다.",
    "fix_en": "Redraw the tie pattern to feature evenly spaced single dots instead of clusters. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "minor",
    "observation_index": 3
   },
   {
    "issue_ko": "얼굴 클로즈업이 아니라 상반신과 바가 넓게 보이는 미디엄 클로즈업이다.",
    "fix_en": "Push the camera in to frame a tight facial close-up. Preserve the people present, their clothing, the set, and the light.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "프레임 오른쪽에 상대의 머리와 넓은 어깨·등이 들어와 좁은 어깨 가장자리가 아니다.",
    "fix_en": "Adjust the staging on the right to show only a narrow shoulder edge instead of a head and wide back. Preserve the main character's position, his clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the main character's face to exactly match the reference identity's facial structure, eye shape, and proportions. Preserve the people present and their positions, their clothing, the hair, the set, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 앵글과 인물의 표정은 좋으나, 유리잔을 쥔 손의 손가락이 6개로 묘사된 치명적인 해부학적 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A의 다지증 오류를 원본의 훼손 없이 완벽하게 수정하였으며, 클로즈업 앵글과 동공이 확장된 표정을 훌륭하게 구현했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선은 화면 우측에 걸쳐진 전택수의 어깨 실루엣을 향해 고정되어 있음.",
      "built_space": "칵테일 바 내부로, 인물 앞의 목재 카운터와 배경의 술병 진열장이 카메라 앵글에 맞게 배치됨.",
      "entities": "장원섭은 레퍼런스와 일치하는 외모와 복장(수트, 넥타이)을 갖추고 있으나, 카운터 위에서 사용 중인 술병은 명확히 보이지 않음.",
      "hard_violations": [
       "해부학적 오류: 유리잔을 쥐고 있는 왼쪽 손의 손가락이 6개로 묘사됨."
      ],
      "physics": "왼팔이 카운터에 지지된 채 유리잔을 쥐고 있으나 손가락 구조가 물리적으로 불가능함."
     },
     {
      "label": "B",
      "direction": "장원섭의 시선과 고개가 화면 우측 전택수를 향해 정확히 조준됨.",
      "built_space": "칵테일 바 내부 구조와 카운터, 배경 진열장이 A와 동일하게 올바른 비율로 렌더링됨.",
      "entities": "장원섭의 얼굴과 복장이 레퍼런스와 완벽히 일치하며, 우측 전택수의 어깨 실루엣도 지시대로 포함됨.",
      "hard_violations": [],
      "physics": "왼손이 유리잔을 자연스럽고 해부학적으로 올바르게 쥐고 있으며 카운터에 안정적으로 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 앵글과 인물의 표정은 좋으나, 유리잔을 쥔 손의 손가락이 6개로 묘사된 치명적인 해부학적 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "A의 다지증 오류를 원본의 훼손 없이 완벽하게 수정하였으며, 클로즈업 앵글과 동공이 확장된 표정을 훌륭하게 구현했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선은 화면 우측에 걸쳐진 전택수의 어깨 실루엣을 향해 고정되어 있음.",
      "built_space": "칵테일 바 내부로, 인물 앞의 목재 카운터와 배경의 술병 진열장이 카메라 앵글에 맞게 배치됨.",
      "entities": "장원섭은 레퍼런스와 일치하는 외모와 복장(수트, 넥타이)을 갖추고 있으나, 카운터 위에서 사용 중인 술병은 명확히 보이지 않음.",
      "hard_violations": [
       "해부학적 오류: 유리잔을 쥐고 있는 왼쪽 손의 손가락이 6개로 묘사됨."
      ],
      "physics": "왼팔이 카운터에 지지된 채 유리잔을 쥐고 있으나 손가락 구조가 물리적으로 불가능함."
     },
     {
      "label": "B",
      "direction": "장원섭의 시선과 고개가 화면 우측 전택수를 향해 정확히 조준됨.",
      "built_space": "칵테일 바 내부 구조와 카운터, 배경 진열장이 A와 동일하게 올바른 비율로 렌더링됨.",
      "entities": "장원섭의 얼굴과 복장이 레퍼런스와 완벽히 일치하며, 우측 전택수의 어깨 실루엣도 지시대로 포함됨.",
      "hard_violations": [],
      "physics": "왼손이 유리잔을 자연스럽고 해부학적으로 올바르게 쥐고 있으며 카운터에 안정적으로 지지됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "프롬프트가 강조한 '확장된 동공'을 정확히 구현하여 캐릭터의 놀란 상태를 매우 충실하게 표현함."
     },
     {
      "label": "A",
      "score": 1778,
      "verdict_ko": "전반적인 구도와 인물의 외양은 훌륭하나, 프롬프트에서 구체적으로 요구한 '확장된 동공'이 묘사되지 않음."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.778,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.778,
      "B": 1.75
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1750,
      "verdict_ko": "프롬프트가 강조한 '확장된 동공'을 정확히 구현하여 캐릭터의 놀란 상태를 매우 충실하게 표현함."
     },
     {
      "label": "B",
      "score": 1778,
      "verdict_ko": "전반적인 구도와 인물의 외양은 훌륭하나, 프롬프트에서 구체적으로 요구한 '확장된 동공'이 묘사되지 않음."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1754,
     "B": 1785
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S60sh2"
  }
 },
 "S60sh10::cine": {
  "applied": true,
  "fingerprint": "1bf7c4fc7ccebad2f81cb7fd0817a9ccbb36e585084cf6379e6fd227b6e3b9b8",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S60sh10_sel.png",
  "source_sha256": "6c1771b2620bfdff633e21de0ebd76d4cbd2898699f522f99b3b97a1f4106ca7",
  "file": "S60sh10_cine.png",
  "latency_ms": 10888
 },
 "S60sh12::signage": {
  "fp": "273f8903c499b656",
  "inscriptions": []
 },
 "S60sh12": {
  "input_fingerprint": "ea54ccf51be98884",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽으로 상체를 바짝 내민 채 절실한 눈빛을 쏘아보는 전택수의 얼굴\n\nLOCATION (lock): At adjacent seats along the cocktail bar counter. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind 장원섭's near shoulder at a low seated eye line, the camera preserves the established diagonal and advances only into a tighter close framing of 전택수. 전택수 occupies the center-right with his face marginally above the lens, torso pitched toward 장원섭 and eyes fixed on him rather than the camera; 장원섭's shoulder remains a narrow foreground anchor at the left edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: bar counter edge (partially visible) — Its long edge recedes diagonally behind the seated men; used as A narrow, softly held strip behind 전택수 preserves the bar location without drawing focus from his face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Existing nighttime bar ambience is restrained in a subdued neutral-to-cool palette with natural facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same dark bar counter, liquor glasses, warm light, and both men's suits. Exclude the prosecutor's downcast pose and frame the senior officer leaning forward urgently.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newly refilled liquor glasses remain in front of Taeksu and Wonseop as Taeksu presses his proposal.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽으로 상체를 바짝 내민 채 절실한 눈빛을 쏘아보는 전택수의 얼굴\n\nLOCATION (lock): At adjacent seats along the cocktail bar counter. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind 장원섭's near shoulder at a low seated eye line, the camera preserves the established diagonal and advances only into a tighter close framing of 전택수. 전택수 occupies the center-right with his face marginally above the lens, torso pitched toward 장원섭 and eyes fixed on him rather than the camera; 장원섭's shoulder remains a narrow foreground anchor at the left edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: bar counter edge (partially visible) — Its long edge recedes diagonally behind the seated men; used as A narrow, softly held strip behind 전택수 preserves the bar location without drawing focus from his face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Existing nighttime bar ambience is restrained in a subdued neutral-to-cool palette with natural facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same dark bar counter, liquor glasses, warm light, and both men's suits. Exclude the prosecutor's downcast pose and frame the senior officer leaning forward urgently.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newly refilled liquor glasses remain in front of Taeksu and Wonseop as Taeksu presses his proposal.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽으로 상체를 바짝 내민 채 절실한 눈빛을 쏘아보는 전택수의 얼굴\n\nLOCATION (lock): At adjacent seats along the cocktail bar counter. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just behind 장원섭's near shoulder at a low seated eye line, the camera preserves the established diagonal and advances only into a tighter close framing of 전택수. 전택수 occupies the center-right with his face marginally above the lens, torso pitched toward 장원섭 and eyes fixed on him rather than the camera; 장원섭's shoulder remains a narrow foreground anchor at the left edge.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: bar counter edge (partially visible) — Its long edge recedes diagonally behind the seated men; used as A narrow, softly held strip behind 전택수 preserves the bar location without drawing focus from his face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Existing nighttime bar ambience is restrained in a subdued neutral-to-cool palette with natural facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same dark bar counter, liquor glasses, warm light, and both men's suits. Exclude the prosecutor's downcast pose and frame the senior officer leaning forward urgently.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newly refilled liquor glasses remain in front of Taeksu and Wonseop as Taeksu presses his proposal.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수의 시선과 몸의 방향이 좌측 전경의 장원섭을 정확히 향하고 있음.",
    "built_space": "바 카운터의 대각선 구도와 배경의 술병들이 올바르게 배치됨.",
    "entities": "전택수의 얼굴과 머리 스타일이 캐릭터 레퍼런스와 일치하며 지정된 복장을 갖춤.",
    "hard_violations": [],
    "physics": "바 카운터에 팔을 기대어 상체를 앞쪽으로 내민 자세를 자연스럽게 지지함."
   },
   {
    "label": "B",
    "direction": "전택수의 시선이 좌측 전경의 장원섭을 향함.",
    "built_space": "바 카운터와 배경 구조가 지침에 맞게 배치됨.",
    "entities": "전택수의 외모가 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "카운터 위에 양손으로 잔을 잡고 있으나, 상체를 내밀지 않고 곧게 앉은 자세임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "텍스트에 명시된 '상체를 바짝 내민 채 절실한 눈빛을 쏘아보는' 핵심 동작을 완벽히 구현했으며, 프레이밍과 조명 지침도 훌륭하게 따랐습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "인물의 외형과 카메라 구도는 지침을 따랐으나, 가장 중요한 '상체를 바짝 내민' 연기를 반영하지 못하고 곧게 앉아 있어 아쉽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S60sh10_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 우측에 있는 전택수의 얼굴이 캐릭터 레퍼런스 이미지의 인물(연령대, 이목구비, 얼굴형)과 전혀 다릅니다.",
     "fix_en": "Replace the face and hair of the man facing the camera to exactly match the character reference for 전택수 (mid-50s Korean man, square face, short neat black hair). Preserve his desperate expression, his dark suit and tie, the shoulder of the man on the left, the glass of liquor, the bar counter, and the background lighting.",
     "severity": "critical",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 우측에 있는 전택수의 얼굴이 캐릭터 레퍼런스 이미지의 인물(연령대, 이목구비, 얼굴형)과 전혀 다릅니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 중앙 전택수의 얼굴·나이·머리가 지정 레퍼런스(50대 중반, 각진 얼굴, 짧은 머리)와 전혀 다른 인물이다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the face and hair of the man facing the camera to exactly match the character reference for 전택수 (mid-50s Korean man, square face, short neat black hair). Preserve his desperate expression, his dark suit and tie, the shoulder of the man on the left, the glass of liquor, the bar counter, and the background lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도, 인물 외형, 조명이 지시사항을 잘 따랐으나, 술잔을 쥔 손가락이 다소 길게 왜곡되어 렌더링된 점이 아쉬움."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "A와 전반적으로 동일하나, 술잔을 쥔 손이 유리와 뭉개지듯 융합되는 물리적으로 불가능한 형태 오류가 발생하여 더 나쁨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선이 앞쪽에 앉은 장원섭의 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사선으로 뻗은 바 카운터와 배경의 술병 진열장이 올바르게 배치됨.",
      "entities": "전택수(50대 중반 한국 남성, 짧은 머리), 장원섭(어깨와 뒷모습), 술잔 모두 지시와 일치함.",
      "hard_violations": [],
      "physics": "전택수가 앞을 향해 상체를 기울인 채 앉아 있으며, 오른손으로 술잔을 쥐고 있으나 손가락 관절이 다소 비현실적으로 김."
     },
     {
      "label": "B",
      "direction": "전택수의 시선이 앞쪽에 앉은 장원섭의 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사선으로 뻗은 바 카운터와 배경의 술병 진열장이 올바르게 배치됨.",
      "entities": "전택수(50대 중반 한국 남성, 짧은 머리), 장원섭(어깨와 뒷모습), 술잔 모두 지시와 일치함.",
      "hard_violations": [
       "physically impossible anatomy (술잔을 쥔 손가락이 유리잔의 형태와 녹아붙어 뭉개짐)"
      ],
      "physics": "전택수가 앞을 향해 상체를 기울인 채 앉아 있으나, 술잔을 쥐고 있는 손이 지지체(술잔)와 융합되어 물리적으로 불가능한 형태를 보임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "구도, 인물 외형, 조명이 지시사항을 잘 따랐으나, 술잔을 쥔 손가락이 다소 길게 왜곡되어 렌더링된 점이 아쉬움."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "A와 전반적으로 동일하나, 술잔을 쥔 손이 유리와 뭉개지듯 융합되는 물리적으로 불가능한 형태 오류가 발생하여 더 나쁨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선이 앞쪽에 앉은 장원섭의 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사선으로 뻗은 바 카운터와 배경의 술병 진열장이 올바르게 배치됨.",
      "entities": "전택수(50대 중반 한국 남성, 짧은 머리), 장원섭(어깨와 뒷모습), 술잔 모두 지시와 일치함.",
      "hard_violations": [],
      "physics": "전택수가 앞을 향해 상체를 기울인 채 앉아 있으며, 오른손으로 술잔을 쥐고 있으나 손가락 관절이 다소 비현실적으로 김."
     },
     {
      "label": "B",
      "direction": "전택수의 시선이 앞쪽에 앉은 장원섭의 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "사선으로 뻗은 바 카운터와 배경의 술병 진열장이 올바르게 배치됨.",
      "entities": "전택수(50대 중반 한국 남성, 짧은 머리), 장원섭(어깨와 뒷모습), 술잔 모두 지시와 일치함.",
      "hard_violations": [
       "physically impossible anatomy (술잔을 쥔 손가락이 유리잔의 형태와 녹아붙어 뭉개짐)"
      ],
      "physics": "전택수가 앞을 향해 상체를 기울인 채 앉아 있으나, 술잔을 쥐고 있는 손이 지지체(술잔)와 융합되어 물리적으로 불가능한 형태를 보임."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1750,
      "verdict_ko": "배경 금속 셰이커에 생성되었던 부자연스러운 인공물(손잡이)을 제거하여 소품의 물리적 사실성을 복원했으며, 캐릭터의 정체성과 절실한 시선 처리, 프레이밍 등 지시문의 모든 요구사항을 훌륭하게 충족했습니다."
     },
     {
      "label": "A",
      "score": 1528,
      "verdict_ko": "지시된 프레이밍, 캐릭터의 외양 및 무게감이 실린 포즈는 매우 잘 구현되었으나, 배경 좌측의 칵테일 셰이커에 용도를 알 수 없는 어색한 금속 고리가 달려 있어 물리적 사실성이 다소 떨어집니다.  ★위반: [gemini-pro] 배경 금속 셰이커에 부착된 물리적으로 불가능하거나 어색한 형태의 얇은 손잡이(인공물 생성 오류)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.778,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.528,
      "B": 1.75
     },
     "violations": {
      "A": [
       "[gemini-pro] 배경 금속 셰이커에 부착된 물리적으로 불가능하거나 어색한 형태의 얇은 손잡이(인공물 생성 오류)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1750,
      "verdict_ko": "배경 금속 셰이커에 생성되었던 부자연스러운 인공물(손잡이)을 제거하여 소품의 물리적 사실성을 복원했으며, 캐릭터의 정체성과 절실한 시선 처리, 프레이밍 등 지시문의 모든 요구사항을 훌륭하게 충족했습니다."
     },
     {
      "label": "B",
      "score": 1528,
      "verdict_ko": "지시된 프레이밍, 캐릭터의 외양 및 무게감이 실린 포즈는 매우 잘 구현되었으나, 배경 좌측의 칵테일 셰이커에 용도를 알 수 없는 어색한 금속 고리가 달려 있어 물리적 사실성이 다소 떨어집니다.  ★위반: [gemini-pro] 배경 금속 셰이커에 부착된 물리적으로 불가능하거나 어색한 형태의 얇은 손잡이(인공물 생성 오류)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1757,
     "B": 1533
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S60sh10"
  }
 },
 "S60sh12::cine": {
  "applied": true,
  "fingerprint": "77e874ea6439d7b4cd81814f9bef64264f8a412b8822e134cd9937b8251250ca",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S60sh12_sel.png",
  "source_sha256": "a611da53e45fb1d7f648fef8ad499e07a9e839a0314c9da2cb9e6a600e057b92",
  "file": "S60sh12_cine.png",
  "latency_ms": 10105
 },
 "S61sh3::signage": {
  "fp": "183a701cec6a283d",
  "inscriptions": []
 },
 "S61sh3": {
  "input_fingerprint": "1e5fb6f935dd3dd8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 목에 핏대를 세운 채 크게 입을 벌리고 있는 차장검사의 얼굴 클로즈업\n\nLOCATION (lock): In the seated meeting area of the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the established position just behind and beside 장원섭, the dolly-in ends on an extreme facial close-up of 차장검사 while retaining the same diagonal sightline and seated eye level. His open mouth and strained neck occupy most of the frame, his face turned three-quarters toward the off-screen 장원섭 rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-center of the frame, foreground, looks toward 장원섭 beyond the frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime office retains moderate contrast and a subdued neutral-to-cool rendering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 차장검사 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same chief prosecutor's office, desk, paperwork, daylight, and formal attire. Exclude the subordinate prosecutor from the close frame and show the chief shouting.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 목에 핏대를 세운 채 크게 입을 벌리고 있는 차장검사의 얼굴 클로즈업\n\nLOCATION (lock): In the seated meeting area of the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the established position just behind and beside 장원섭, the dolly-in ends on an extreme facial close-up of 차장검사 while retaining the same diagonal sightline and seated eye level. His open mouth and strained neck occupy most of the frame, his face turned three-quarters toward the off-screen 장원섭 rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-center of the frame, foreground, looks toward 장원섭 beyond the frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime office retains moderate contrast and a subdued neutral-to-cool rendering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 차장검사 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same chief prosecutor's office, desk, paperwork, daylight, and formal attire. Exclude the subordinate prosecutor from the close frame and show the chief shouting.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 목에 핏대를 세운 채 크게 입을 벌리고 있는 차장검사의 얼굴 클로즈업\n\nLOCATION (lock): In the seated meeting area of the deputy chief prosecutor's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the established position just behind and beside 장원섭, the dolly-in ends on an extreme facial close-up of 차장검사 while retaining the same diagonal sightline and seated eye level. His open mouth and strained neck occupy most of the frame, his face turned three-quarters toward the off-screen 장원섭 rather than toward the lens.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-center of the frame, foreground, looks toward 장원섭 beyond the frame.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient illumination appropriate to the daytime office retains moderate contrast and a subdued neutral-to-cool rendering.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 차장검사 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same chief prosecutor's office, desk, paperwork, daylight, and formal attire. Exclude the subordinate prosecutor from the close frame and show the chief shouting.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "차장검사의 시선과 고개가 화면 좌측 전경에 배치된 인물(장원섭)의 어깨 너머를 향하고 있음.",
    "built_space": "사무실 내부. 책상, 명패, 붓글씨 액자 및 레퍼런스와 동일한 인물들이 띄워진 TV 화면이 올바른 위치에 배치됨.",
    "entities": "차장검사의 외모, 헤어스타일, 정장 복장이 레퍼런스와 일치하며, 땀이 맺힌 얼굴과 벌린 입 등 묘사된 감정 상태가 잘 구현됨.",
    "hard_violations": [],
    "physics": "의자에 앉아 상체를 약간 앞으로 기울인 채 소리치는 자연스러운 자세이며, 물리적으로 올바르게 지탱됨."
   },
   {
    "label": "B",
    "direction": "차장검사의 시선이 화면 좌측 전경의 인물을 향해 정확히 고정되어 있음.",
    "built_space": "사무실 환경은 유지되었으나, 뒷배경의 TV 모니터 화면이 이전 샷의 군중 이미지에서 도시 야경으로 임의로 변경됨.",
    "entities": "차장검사의 기본 인상과 복장은 일치하나, 텍스트가 요구한 '목에 세운 핏대'가 마치 붉은 펜으로 그은 선이나 피부 질환처럼 비현실적으로 렌더링됨.",
    "hard_violations": [],
    "physics": "의자에 앉아 있는 자세로, 신체와 물체의 지탱 상태가 물리적으로 정상임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "이전 샷의 배경(TV 화면 속 인물들)을 완벽하게 유지했으며, 핏대가 서고 크게 벌린 입의 긴장감이 사실적인 질감으로 훌륭하게 묘사되었습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "목의 핏대가 해부학적으로 부자연스러운 붉은 상처나 그래픽처럼 묘사되었으며, 이전 샷에서 연속되어야 할 TV 화면의 내용이 도시 풍경으로 변경되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 차장검사 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S59sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869633>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 샷 래퍼런스에서 배제하라고 명시한 인물(화면 왼쪽 전경의 뒷모습)이 여전히 프레임에 포함되어 있습니다.",
     "fix_en": "Remove the out-of-focus person's shoulder in the left foreground entirely, replacing that area with the out-of-focus background wall and framed calligraphy, preserving the main character, his expression, suit, and the rest of the office setting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "극단 얼굴 클로즈업이 아니며 벌린 입과 긴장한 목이 프레임 대부분을 차지하지 않고 상반신·책상 명패·배경이 크게 보임",
     "fix_en": "Blur the desk, nameplate, and lower torso to simulate a tighter focus on the character's face, preserving the character's facial expression, lighting, and the existing camera angle.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "목에 핏대가 서 있지 않음",
     "fix_en": "Add prominent, bulging veins to the character's strained neck, preserving his face, open mouth, expression, suit, and the background.",
     "severity": "minor",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라 프레이밍이 지시된 '얼굴 클로즈업(extreme facial close-up)'이 아닌 미디엄 클로즈업(가슴까지 보임)으로 촬영되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷 래퍼런스에서 배제하라고 명시한 인물(화면 왼쪽 전경의 뒷모습)이 여전히 프레임에 포함되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "극단 얼굴 클로즈업이 아니며 벌린 입과 긴장한 목이 프레임 대부분을 차지하지 않고 상반신·책상 명패·배경이 크게 보임",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 왼쪽 전경에 샷에 없어야 할 다른 인물의 어깨가 크게 들어와 있음",
     "severity": "major"
    },
    {
     "issue_ko": "목에 핏대가 서 있지 않음",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "목에 핏대가 서 있지 않음",
    "fix_en": "Add prominent, bulging veins to the character's strained neck, preserving his face, open mouth, expression, suit, and the background.",
    "severity": "minor",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the out-of-focus person's shoulder in the left foreground entirely, replacing that area with the out-of-focus background wall and framed calligraphy, preserving the main character, his expression, suit, and the rest of the office setting.\n- Blur the desk, nameplate, and lower torso to simulate a tighter focus on the character's face, preserving the character's facial expression, lighting, and the existing camera angle.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시대로 프레임에서 부하 직원을 성공적으로 제외하여 대상을 온전히 오프스크린으로 처리했으며, 목에 핏대를 세우고 고함치는 표정을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "부하 직원을 클로즈업 프레임에서 제외하라는 명시적인 지시에도 불구하고 여전히 왼쪽 화면에 부하 직원의 어깨가 크게 남아있어 프레임 구성에 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "차장검사의 시선과 고개가 화면 왼쪽 바깥(오프스크린)에 위치한 대상을 향하고 있습니다.",
      "built_space": "사무실 내부이며, 책상과 명패가 앞에 있고 뒤쪽 벽에는 액자와 화면이 정상적으로 배치되어 있습니다.",
      "entities": "차장검사는 레퍼런스와 동일한 외모와 복장(스트라이프 정장, 넥타이)을 갖추고 있으며, 부하 직원의 모습은 화면에서 완전히 제거되었습니다.",
      "hard_violations": [],
      "physics": "인물은 의자에 앉아 책상에 지지된 안정적인 자세를 취하고 있습니다."
     },
     {
      "label": "A",
      "direction": "차장검사의 시선과 고개가 화면 왼쪽 바깥을 향하고 있습니다.",
      "built_space": "사무실 내부이며, 책상, 명패, 액자, 뒤쪽 화면 등 배경 요소가 제대로 배치되어 있습니다.",
      "entities": "차장검사의 외모와 복장은 레퍼런스와 일치하나, 제외해야 할 부하 직원의 어깨가 화면 왼쪽을 크게 차지하고 있습니다.",
      "hard_violations": [],
      "physics": "인물은 책상을 앞에 두고 의자에 정상적으로 앉아 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시대로 프레임에서 부하 직원을 성공적으로 제외하여 대상을 온전히 오프스크린으로 처리했으며, 목에 핏대를 세우고 고함치는 표정을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "부하 직원을 클로즈업 프레임에서 제외하라는 명시적인 지시에도 불구하고 여전히 왼쪽 화면에 부하 직원의 어깨가 크게 남아있어 프레임 구성에 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "차장검사의 시선과 고개가 화면 왼쪽 바깥(오프스크린)에 위치한 대상을 향하고 있습니다.",
      "built_space": "사무실 내부이며, 책상과 명패가 앞에 있고 뒤쪽 벽에는 액자와 화면이 정상적으로 배치되어 있습니다.",
      "entities": "차장검사는 레퍼런스와 동일한 외모와 복장(스트라이프 정장, 넥타이)을 갖추고 있으며, 부하 직원의 모습은 화면에서 완전히 제거되었습니다.",
      "hard_violations": [],
      "physics": "인물은 의자에 앉아 책상에 지지된 안정적인 자세를 취하고 있습니다."
     },
     {
      "label": "A",
      "direction": "차장검사의 시선과 고개가 화면 왼쪽 바깥을 향하고 있습니다.",
      "built_space": "사무실 내부이며, 책상, 명패, 액자, 뒤쪽 화면 등 배경 요소가 제대로 배치되어 있습니다.",
      "entities": "차장검사의 외모와 복장은 레퍼런스와 일치하나, 제외해야 할 부하 직원의 어깨가 화면 왼쪽을 크게 차지하고 있습니다.",
      "hard_violations": [],
      "physics": "인물은 책상을 앞에 두고 의자에 정상적으로 앉아 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "부하 검사를 프레임에서 제외하라는 명시적 지시를 충실히 이행했으며, 핏대가 선 목과 크게 벌린 입 등 요구된 표정 연출을 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "차장검사의 표정 연출은 좋으나, 지시문에서 명시적으로 배제하라고 한 이전 샷의 인물(부하 검사의 등/어깨)을 화면 좌측에 그대로 남겨두어 핵심 조건을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 프레임 좌측 바깥(장원섭이 있는 방향)을 향하고 있습니다.",
      "built_space": "사무실 책상 앞에 앉아 있으며, 배경의 서예 액자와 TV 화면 배치가 레퍼런스와 정확히 일치합니다.",
      "entities": "차장검사의 얼굴, 헤어스타일, 스트라이프 정장이 레퍼런스와 일치하며, 지시대로 부하 검사는 화면에 등장하지 않습니다.",
      "hard_violations": [],
      "physics": "의자에 앉아 상체를 안정적으로 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "시선은 프레임 좌측 바깥을 향하고 있습니다.",
      "built_space": "사무실 책상 앞에 앉아 있으며, 배경 요소가 레퍼런스와 일치합니다.",
      "entities": "차장검사의 외양은 일치하나, 화면에 등장해서는 안 되는 부하 검사의 어깨가 좌측에 나타납니다.",
      "hard_violations": [
       "프레임에서 제외하라고 명시된 인물(부하 검사의 어깨 및 등)이 화면 좌측에 포함됨"
      ],
      "physics": "의자에 앉아 상체를 안정적으로 지탱하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "부하 검사를 프레임에서 제외하라는 명시적 지시를 충실히 이행했으며, 핏대가 선 목과 크게 벌린 입 등 요구된 표정 연출을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "차장검사의 표정 연출은 좋으나, 지시문에서 명시적으로 배제하라고 한 이전 샷의 인물(부하 검사의 등/어깨)을 화면 좌측에 그대로 남겨두어 핵심 조건을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 프레임 좌측 바깥(장원섭이 있는 방향)을 향하고 있습니다.",
      "built_space": "사무실 책상 앞에 앉아 있으며, 배경의 서예 액자와 TV 화면 배치가 레퍼런스와 정확히 일치합니다.",
      "entities": "차장검사의 얼굴, 헤어스타일, 스트라이프 정장이 레퍼런스와 일치하며, 지시대로 부하 검사는 화면에 등장하지 않습니다.",
      "hard_violations": [],
      "physics": "의자에 앉아 상체를 안정적으로 지탱하고 있습니다."
     },
     {
      "label": "A",
      "direction": "시선은 프레임 좌측 바깥을 향하고 있습니다.",
      "built_space": "사무실 책상 앞에 앉아 있으며, 배경 요소가 레퍼런스와 일치합니다.",
      "entities": "차장검사의 외양은 일치하나, 화면에 등장해서는 안 되는 부하 검사의 어깨가 좌측에 나타납니다.",
      "hard_violations": [
       "프레임에서 제외하라고 명시된 인물(부하 검사의 어깨 및 등)이 화면 좌측에 포함됨"
      ],
      "physics": "의자에 앉아 상체를 안정적으로 지탱하고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S59sh4"
  }
 },
 "S61sh3::cine": {
  "applied": true,
  "fingerprint": "2537ed1170aedd713d856d6e7af784b204d1ffbbb74ccb8786b7c8275c3f54a2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S61sh3_sel.png",
  "source_sha256": "1f96ed5e8bc18fe68bee9c1ee4b78a84495050a15887ae664f1b72de41089e75",
  "file": "S61sh3_cine.png",
  "latency_ms": 13908
 },
 "S61sh4::signage": {
  "fp": "8e3e166991e76ba2",
  "inscriptions": []
 },
 "S61sh4": {
  "input_fingerprint": "fd6b3dda88e23f65",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입을 굳게 다문 채 매섭게 쏘아보는 장원섭의 측면\n\nLOCATION (lock): Across from the deputy chief prosecutor in his office meeting area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the axis cut, hold statically close beside 장원섭 at slightly below his seated eye line, presenting a taut profile against the off-screen position of 차장검사. 장원섭 occupies the right two-thirds, lips sealed and jaw locked, with his unwavering gaze traveling leftward out of frame; the pause is allowed to outlast the final words.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office chair back (occupied by 장원섭) — Its upper edge is seen obliquely behind 장원섭's shoulder; used as A restrained background contour behind 장원섭 maintains the office setting without interrupting his profile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime office illumination keeps the profile naturally modeled in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office background, desk arrangement, daylight, and the prosecutor's suit. Exclude the chief prosecutor from the close frame and isolate the subordinate's hard stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입을 굳게 다문 채 매섭게 쏘아보는 장원섭의 측면\n\nLOCATION (lock): Across from the deputy chief prosecutor in his office meeting area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the axis cut, hold statically close beside 장원섭 at slightly below his seated eye line, presenting a taut profile against the off-screen position of 차장검사. 장원섭 occupies the right two-thirds, lips sealed and jaw locked, with his unwavering gaze traveling leftward out of frame; the pause is allowed to outlast the final words.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office chair back (occupied by 장원섭) — Its upper edge is seen obliquely behind 장원섭's shoulder; used as A restrained background contour behind 장원섭 maintains the office setting without interrupting his profile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime office illumination keeps the profile naturally modeled in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office background, desk arrangement, daylight, and the prosecutor's suit. Exclude the chief prosecutor from the close frame and isolate the subordinate's hard stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 입을 굳게 다문 채 매섭게 쏘아보는 장원섭의 측면\n\nLOCATION (lock): Across from the deputy chief prosecutor in his office meeting area. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: After the axis cut, hold statically close beside 장원섭 at slightly below his seated eye line, presenting a taut profile against the off-screen position of 차장검사. 장원섭 occupies the right two-thirds, lips sealed and jaw locked, with his unwavering gaze traveling leftward out of frame; the pause is allowed to outlast the final words.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: office chair back (occupied by 장원섭) — Its upper edge is seen obliquely behind 장원섭's shoulder; used as A restrained background contour behind 장원섭 maintains the office setting without interrupting his profile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime office illumination keeps the profile naturally modeled in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office background, desk arrangement, daylight, and the prosecutor's suit. Exclude the chief prosecutor from the close frame and isolate the subordinate's hard stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "인물이 화면 왼쪽 밖을 향해 굳은 표정으로 시선을 두고 있음.",
    "built_space": "사무실 배경으로 벽면의 서예 액자, TV 화면, 그리고 인물 어깨 뒤로 의자 등받이가 보임.",
    "entities": "캐릭터 레퍼런스에 제시된 장원섭의 얼굴, 헤어스타일, 회색 정장이 정확히 일치하게 렌더링됨.",
    "hard_violations": [],
    "physics": "인물이 의자에 안정적으로 기대어 앉아 있음."
   },
   {
    "label": "A",
    "direction": "인물이 화면 왼쪽 밖을 향해 매섭게 시선을 두고 있음.",
    "built_space": "사무실 배경으로 서예 액자, 블라인드, 그리고 화면 전경에 이전 샷의 명패가 보임.",
    "entities": "장원섭 레퍼런스가 아닌 이전 샷 스틸의 인물(얼굴, 헤어스타일)과 짙은 줄무늬 정장이 나타남.",
    "hard_violations": [
     "지정된 인물(장원섭)이 아닌 이전 샷의 인물을 그대로 렌더링함 (Wrong identity)."
    ],
    "physics": "인물이 의자에 앉아 자세를 유지하고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 캐릭터 레퍼런스를 정확히 반영하여 인물의 측면 앵글과 시선 방향을 성공적으로 연출함."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "이전 샷의 인물 얼굴을 그대로 가져와 명시된 '장원섭' 캐릭터 지시를 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S61sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 속 남성의 얼굴·헤어·이목구비가 장원섭 캐릭터 레퍼런스와 다른 사람이다.",
     "fix_en": "Replace the man's face and hair to match the character reference for Jang Won-seop, featuring a slimmer face and sharper jawline. Preserve the man's position, his grey suit, light blue shirt, dotted tie, the office background, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "입을 굳게 다문 매서운 노려봄이 아니라 입이 느슨하고 시선이 약한 측면 표정이다.",
     "fix_en": "Adjust the man's facial expression so his lips are pressed tightly together and his jaw is visibly locked in a fierce glare. Preserve the man's identity, position, clothing, the office background, the lighting, and the framing.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 속 남성의 얼굴·헤어·이목구비가 장원섭 캐릭터 레퍼런스와 다른 사람이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "입을 굳게 다문 매서운 노려봄이 아니라 입이 느슨하고 시선이 약한 측면 표정이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 0,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "입을 굳게 다문 매서운 노려봄이 아니라 입이 느슨하고 시선이 약한 측면 표정이다.",
    "fix_en": "Adjust the man's facial expression so his lips are pressed tightly together and his jaw is visibly locked in a fierce glare. Preserve the man's identity, position, clothing, the office background, the lighting, and the framing.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the man's face and hair to match the character reference for Jang Won-seop, featuring a slimmer face and sharper jawline. Preserve the man's position, his grey suit, light blue shirt, dotted tie, the office background, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "프롬프트가 요구한 '측면' 구도에 더 부합하며, 미간을 찌푸린 연기를 통해 '매섭게 쏘아보는' 표정을 매우 충실하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "시선 방향은 맞으나 완벽한 측면이 아닌 반측면에 가깝고, 표정이 다소 평온하여 '매섭게 쏘아보는' 지시를 충족하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 화면 왼쪽 프레임 밖을 향하고 있습니다.",
      "built_space": "사무실 배경으로, 인물의 오른쪽 어깨 뒤로 가죽 의자의 등받이가 위치해 있으며, 뒤쪽 벽면에 액자와 화면이 보입니다.",
      "entities": "장원섭(40대 초반 한국 남성)이 캐릭터 레퍼런스와 일치하는 짙은 회색 정장과 도트 무늬 넥타이를 착용하고 있습니다.",
      "hard_violations": [],
      "physics": "인물이 의자에 앉아 있는 자세가 자연스럽게 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 화면 왼쪽 프레임 밖을 향하고 있습니다.",
      "built_space": "사무실 배경으로, 인물의 오른쪽 어깨 뒤로 가죽 의자의 등받이가 위치해 있으며, 뒤쪽 벽면에 액자와 화면이 보입니다.",
      "entities": "장원섭(40대 초반 한국 남성)이 캐릭터 레퍼런스와 일치하는 짙은 회색 정장과 도트 무늬 넥타이를 착용하고 있습니다.",
      "hard_violations": [],
      "physics": "인물이 의자에 앉아 있는 자세가 자연스럽게 지탱되고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "프롬프트가 요구한 '측면' 구도에 더 부합하며, 미간을 찌푸린 연기를 통해 '매섭게 쏘아보는' 표정을 매우 충실하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "시선 방향은 맞으나 완벽한 측면이 아닌 반측면에 가깝고, 표정이 다소 평온하여 '매섭게 쏘아보는' 지시를 충족하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 화면 왼쪽 프레임 밖을 향하고 있습니다.",
      "built_space": "사무실 배경으로, 인물의 오른쪽 어깨 뒤로 가죽 의자의 등받이가 위치해 있으며, 뒤쪽 벽면에 액자와 화면이 보입니다.",
      "entities": "장원섭(40대 초반 한국 남성)이 캐릭터 레퍼런스와 일치하는 짙은 회색 정장과 도트 무늬 넥타이를 착용하고 있습니다.",
      "hard_violations": [],
      "physics": "인물이 의자에 앉아 있는 자세가 자연스럽게 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 화면 왼쪽 프레임 밖을 향하고 있습니다.",
      "built_space": "사무실 배경으로, 인물의 오른쪽 어깨 뒤로 가죽 의자의 등받이가 위치해 있으며, 뒤쪽 벽면에 액자와 화면이 보입니다.",
      "entities": "장원섭(40대 초반 한국 남성)이 캐릭터 레퍼런스와 일치하는 짙은 회색 정장과 도트 무늬 넥타이를 착용하고 있습니다.",
      "hard_violations": [],
      "physics": "인물이 의자에 앉아 있는 자세가 자연스럽게 지탱되고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "명시된 '측면(profile)' 구도를 정확히 구현했으며, 미간을 찌푸리고 턱을 앙다문 표정으로 '매섭게 쏘아보는' 감정선까지 훌륭하게 표현하여 프롬프트 지시를 완벽히 따랐습니다."
     },
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "요구된 측면 구도라기보다는 반측면(3/4 앵글)에 가깝게 얼굴이 렌더링되었으며, 표정 또한 상대적으로 부드러워 매섭게 쏘아보는 느낌이 다소 부족합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "남자의 시선이 프레임 왼쪽 밖의 보이지 않는 대상을 향해 고정되어 있음.",
      "built_space": "사무실 배경. 인물의 등 뒤로 갈색 가죽 의자의 상단이 보이고, 배경 좌측에는 붓글씨 액자가, 우측에는 인물들이 띄워진 화면이 배치됨.",
      "entities": "장원섭의 얼굴, 헤어스타일, 회색 정장 및 점무늬 넥타이가 캐릭터 레퍼런스와 일치함. 40대 초반 남성의 외양이 잘 반영됨.",
      "hard_violations": [],
      "physics": "남자는 의자에 자연스럽게 앉아있으며 신체나 의복에 중력과 지지가 정상적으로 작용함."
     },
     {
      "label": "B",
      "direction": "남자의 시선이 프레임 왼쪽 밖을 향해 매섭게 고정되어 있음.",
      "built_space": "사무실 배경. 인물의 등 뒤로 갈색 가죽 의자 등받이가 올바르게 위치하며, 좌측 붓글씨 액자와 우측의 화면 등 이전 샷의 요소들이 잘 유지됨.",
      "entities": "장원섭의 얼굴과 의상(회색 정장, 점무늬 넥타이)이 레퍼런스와 일치하며, 지시된 '측면'의 특징이 매우 정확하게 나타남.",
      "hard_violations": [],
      "physics": "남자는 의자에 무게를 싣고 앉아있으며, 어색함 없이 안정적인 자세를 유지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "명시된 '측면(profile)' 구도를 정확히 구현했으며, 미간을 찌푸리고 턱을 앙다문 표정으로 '매섭게 쏘아보는' 감정선까지 훌륭하게 표현하여 프롬프트 지시를 완벽히 따랐습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "요구된 측면 구도라기보다는 반측면(3/4 앵글)에 가깝게 얼굴이 렌더링되었으며, 표정 또한 상대적으로 부드러워 매섭게 쏘아보는 느낌이 다소 부족합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "남자의 시선이 프레임 왼쪽 밖의 보이지 않는 대상을 향해 고정되어 있음.",
      "built_space": "사무실 배경. 인물의 등 뒤로 갈색 가죽 의자의 상단이 보이고, 배경 좌측에는 붓글씨 액자가, 우측에는 인물들이 띄워진 화면이 배치됨.",
      "entities": "장원섭의 얼굴, 헤어스타일, 회색 정장 및 점무늬 넥타이가 캐릭터 레퍼런스와 일치함. 40대 초반 남성의 외양이 잘 반영됨.",
      "hard_violations": [],
      "physics": "남자는 의자에 자연스럽게 앉아있으며 신체나 의복에 중력과 지지가 정상적으로 작용함."
     },
     {
      "label": "A",
      "direction": "남자의 시선이 프레임 왼쪽 밖을 향해 매섭게 고정되어 있음.",
      "built_space": "사무실 배경. 인물의 등 뒤로 갈색 가죽 의자 등받이가 올바르게 위치하며, 좌측 붓글씨 액자와 우측의 화면 등 이전 샷의 요소들이 잘 유지됨.",
      "entities": "장원섭의 얼굴과 의상(회색 정장, 점무늬 넥타이)이 레퍼런스와 일치하며, 지시된 '측면'의 특징이 매우 정확하게 나타남.",
      "hard_violations": [],
      "physics": "남자는 의자에 무게를 싣고 앉아있으며, 어색함 없이 안정적인 자세를 유지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 14
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S61sh3"
  }
 },
 "S61sh4::cine": {
  "applied": true,
  "fingerprint": "2c57f2028aa8e811b85d0e0487f2e8f226f4cd19f83786212dd218e4d78249ac",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S61sh4_sel.png",
  "source_sha256": "d86f1be8611f27c839d6ee08eefbb8a188bc5e05d3e84b16d57b7de46862934d",
  "file": "S61sh4_cine.png",
  "latency_ms": 11421
 },
 "S62sh2::confined_fp_apt": {
  "applies": false,
  "reason_ko": "엘리베이터 내부는 복잡한 좌석 배치나 조종석이 있는 공간이 아니며, 단순히 닫히는 문 사이로 들어오는 인물을 묘사하는 숏이므로 평면도 레이아웃 보조가 필요하지 않음.",
  "input_fingerprint": "d26f91b74d690a82"
 },
 "S62sh2::signage": {
  "fp": "11321a247e16280c",
  "inscriptions": [
   {
    "surface_native": "엘리베이터 문 안전 스티커",
    "text_native": "손끼임 주의",
    "reason_ko": "닫히는 엘리베이터 문에 몸을 들이미는 다급한 상황에서 문 틈새에 부착된 실제 한국의 안전 경고 스티커가 장면의 현실성을 높여줍니다."
   }
  ]
 },
 "era_assess::cb36982750678aa8": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 관공서 및 검찰청 엘리베이터 내부",
    "search_terms_native": [
     "검찰청 엘리베이터 내부",
     "관공서 엘리베이터 버튼",
     "현대엘리베이터 2000년대",
     "엘리베이터 조작반 닫힘"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며 다른 언어로 번역하거나 결합하지 마십시오.",
    "reason_ko": "2000년대 한국 관공서 엘리베이터 내부의 특유의 스테인리스 스틸 마감, 열림/닫힘 한글 버튼 디자인, 상단 LED 층수 표시기는 범용 이미지 모델이 재현하기 어렵습니다."
   }
  ]
 },
 "era_ref::4ff8a73bb7acda72": {
  "subject": "2000년대~2010년대 한국 관공서 및 검찰청 엘리베이터 내부",
  "terms": [
   "검찰청 엘리베이터 내부",
   "관공서 엘리베이터 버튼",
   "현대엘리베이터 2000년대",
   "엘리베이터 조작반 닫힘"
  ],
  "queries": [
   [
    "검찰청 엘리베이터 내부 관공서 엘리베이터 버튼 현대엘리베이터 2000년대 조작반 닫힘",
    "2000년대 2010년대 한국 관공서 검찰청 엘리베이터 내부 버튼 조작반"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDA1MDZfMTA0/MDAxNzE0OTYxNTYwMzMw.TnYPiO3I78oU9khps9-EO1plcMKe2JWSiO_TMICjbZEg.OLOeMtFPox70USqVfATy9lnEyJAWiaDjeS3I3bX-8P4g.JPEG/KakaoTalk_20240506_111207055_01.jpg?type=w400",
  "picked_reason_ko": "4번은 전체 실내가 아닌 조작반 중심의 사진이지만, 상업광고나 차량용 설비 없이 한국 관공서에서 볼 법한 소박한 스테인리스 조작반과 버튼 구성을 가장 명료하게 보여준다.",
  "sha256": "e7c964d4ae48c6c18937bfdc95acf1bd5f631b852480429c1845ecd44c1f24f6",
  "file": "eraref_4ff8a73bb7acda72.png"
 },
 "S62sh2::bgfirst_bg": {
  "input_fingerprint": "c43748ec57e7e67d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 닫히려는 엘리베이터 문 사이로 다급하게 상체를 들이민 장원섭의 모습\n\nLOCATION (lock): Inside the compact prosecution-building elevator at its closing doorway, with doors, control panel, and fixed standing positions close together.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height just outside and to one side of the elevator entrance, frame 장원섭 tightly in mid-lunge through the narrowing diagonal opening while keeping both converging door edges visible. His upper body occupies the middle-left as one foot drives him inward, while 차장검사 remains deeper inside at the right, caught looking down as the interruption reaches him.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, moves toward elevator interior; narrowing elevator doorway in the middle-center of the frame, midground; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator door panels (closing) — Their front faces angle toward the center seam as they close around the entrance; used as The two edges form a converging frame around 장원섭's urgent entry; elevator interior (occupied) — Viewed diagonally through the narrowing entrance; used as Provides the deeper spatial plane holding 차장검사 beyond 장원섭.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination is rendered with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 관공서 및 검찰청 엘리베이터 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 닫히려는 엘리베이터 문 사이로 다급하게 상체를 들이민 장원섭의 모습\n\nLOCATION (lock): Inside the compact prosecution-building elevator at its closing doorway, with doors, control panel, and fixed standing positions close together.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From chest height just outside and to one side of the elevator entrance, frame 장원섭 tightly in mid-lunge through the narrowing diagonal opening while keeping both converging door edges visible. His upper body occupies the middle-left as one foot drives him inward, while 차장검사 remains deeper inside at the right, caught looking down as the interruption reaches him.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, moves toward elevator interior; narrowing elevator doorway in the middle-center of the frame, midground; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator door panels (closing) — Their front faces angle toward the center seam as they close around the entrance; used as The two edges form a converging frame around 장원섭's urgent entry; elevator interior (occupied) — Viewed diagonally through the narrowing entrance; used as Provides the deeper spatial plane holding 차장검사 beyond 장원섭.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination is rendered with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 관공서 및 검찰청 엘리베이터 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh2__bgfirst_bg.png",
  "asset_id": "f5fac2f7-846d-4fff-8ee7-8161731a0ba5",
  "input_asset_ids": [
   "75c285ae-cc8f-430c-9b41-a0738b9f55af",
   "3d62c13f-094e-4df1-92c2-bab2967d17b9"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국 관공서 및 검찰청 엘리베이터 내부",
   "queries": [
    [
     "검찰청 엘리베이터 내부 관공서 엘리베이터 버튼 현대엘리베이터 2000년대 조작반 닫힘",
     "2000년대 2010년대 한국 관공서 검찰청 엘리베이터 내부 버튼 조작반"
    ]
   ],
   "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDA1MDZfMTA0/MDAxNzE0OTYxNTYwMzMw.TnYPiO3I78oU9khps9-EO1plcMKe2JWSiO_TMICjbZEg.OLOeMtFPox70USqVfATy9lnEyJAWiaDjeS3I3bX-8P4g.JPEG/KakaoTalk_20240506_111207055_01.jpg?type=w400",
   "sha256": "e7c964d4ae48c6c18937bfdc95acf1bd5f631b852480429c1845ecd44c1f24f6",
   "file": "eraref_4ff8a73bb7acda72.png"
  }
 },
 "S62sh2": {
  "input_fingerprint": "184b9b01c078ac28",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 닫히려는 엘리베이터 문 사이로 다급하게 상체를 들이민 장원섭의 모습\n\nLOCATION (lock): Inside the compact prosecution-building elevator at its closing doorway, with doors, control panel, and fixed standing positions close together. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height just outside and to one side of the elevator entrance, frame 장원섭 tightly in mid-lunge through the narrowing diagonal opening while keeping both converging door edges visible. His upper body occupies the middle-left as one foot drives him inward, while 차장검사 remains deeper inside at the right, caught looking down as the interruption reaches him.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, moves toward elevator interior; narrowing elevator doorway in the middle-center of the frame, midground; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator door panels (closing) — Their front faces angle toward the center seam as they close around the entrance; used as The two edges form a converging frame around 장원섭's urgent entry; elevator interior (occupied) — Viewed diagonally through the narrowing entrance; used as Provides the deeper spatial plane holding 차장검사 beyond 장원섭.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination is rendered with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 엘리베이터 문 안전 스티커: \"손끼임 주의\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective, every panel, control, seat, mirror, window and fixture stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: how many steering wheels or control surfaces, how many seats and which way each faces, where each mirror sits and what it could reflect from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second rim, sliding a seat, turning a mirror or growing a new panel is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: which seat the body occupies, which side of the wheel the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 닫히려는 엘리베이터 문 사이로 다급하게 상체를 들이민 장원섭의 모습\n\nLOCATION (lock): Inside the compact prosecution-building elevator at its closing doorway, with doors, control panel, and fixed standing positions close together. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height just outside and to one side of the elevator entrance, frame 장원섭 tightly in mid-lunge through the narrowing diagonal opening while keeping both converging door edges visible. His upper body occupies the middle-left as one foot drives him inward, while 차장검사 remains deeper inside at the right, caught looking down as the interruption reaches him.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, moves toward elevator interior; narrowing elevator doorway in the middle-center of the frame, midground; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator door panels (closing) — Their front faces angle toward the center seam as they close around the entrance; used as The two edges form a converging frame around 장원섭's urgent entry; elevator interior (occupied) — Viewed diagonally through the narrowing entrance; used as Provides the deeper spatial plane holding 차장검사 beyond 장원섭.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination is rendered with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 엘리베이터 문 안전 스티커: \"손끼임 주의\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 닫히려는 엘리베이터 문 사이로 다급하게 상체를 들이민 장원섭의 모습\n\nLOCATION (lock): Inside the compact prosecution-building elevator at its closing doorway, with doors, control panel, and fixed standing positions close together. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height just outside and to one side of the elevator entrance, frame 장원섭 tightly in mid-lunge through the narrowing diagonal opening while keeping both converging door edges visible. His upper body occupies the middle-left as one foot drives him inward, while 차장검사 remains deeper inside at the right, caught looking down as the interruption reaches him.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, moves toward elevator interior; narrowing elevator doorway in the middle-center of the frame, midground; 차장검사 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator door panels (closing) — Their front faces angle toward the center seam as they close around the entrance; used as The two edges form a converging frame around 장원섭's urgent entry; elevator interior (occupied) — Viewed diagonally through the narrowing entrance; used as Provides the deeper spatial plane holding 차장검사 beyond 장원섭.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination is rendered with restrained neutral-to-cool color and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 엘리베이터 문 안전 스티커: \"손끼임 주의\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched panels, seats, controls and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S62sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L57B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1321,
      "verdict_ko": "다급한 돌진 동작과 텍스트는 훌륭하나, 카메라 위치가 엘리베이터 내부로 완전히 뒤바뀌는 심각한 공간적 오류가 있습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "돌진하는 역동성이 부족하며, 카메라 위치가 엘리베이터 내부로 역전되었고 스티커 위치도 잘못되었습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.571
     },
     "adjusted": {
      "A": 1.5,
      "B": 1.321
     },
     "violations": {
      "B": [
       "[gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
      ],
      "A": [
       "[gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1321,
      "verdict_ko": "다급한 돌진 동작과 텍스트는 훌륭하나, 카메라 위치가 엘리베이터 내부로 완전히 뒤바뀌는 심각한 공간적 오류가 있습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "돌진하는 역동성이 부족하며, 카메라 위치가 엘리베이터 내부로 역전되었고 스티커 위치도 잘못되었습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 850,
      "verdict_ko": "카메라 시점이 지시와 반대로 적용되었으며, 전경과 배경에 동일한 얼굴의 인물이 중복 생성되는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 카메라 시점 반전 (내부에서 외부 복도를 렌더링함) / [gemini-pro] 인물 복제 및 중복 (배경 인물이 전경 인물과 동일하게 생성됨) / [openrouter:x-ai/grok-4.6] 약정된 문 밖 시점이 아니라 내부측에서 로비를 찍고 차장검사를 엘리베이터 안이 아닌 창 있는 로비에 둠"
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "카메라 위치가 지정된 방향과 정반대이며, 전경의 주인공 얼굴이 지정된 레퍼런스와 완전히 다릅니다.  ★위반: [gemini-pro] 카메라 시점 및 배경 반전 (외부에서 내부를 봐야 하나, 내부에서 외부 복도를 렌더링함) / [gemini-pro] 지정된 인물 신원 위반 (전경 피사체가 지정된 레퍼런스와 다름) / [openrouter:x-ai/grok-4.6] 약정된 문 밖 시점이 아니라 내부측에서 로비를 찍고 차장검사를 엘리베이터 안이 아닌 창 있는 로비에 둠"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 2.0,
      "B": 1.6
     },
     "adjusted": {
      "A": 1.25,
      "B": 0.85
     },
     "violations": {
      "A": [
       "[gemini-pro] 카메라 시점 및 배경 반전 (외부에서 내부를 봐야 하나, 내부에서 외부 복도를 렌더링함)",
       "[gemini-pro] 지정된 인물 신원 위반 (전경 피사체가 지정된 레퍼런스와 다름)",
       "[openrouter:x-ai/grok-4.6] 약정된 문 밖 시점이 아니라 내부측에서 로비를 찍고 차장검사를 엘리베이터 안이 아닌 창 있는 로비에 둠"
      ],
      "B": [
       "[gemini-pro] 카메라 시점 반전 (내부에서 외부 복도를 렌더링함)",
       "[gemini-pro] 인물 복제 및 중복 (배경 인물이 전경 인물과 동일하게 생성됨)",
       "[openrouter:x-ai/grok-4.6] 약정된 문 밖 시점이 아니라 내부측에서 로비를 찍고 차장검사를 엘리베이터 안이 아닌 창 있는 로비에 둠"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.4,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 850,
      "verdict_ko": "카메라 시점이 지시와 반대로 적용되었으며, 전경과 배경에 동일한 얼굴의 인물이 중복 생성되는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 카메라 시점 반전 (내부에서 외부 복도를 렌더링함) / [gemini-pro] 인물 복제 및 중복 (배경 인물이 전경 인물과 동일하게 생성됨) / [openrouter:x-ai/grok-4.6] 약정된 문 밖 시점이 아니라 내부측에서 로비를 찍고 차장검사를 엘리베이터 안이 아닌 창 있는 로비에 둠"
     },
     {
      "label": "B",
      "score": 1250,
      "verdict_ko": "카메라 위치가 지정된 방향과 정반대이며, 전경의 주인공 얼굴이 지정된 레퍼런스와 완전히 다릅니다.  ★위반: [gemini-pro] 카메라 시점 및 배경 반전 (외부에서 내부를 봐야 하나, 내부에서 외부 복도를 렌더링함) / [gemini-pro] 지정된 인물 신원 위반 (전경 피사체가 지정된 레퍼런스와 다름) / [openrouter:x-ai/grok-4.6] 약정된 문 밖 시점이 아니라 내부측에서 로비를 찍고 차장검사를 엘리베이터 안이 아닌 창 있는 로비에 둠"
     }
    ],
    "all_candidates_fail": true
   },
   "combined": {
    "totals": {
     "A": 2350,
     "B": 2571
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 2350,
   "B": 2571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "다급한 돌진 동작과 텍스트는 훌륭하나, 카메라 위치가 엘리베이터 내부로 완전히 뒤바뀌는 심각한 공간적 오류가 있습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
   },
   {
    "label": "A",
    "score": 1500,
    "verdict_ko": "돌진하는 역동성이 부족하며, 카메라 위치가 엘리베이터 내부로 역전되었고 스티커 위치도 잘못되었습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(외부에서 내부로)를 위반하고 내부에서 외부 복도를 바라보는 시점으로 렌더링함"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L57B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트 상 전경에서 문틈으로 뛰어들어야 할 '장원섭'의 얼굴이 다른 인물로 렌더링되었고, 오히려 배경에 있어야 할 '차장검사'의 얼굴이 장원섭 레퍼런스와 일치하게 렌더링되어 캐릭터 배역이 완전히 뒤바뀜.",
     "fix_en": "Replace the foreground man's face with the reference face for 장원섭, and change the background man's face to a different Korean male. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "로케이션 레퍼런스 상 창문이 있는 곳은 엘리베이터 외부 복도이나, 카메라가 외부에서 내부를 바라보는 구도임에도 배경(엘리베이터 내부)에 복도의 창문 공간이 렌더링되어 공간 구조가 안팎으로 뒤집힘.",
     "fix_en": "Replace the background corridor and window with solid, windowless metal elevator interior walls. Preserve the people present and their positions, their clothing, the foreground doors, the light, and the framing.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "왼쪽 엘리베이터 문에 적힌 '손끼임 주의' 표지판이 물리적 질감을 가진 실제 스티커가 아닌, 화면 위에 단순히 합성된 평면적인 디지털 그래픽 오버레이처럼 렌더링됨.",
     "fix_en": "Apply physical surface texture and edge thickness to the yellow sign so it reads as a real sticker on the metal. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "왼쪽 전경의 장원섭이 안으로 들이밀지 않고 얼굴과 시선을 카메라(렌즈) 쪽으로 돌린 채 문 밖/복도 쪽으로 몸을 던지고 있다.",
     "fix_en": "Turn the foreground man's head and gaze inward toward the gap between the doors rather than at the camera. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "닫히는 양쪽 문 가장자리가 몸을 조이는 좁은 수렴 틈이 아니고 개구부가 넓게 열려 있다.",
     "fix_en": "Extend the left and right door panels further inward to create a narrower, tightly converging gap around the foreground man. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트 상 전경에서 문틈으로 뛰어들어야 할 '장원섭'의 얼굴이 다른 인물로 렌더링되었고, 오히려 배경에 있어야 할 '차장검사'의 얼굴이 장원섭 레퍼런스와 일치하게 렌더링되어 캐릭터 배역이 완전히 뒤바뀜.",
     "severity": "critical"
    },
    {
     "issue_ko": "로케이션 레퍼런스 상 창문이 있는 곳은 엘리베이터 외부 복도이나, 카메라가 외부에서 내부를 바라보는 구도임에도 배경(엘리베이터 내부)에 복도의 창문 공간이 렌더링되어 공간 구조가 안팎으로 뒤집힘.",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 엘리베이터 문에 적힌 '손끼임 주의' 표지판이 물리적 질감을 가진 실제 스티커가 아닌, 화면 위에 단순히 합성된 평면적인 디지털 그래픽 오버레이처럼 렌더링됨.",
     "severity": "major"
    },
    {
     "issue_ko": "문 너머 중앙~우측이 엘리베이터 내부가 아니라 창문이 있는 복도이고, 차장검사가 엘리베이터 안이 아닌 그 복도에 서 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 전경의 장원섭이 안으로 들이밀지 않고 얼굴과 시선을 카메라(렌즈) 쪽으로 돌린 채 문 밖/복도 쪽으로 몸을 던지고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "전경 장원섭의 얼굴이 제공된 인물 레퍼런스와 다른 사람이다.",
     "severity": "major"
    },
    {
     "issue_ko": "닫히는 양쪽 문 가장자리가 몸을 조이는 좁은 수렴 틈이 아니고 개구부가 넓게 열려 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 엘리베이터 문에 적힌 '손끼임 주의' 표지판이 물리적 질감을 가진 실제 스티커가 아닌, 화면 위에 단순히 합성된 평면적인 디지털 그래픽 오버레이처럼 렌더링됨.",
    "fix_en": "Apply physical surface texture and edge thickness to the yellow sign so it reads as a real sticker on the metal. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "왼쪽 전경의 장원섭이 안으로 들이밀지 않고 얼굴과 시선을 카메라(렌즈) 쪽으로 돌린 채 문 밖/복도 쪽으로 몸을 던지고 있다.",
    "fix_en": "Turn the foreground man's head and gaze inward toward the gap between the doors rather than at the camera. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "닫히는 양쪽 문 가장자리가 몸을 조이는 좁은 수렴 틈이 아니고 개구부가 넓게 열려 있다.",
    "fix_en": "Extend the left and right door panels further inward to create a narrower, tightly converging gap around the foreground man. Preserve the people present and their positions, their clothing, the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the foreground man's face with the reference face for 장원섭, and change the background man's face to a different Korean male. Preserve the people present and their positions, their clothing, the set, the light, and the framing.\n- Replace the background corridor and window with solid, windowless metal elevator interior walls. Preserve the people present and their positions, their clothing, the foreground doors, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 583,
      "verdict_ko": "카메라 시점이 반대로 적용되고 캐릭터 정체성이 뒤바뀌는 치명적 오류가 있으나, 배경 인물의 시선 지시와 복도 공간 묘사는 상대적으로 정확합니다.  ★위반: [gemini-pro] 지정된 카메라 위치 위반 (외부가 아닌 내부에서 촬영) / [gemini-pro] 캐릭터 정체성 오류 (레퍼런스 인물 교체 및 오배치) / [openrouter:x-ai/grok-4.6] 카메라가 바깥에서 엘리베이터 안을 보도록 고정됐는데 문 너머가 복도여서 시점이 반대이고 진입 방향이 뒤집힘"
     },
     {
      "label": "B",
      "score": 1000,
      "verdict_ko": "시점 위반과 캐릭터 오류에 더해 공간 구조가 금속 벽으로 왜곡되었으며, 배경 인물의 시선 연출도 누락되었습니다.  ★위반: [gemini-pro] 지정된 카메라 위치 위반 / [gemini-pro] 캐릭터 정체성 오류 / [gemini-pro] 불가능한 공간 구조 및 왜곡된 배경"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.333,
      "B": 1.75
     },
     "adjusted": {
      "A": 0.583,
      "B": 1.0
     },
     "violations": {
      "A": [
       "[gemini-pro] 지정된 카메라 위치 위반 (외부가 아닌 내부에서 촬영)",
       "[gemini-pro] 캐릭터 정체성 오류 (레퍼런스 인물 교체 및 오배치)",
       "[openrouter:x-ai/grok-4.6] 카메라가 바깥에서 엘리베이터 안을 보도록 고정됐는데 문 너머가 복도여서 시점이 반대이고 진입 방향이 뒤집힘"
      ],
      "B": [
       "[gemini-pro] 지정된 카메라 위치 위반",
       "[gemini-pro] 캐릭터 정체성 오류",
       "[gemini-pro] 불가능한 공간 구조 및 왜곡된 배경"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 583,
      "verdict_ko": "카메라 시점이 반대로 적용되고 캐릭터 정체성이 뒤바뀌는 치명적 오류가 있으나, 배경 인물의 시선 지시와 복도 공간 묘사는 상대적으로 정확합니다.  ★위반: [gemini-pro] 지정된 카메라 위치 위반 (외부가 아닌 내부에서 촬영) / [gemini-pro] 캐릭터 정체성 오류 (레퍼런스 인물 교체 및 오배치) / [openrouter:x-ai/grok-4.6] 카메라가 바깥에서 엘리베이터 안을 보도록 고정됐는데 문 너머가 복도여서 시점이 반대이고 진입 방향이 뒤집힘"
     },
     {
      "label": "B",
      "score": 1000,
      "verdict_ko": "시점 위반과 캐릭터 오류에 더해 공간 구조가 금속 벽으로 왜곡되었으며, 배경 인물의 시선 연출도 누락되었습니다.  ★위반: [gemini-pro] 지정된 카메라 위치 위반 / [gemini-pro] 캐릭터 정체성 오류 / [gemini-pro] 불가능한 공간 구조 및 왜곡된 배경"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 750,
      "verdict_ko": "배경 인물의 시선(아래를 봄)과 창문 디테일은 지시를 따랐으나, 카메라 위치가 완전히 뒤바뀌고 두 인물의 신원이 교차 적용되어 실패한 결과물입니다.  ★위반: [gemini-pro] 지정된 카메라 위치(엘리베이터 외부에서 내부를 바라봄)를 위반하고, 엘리베이터 내부에서 외부를 바라보는 뷰로 렌더링됨. / [openrouter:x-ai/grok-4.6] 카메라를 입구 바깥으로 고정했는데 문 너머가 승강장이라 약정 시점이 반대다. / [openrouter:x-ai/grok-4.6] 차장검사가 엘리베이터 안이 아니라 복도에 배치되었다."
     },
     {
      "label": "A",
      "score": 1417,
      "verdict_ko": "카메라 시점 오류 및 인물 신원 교차 적용이라는 치명적 오류를 공유하며, 배경 인물의 시선 지시마저 누락되었습니다.  ★위반: [gemini-pro] 지정된 카메라 위치(엘리베이터 외부에서 내부를 바라봄)를 위반하고, 엘리베이터 내부에서 외부를 바라보는 뷰로 렌더링됨."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.5
     },
     "adjusted": {
      "A": 1.417,
      "B": 0.75
     },
     "violations": {
      "A": [
       "[gemini-pro] 지정된 카메라 위치(엘리베이터 외부에서 내부를 바라봄)를 위반하고, 엘리베이터 내부에서 외부를 바라보는 뷰로 렌더링됨."
      ],
      "B": [
       "[gemini-pro] 지정된 카메라 위치(엘리베이터 외부에서 내부를 바라봄)를 위반하고, 엘리베이터 내부에서 외부를 바라보는 뷰로 렌더링됨.",
       "[openrouter:x-ai/grok-4.6] 카메라를 입구 바깥으로 고정했는데 문 너머가 승강장이라 약정 시점이 반대다.",
       "[openrouter:x-ai/grok-4.6] 차장검사가 엘리베이터 안이 아니라 복도에 배치되었다."
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 750,
      "verdict_ko": "배경 인물의 시선(아래를 봄)과 창문 디테일은 지시를 따랐으나, 카메라 위치가 완전히 뒤바뀌고 두 인물의 신원이 교차 적용되어 실패한 결과물입니다.  ★위반: [gemini-pro] 지정된 카메라 위치(엘리베이터 외부에서 내부를 바라봄)를 위반하고, 엘리베이터 내부에서 외부를 바라보는 뷰로 렌더링됨. / [openrouter:x-ai/grok-4.6] 카메라를 입구 바깥으로 고정했는데 문 너머가 승강장이라 약정 시점이 반대다. / [openrouter:x-ai/grok-4.6] 차장검사가 엘리베이터 안이 아니라 복도에 배치되었다."
     },
     {
      "label": "B",
      "score": 1417,
      "verdict_ko": "카메라 시점 오류 및 인물 신원 교차 적용이라는 치명적 오류를 공유하며, 배경 인물의 시선 지시마저 누락되었습니다.  ★위반: [gemini-pro] 지정된 카메라 위치(엘리베이터 외부에서 내부를 바라봄)를 위반하고, 엘리베이터 내부에서 외부를 바라보는 뷰로 렌더링됨."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1333,
     "B": 2417
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh2__bgfirst_bg.png",
   "bg_asset_id": "f5fac2f7-846d-4fff-8ee7-8161731a0ba5",
   "bg_record_key": "S62sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S62sh2::cine": {
  "applied": true,
  "fingerprint": "9a40d84ab343c20c965033a9d53aea28ec4ce7978c51eea4ed8c1da085366ec2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S62sh2_sel.png",
  "source_sha256": "5a4896ef7f15a085de733de4be02a493be60c8f3e44637d31aed99c9b0fea4a3",
  "file": "S62sh2_cine.png",
  "latency_ms": 10522
 },
 "S62sh6::confined_fp_apt": {
  "applies": false,
  "reason_ko": "엘리베이터 내부는 조종실이나 기계 제어실처럼 복잡한 제어 장치와 지정된 좌석 배치가 중요한 공간이 아니며, 단순히 두 인물이 서 있는 위치가 이야기의 개연성에 큰 영향을 미치지 않으므로 평면도 레이아웃 보조가 필요하지 않습니다.",
  "input_fingerprint": "272c83d0c4cf66f5"
 },
 "S62sh6::signage": {
  "fp": "a56abbb4fa24021a",
  "inscriptions": [
   {
    "surface_native": "엘리베이터 조작반 표시판",
    "text_native": "정원 15인\n적재하중 1000kg",
    "reason_ko": "한국의 2000년대 엘리베이터 내부 조작반에 필수적으로 부착되는 정원 및 적재하중 표시판으로, 사실적인 공간 묘사를 위해 필요합니다."
   }
  ]
 },
 "era_assess::71e78c630e17facd": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 엘리베이터 내부 조작반 및 의장",
    "search_terms_native": [
     "엘리베이터 내부 버튼",
     "현대엘리베이터 의장",
     "한국 아파트 승강기 내부",
     "엘리베이터 운전반"
    ],
    "language_lock_native": "이 검색은 반드시 한국어로만 수행되어야 하며, 다른 언어로 번역하거나 검색어를 영어로 수정하지 마십시오.",
    "reason_ko": "2000년대 한국 엘리베이터 내부의 독특한 한글 버튼 표기(열림/닫힘), 점자 버튼 형태, 비상 통화 장치 및 특유의 금속 벽 패널 디자인을 정확히 묘사하기 위함입니다."
   }
  ]
 },
 "era_ref::e3c63f5c1352d905": {
  "subject": "2000년대~2010년대 한국 엘리베이터 내부 조작반 및 의장",
  "terms": [
   "엘리베이터 내부 버튼",
   "현대엘리베이터 의장",
   "한국 아파트 승강기 내부",
   "엘리베이터 운전반"
  ],
  "queries": [
   [
    "2000년대 2010년대 현대엘리베이터 의장 엘리베이터 내부 버튼 운전반 한국 아파트 승강기 내부",
    "현대엘리베이터 카내 조작반 COP 의장 2000년대 2010년대 아파트"
   ],
   [
    "현대엘리베이터 2000년대 카내 운전반 의장 아파트",
    "현대엘리베이터 2010년대 엘리베이터 내부 조작반 의장",
    "구형 현대엘리베이터 내부 운전반 아파트 승강기",
    "현대엘리베이터 카내 조작반 버튼 디자인 역사"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://techblog.co.kr/tistory/1043/%EA%B5%AC%EC%B6%95%EC%95%84%ED%8C%8C%ED%8A%B8%EC%97%98%EB%A6%AC%EB%B2%A0%EC%9D%B4%ED%84%B0%EA%B5%90%EC%B2%B4%EA%B8%B0-07.jpg",
  "picked_reason_ko": "한국어 안내문과 전형적인 공동주택용 금속 의장, 층 버튼 조작반, 손잡이와 출입문까지 함께 보여 2001~2017년 한국의 일상적인 엘리베이터 내부 참고로 가장 적합하다.",
  "sha256": "00b2908d7937d8d57821bde1cbbc1e46651a8b90ea5b831749d397657f04ae15",
  "file": "eraref_e3c63f5c1352d905.png"
 },
 "S62sh6::bgfirst_bg": {
  "input_fingerprint": "7bbfd069ceed15f4",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 엘리베이터 문이 열린 채 나란히 서서 팽팽하게 대치하는 차장검사와 장원섭의 전경\n\nLOCATION (lock): Inside the compact elevator car at the open doors, with the two men standing beside its controls and fixed wall panels.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The backward move settles several steps into the hallway at slightly below chest height, opening into a wide diagonal view of the fully open elevator doorway. 차장검사 holds the left side and 장원섭 the right, both seen full-body with planted feet and torsos turned toward one another, their locked attention tightening the otherwise exposed doorway.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, midground, looks toward 장원섭 across the elevator doorway; 장원섭 in the middle-right of the frame, midground, looks toward 차장검사 across the elevator doorway; open elevator doorway shared by both men in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator doorway (fully open) — Its open front faces the camera on a diagonal three-quarter axis; used as Frames both men together and exposes their side-by-side placement within one confined opening; hallway area outside elevator (unoccupied in the immediate foreground); used as Creates the foreground clearance produced by the completed dolly-out.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination keeps the wide confrontation restrained, neutral-to-cool, and naturally contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 엘리베이터 내부 조작반 및 의장: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 엘리베이터 문이 열린 채 나란히 서서 팽팽하게 대치하는 차장검사와 장원섭의 전경\n\nLOCATION (lock): Inside the compact elevator car at the open doors, with the two men standing beside its controls and fixed wall panels.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: The backward move settles several steps into the hallway at slightly below chest height, opening into a wide diagonal view of the fully open elevator doorway. 차장검사 holds the left side and 장원섭 the right, both seen full-body with planted feet and torsos turned toward one another, their locked attention tightening the otherwise exposed doorway.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, midground, looks toward 장원섭 across the elevator doorway; 장원섭 in the middle-right of the frame, midground, looks toward 차장검사 across the elevator doorway; open elevator doorway shared by both men in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator doorway (fully open) — Its open front faces the camera on a diagonal three-quarter axis; used as Frames both men together and exposes their side-by-side placement within one confined opening; hallway area outside elevator (unoccupied in the immediate foreground); used as Creates the foreground clearance produced by the completed dolly-out.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination keeps the wide confrontation restrained, neutral-to-cool, and naturally contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 엘리베이터 내부 조작반 및 의장: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh6__bgfirst_bg.png",
  "asset_id": "7b9ba9d5-960e-466b-bd60-9f225fa84974",
  "input_asset_ids": [
   "79a2bb5f-9e36-4a5c-9b1b-d6630635d69d",
   "3d62c13f-094e-4df1-92c2-bab2967d17b9"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국 엘리베이터 내부 조작반 및 의장",
   "queries": [
    [
     "2000년대 2010년대 현대엘리베이터 의장 엘리베이터 내부 버튼 운전반 한국 아파트 승강기 내부",
     "현대엘리베이터 카내 조작반 COP 의장 2000년대 2010년대 아파트"
    ],
    [
     "현대엘리베이터 2000년대 카내 운전반 의장 아파트",
     "현대엘리베이터 2010년대 엘리베이터 내부 조작반 의장",
     "구형 현대엘리베이터 내부 운전반 아파트 승강기",
     "현대엘리베이터 카내 조작반 버튼 디자인 역사"
    ]
   ],
   "picked_url": "https://techblog.co.kr/tistory/1043/%EA%B5%AC%EC%B6%95%EC%95%84%ED%8C%8C%ED%8A%B8%EC%97%98%EB%A6%AC%EB%B2%A0%EC%9D%B4%ED%84%B0%EA%B5%90%EC%B2%B4%EA%B8%B0-07.jpg",
   "sha256": "00b2908d7937d8d57821bde1cbbc1e46651a8b90ea5b831749d397657f04ae15",
   "file": "eraref_e3c63f5c1352d905.png"
  }
 },
 "S62sh6": {
  "input_fingerprint": "2d0aa3241c4a949d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엘리베이터 문이 열린 채 나란히 서서 팽팽하게 대치하는 차장검사와 장원섭의 전경\n\nLOCATION (lock): Inside the compact elevator car at the open doors, with the two men standing beside its controls and fixed wall panels. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The backward move settles several steps into the hallway at slightly below chest height, opening into a wide diagonal view of the fully open elevator doorway. 차장검사 holds the left side and 장원섭 the right, both seen full-body with planted feet and torsos turned toward one another, their locked attention tightening the otherwise exposed doorway.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, midground, looks toward 장원섭 across the elevator doorway; 장원섭 in the middle-right of the frame, midground, looks toward 차장검사 across the elevator doorway; open elevator doorway shared by both men in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator doorway (fully open) — Its open front faces the camera on a diagonal three-quarter axis; used as Frames both men together and exposes their side-by-side placement within one confined opening; hallway area outside elevator (unoccupied in the immediate foreground); used as Creates the foreground clearance produced by the completed dolly-out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination keeps the wide confrontation restrained, neutral-to-cool, and naturally contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same elevator cabin, metallic finishes, open doorway, and daylight spill. Exclude the running entrance and show both men standing in tense opposition.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The elevator doors remain open while Wonseop and the deputy chief prosecutor stand facing each other inside.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 엘리베이터 조작반 표시판: \"정원 15인\n적재하중 1000kg\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective, every panel, control, seat, mirror, window and fixture stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: how many steering wheels or control surfaces, how many seats and which way each faces, where each mirror sits and what it could reflect from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second rim, sliding a seat, turning a mirror or growing a new panel is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: which seat the body occupies, which side of the wheel the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엘리베이터 문이 열린 채 나란히 서서 팽팽하게 대치하는 차장검사와 장원섭의 전경\n\nLOCATION (lock): Inside the compact elevator car at the open doors, with the two men standing beside its controls and fixed wall panels. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The backward move settles several steps into the hallway at slightly below chest height, opening into a wide diagonal view of the fully open elevator doorway. 차장검사 holds the left side and 장원섭 the right, both seen full-body with planted feet and torsos turned toward one another, their locked attention tightening the otherwise exposed doorway.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, midground, looks toward 장원섭 across the elevator doorway; 장원섭 in the middle-right of the frame, midground, looks toward 차장검사 across the elevator doorway; open elevator doorway shared by both men in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator doorway (fully open) — Its open front faces the camera on a diagonal three-quarter axis; used as Frames both men together and exposes their side-by-side placement within one confined opening; hallway area outside elevator (unoccupied in the immediate foreground); used as Creates the foreground clearance produced by the completed dolly-out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination keeps the wide confrontation restrained, neutral-to-cool, and naturally contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The elevator doors remain open while Wonseop and the deputy chief prosecutor stand facing each other inside.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 엘리베이터 문이 열린 채 나란히 서서 팽팽하게 대치하는 차장검사와 장원섭의 전경\n\nLOCATION (lock): Inside the compact elevator car at the open doors, with the two men standing beside its controls and fixed wall panels. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The backward move settles several steps into the hallway at slightly below chest height, opening into a wide diagonal view of the fully open elevator doorway. 차장검사 holds the left side and 장원섭 the right, both seen full-body with planted feet and torsos turned toward one another, their locked attention tightening the otherwise exposed doorway.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, midground, looks toward 장원섭 across the elevator doorway; 장원섭 in the middle-right of the frame, midground, looks toward 차장검사 across the elevator doorway; open elevator doorway shared by both men in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: elevator doorway (fully open) — Its open front faces the camera on a diagonal three-quarter axis; used as Frames both men together and exposes their side-by-side placement within one confined opening; hallway area outside elevator (unoccupied in the immediate foreground); used as Creates the foreground clearance produced by the completed dolly-out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination keeps the wide confrontation restrained, neutral-to-cool, and naturally contrasted.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same elevator cabin, metallic finishes, open doorway, and daylight spill. Exclude the running entrance and show both men standing in tense opposition.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The elevator doors remain open while Wonseop and the deputy chief prosecutor stand facing each other inside.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 엘리베이터 조작반 표시판: \"정원 15인\n적재하중 1000kg\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh6__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched panels, seats, controls and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S62sh6.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L57B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "요구된 한글 텍스트, 인물의 좌우 배치, 엘리베이터의 금속 질감은 완벽하게 구현했으나, 카메라와 인물의 안팎 위치가 지시(카메라가 복도, 인물이 내부)와 반대로 렌더링되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물의 좌우 위치가 바뀌었고 지정된 텍스트와 공간의 금속 질감이 누락되었으며, 안팎의 렌더링 위치도 프롬프트 지시와 반대입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 남자가 서로 시선을 맞추며 마주보고 있음.",
      "built_space": "엘리베이터 내부에서 바깥 복도를 바라보는 구도이며, 레퍼런스의 금속 재질이 아닌 무광 패널 벽면으로 구현됨.",
      "entities": "왼쪽에 장원섭, 오른쪽에 차장검사가 위치하여 프롬프트의 지시와 반대이며, 우측 조작반에 지정된 텍스트가 누락됨.",
      "hard_violations": [
       "카메라가 복도에 있고 인물이 엘리베이터 내부에 있어야 한다는 공간 배치 지시가 완전히 반대로(카메라 내부, 인물 외부) 뒤바뀜"
      ],
      "physics": "두 사람 모두 바닥에 안정적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "두 남자가 서로 시선을 맞추며 마주보고 있음.",
      "built_space": "엘리베이터 내부에서 바깥 복도를 바라보는 구도. 금속 벽면과 조작반의 형태가 레퍼런스와 일치함.",
      "entities": "왼쪽 차장검사, 오른쪽 장원섭의 배치가 정확하며, 조작반에 요구된 '정원 15인 적재하중 1000kg' 텍스트가 완벽히 출력됨.",
      "hard_violations": [
       "카메라가 복도에 있고 인물이 엘리베이터 내부에 있어야 한다는 공간 배치 지시가 완전히 반대로(카메라 내부, 인물 외부) 뒤바뀜"
      ],
      "physics": "두 사람 모두 바닥에 안정적으로 서 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "요구된 한글 텍스트, 인물의 좌우 배치, 엘리베이터의 금속 질감은 완벽하게 구현했으나, 카메라와 인물의 안팎 위치가 지시(카메라가 복도, 인물이 내부)와 반대로 렌더링되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물의 좌우 위치가 바뀌었고 지정된 텍스트와 공간의 금속 질감이 누락되었으며, 안팎의 렌더링 위치도 프롬프트 지시와 반대입니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "두 남자가 서로 시선을 맞추며 마주보고 있음.",
      "built_space": "엘리베이터 내부에서 바깥 복도를 바라보는 구도이며, 레퍼런스의 금속 재질이 아닌 무광 패널 벽면으로 구현됨.",
      "entities": "왼쪽에 장원섭, 오른쪽에 차장검사가 위치하여 프롬프트의 지시와 반대이며, 우측 조작반에 지정된 텍스트가 누락됨.",
      "hard_violations": [
       "카메라가 복도에 있고 인물이 엘리베이터 내부에 있어야 한다는 공간 배치 지시가 완전히 반대로(카메라 내부, 인물 외부) 뒤바뀜"
      ],
      "physics": "두 사람 모두 바닥에 안정적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "두 남자가 서로 시선을 맞추며 마주보고 있음.",
      "built_space": "엘리베이터 내부에서 바깥 복도를 바라보는 구도. 금속 벽면과 조작반의 형태가 레퍼런스와 일치함.",
      "entities": "왼쪽 차장검사, 오른쪽 장원섭의 배치가 정확하며, 조작반에 요구된 '정원 15인 적재하중 1000kg' 텍스트가 완벽히 출력됨.",
      "hard_violations": [
       "카메라가 복도에 있고 인물이 엘리베이터 내부에 있어야 한다는 공간 배치 지시가 완전히 반대로(카메라 내부, 인물 외부) 뒤바뀜"
      ],
      "physics": "두 사람 모두 바닥에 안정적으로 서 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트가 지시한 복도 뷰가 아닌 엘리베이터 내부 뷰로 렌더링되었으나, 공간의 금속 재질, 인물의 좌우 배치 및 조작반의 한글 텍스트를 매우 충실하게 구현함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "엘리베이터 내부 벽면이 흰색 패널로 잘못 표현되었고, 두 인물의 좌우 배치가 프롬프트 지시와 반대이며 요구된 텍스트도 누락됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "차장검사와 장원섭이 나란히 서서 서로의 얼굴을 향해 시선을 고정하고 있음.",
      "built_space": "엘리베이터 내부에서 복도를 바라보는 구도. 우측에 조작반, 좌측에 손잡이가 있으며, 금속 재질의 벽면이 참고 사진의 구조와 정확히 일치함.",
      "entities": "좌측에 핀스트라이프 정장의 차장검사, 우측에 회색 정장의 장원섭이 배치되어 지정된 인상착의와 부합함. 우측 조작반에 '정원 15인 / 직재하중 1000kg' 텍스트가 거의 정확하게 렌더링됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 엘리베이터 바닥에 안정적으로 두 발을 딛고 서 있음."
     },
     {
      "label": "B",
      "direction": "두 인물이 마주보며 시선을 교환하고 있음.",
      "built_space": "엘리베이터 내부에서 복도를 바라보는 구도. 벽면이 금속이 아닌 흰색 패널로 렌더링되어 참고 사진의 공간과 불일치함.",
      "entities": "좌측에 장원섭, 우측에 차장검사가 배치되어 인물 위치가 프롬프트와 반대임. 조작반에 요구된 텍스트가 없음.",
      "hard_violations": [],
      "physics": "두 사람 모두 바닥에 정상적으로 발을 딛고 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프롬프트가 지시한 복도 뷰가 아닌 엘리베이터 내부 뷰로 렌더링되었으나, 공간의 금속 재질, 인물의 좌우 배치 및 조작반의 한글 텍스트를 매우 충실하게 구현함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "엘리베이터 내부 벽면이 흰색 패널로 잘못 표현되었고, 두 인물의 좌우 배치가 프롬프트 지시와 반대이며 요구된 텍스트도 누락됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "차장검사와 장원섭이 나란히 서서 서로의 얼굴을 향해 시선을 고정하고 있음.",
      "built_space": "엘리베이터 내부에서 복도를 바라보는 구도. 우측에 조작반, 좌측에 손잡이가 있으며, 금속 재질의 벽면이 참고 사진의 구조와 정확히 일치함.",
      "entities": "좌측에 핀스트라이프 정장의 차장검사, 우측에 회색 정장의 장원섭이 배치되어 지정된 인상착의와 부합함. 우측 조작반에 '정원 15인 / 직재하중 1000kg' 텍스트가 거의 정확하게 렌더링됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 엘리베이터 바닥에 안정적으로 두 발을 딛고 서 있음."
     },
     {
      "label": "A",
      "direction": "두 인물이 마주보며 시선을 교환하고 있음.",
      "built_space": "엘리베이터 내부에서 복도를 바라보는 구도. 벽면이 금속이 아닌 흰색 패널로 렌더링되어 참고 사진의 공간과 불일치함.",
      "entities": "좌측에 장원섭, 우측에 차장검사가 배치되어 인물 위치가 프롬프트와 반대임. 조작반에 요구된 텍스트가 없음.",
      "hard_violations": [],
      "physics": "두 사람 모두 바닥에 정상적으로 발을 딛고 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "두 남자가 서로 시선을 맞추며 마주보고 있음.",
    "built_space": "엘리베이터 내부에서 바깥 복도를 바라보는 구도이며, 레퍼런스의 금속 재질이 아닌 무광 패널 벽면으로 구현됨.",
    "entities": "왼쪽에 장원섭, 오른쪽에 차장검사가 위치하여 프롬프트의 지시와 반대이며, 우측 조작반에 지정된 텍스트가 누락됨.",
    "hard_violations": [
     "카메라가 복도에 있고 인물이 엘리베이터 내부에 있어야 한다는 공간 배치 지시가 완전히 반대로(카메라 내부, 인물 외부) 뒤바뀜"
    ],
    "physics": "두 사람 모두 바닥에 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "두 남자가 서로 시선을 맞추며 마주보고 있음.",
    "built_space": "엘리베이터 내부에서 바깥 복도를 바라보는 구도. 금속 벽면과 조작반의 형태가 레퍼런스와 일치함.",
    "entities": "왼쪽 차장검사, 오른쪽 장원섭의 배치가 정확하며, 조작반에 요구된 '정원 15인 적재하중 1000kg' 텍스트가 완벽히 출력됨.",
    "hard_violations": [
     "카메라가 복도에 있고 인물이 엘리베이터 내부에 있어야 한다는 공간 배치 지시가 완전히 반대로(카메라 내부, 인물 외부) 뒤바뀜"
    ],
    "physics": "두 사람 모두 바닥에 안정적으로 서 있음."
   }
  ],
  "totals": {
   "A": 7,
   "B": 13
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "요구된 한글 텍스트, 인물의 좌우 배치, 엘리베이터의 금속 질감은 완벽하게 구현했으나, 카메라와 인물의 안팎 위치가 지시(카메라가 복도, 인물이 내부)와 반대로 렌더링되었습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "인물의 좌우 위치가 바뀌었고 지정된 텍스트와 공간의 금속 질감이 누락되었으며, 안팎의 렌더링 위치도 프롬프트 지시와 반대입니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L57B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   },
   {
    "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869633>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라가 복도에서 엘리베이터를 바라보아야 하나, 문 너머 배경에 창문이 있는 복도가 배치되어 안팎의 공간 구조가 거꾸로 뒤바뀜.",
     "fix_en": "Darken and blur the background hallway to deemphasize the incorrect spatial layout. Preserve characters, clothing, set, light, and framing.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "대각선 구도(diagonal three-quarter axis)가 요구되었으나, 엘리베이터를 정면에서 평면적으로 바라보는 구도로 연출됨.",
     "fix_en": "Apply slight shading to the elevator's side panels to enhance depth perception. Preserve characters, clothing, set, light, and framing.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "우측 벽면에 지정된 텍스트 중 '적재하중'이 '직재하중'으로 잘못 표기됨.",
     "fix_en": "Correct the text '직재하중' to '적재하중' on the right panel. Preserve characters, clothing, set, light, and framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "우측 조작반의 버튼 배열과 패널 디자인이 장소 참조와 다르다.",
     "fix_en": "Modify the right button panel to match the reference layout. Preserve characters, clothing, set, light, and framing.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "문 너머 홀의 창호·광택 바닥·우측 난간이 장소 참조와 다르고 가구가 추가되어 있다.",
     "fix_en": "Adjust background window and floor textures to closer match the reference without adding furniture. Preserve characters, clothing, set, light, and framing.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라가 복도에서 엘리베이터를 바라보아야 하나, 문 너머 배경에 창문이 있는 복도가 배치되어 안팎의 공간 구조가 거꾸로 뒤바뀜.",
     "severity": "major"
    },
    {
     "issue_ko": "대각선 구도(diagonal three-quarter axis)가 요구되었으나, 엘리베이터를 정면에서 평면적으로 바라보는 구도로 연출됨.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 벽면에 지정된 텍스트 중 '적재하중'이 '직재하중'으로 잘못 표기됨.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 복도가 아니라 엘리베이터 안에 있어, 전경이 비어 있는 복도가 아니라 내부 벽·손잡이고 복도는 문 너머 배경이다.",
     "severity": "major"
    },
    {
     "issue_ko": "열린 출입구가 사선 3/4가 아니라 정면에 가깝게 카메라와 마주 본다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 조작반의 버튼 배열과 패널 디자인이 장소 참조와 다르다.",
     "severity": "major"
    },
    {
     "issue_ko": "문 너머 홀의 창호·광택 바닥·우측 난간이 장소 참조와 다르고 가구가 추가되어 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "카메라가 복도에서 엘리베이터를 바라보아야 하나, 문 너머 배경에 창문이 있는 복도가 배치되어 안팎의 공간 구조가 거꾸로 뒤바뀜.",
    "fix_en": "Darken and blur the background hallway to deemphasize the incorrect spatial layout. Preserve characters, clothing, set, light, and framing.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "대각선 구도(diagonal three-quarter axis)가 요구되었으나, 엘리베이터를 정면에서 평면적으로 바라보는 구도로 연출됨.",
    "fix_en": "Apply slight shading to the elevator's side panels to enhance depth perception. Preserve characters, clothing, set, light, and framing.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "우측 벽면에 지정된 텍스트 중 '적재하중'이 '직재하중'으로 잘못 표기됨.",
    "fix_en": "Correct the text '직재하중' to '적재하중' on the right panel. Preserve characters, clothing, set, light, and framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "우측 조작반의 버튼 배열과 패널 디자인이 장소 참조와 다르다.",
    "fix_en": "Modify the right button panel to match the reference layout. Preserve characters, clothing, set, light, and framing.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "문 너머 홀의 창호·광택 바닥·우측 난간이 장소 참조와 다르고 가구가 추가되어 있다.",
    "fix_en": "Adjust background window and floor textures to closer match the reference without adding furniture. Preserve characters, clothing, set, light, and framing.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh6__bgfirst_bg.png",
   "bg_asset_id": "7b9ba9d5-960e-466b-bd60-9f225fa84974",
   "bg_record_key": "S62sh6::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S62sh2"
  }
 },
 "S62sh6::cine": {
  "applied": true,
  "fingerprint": "c7845933131d76b955b2271aaa43be2e41ae1514422e17ffaafc89990e0b3505",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S62sh6_sel.png",
  "source_sha256": "0312233ebfbf5264a9c6d96b618f8c4f826de33fb11d72634f033527554ac3ac",
  "file": "S62sh6_cine.png",
  "latency_ms": 15592
 },
 "S62sh9::confined_fp_apt": {
  "applies": false,
  "reason_ko": "엘리베이터는 탑승자의 좌석이나 복잡한 조작 기구의 위치가 중요한 통제실이나 조종석 같은 공간이 아니며, 방향성 객체의 정확한 조준이 요구되는 장면도 아니므로 평면도 배치가 필요하지 않습니다.",
  "input_fingerprint": "dead9c9954a56731"
 },
 "S62sh9::signage": {
  "fp": "f97d8dc4493bf9d8",
  "inscriptions": [
   {
    "surface_native": "엘리베이터 문 안전 경고 스티커",
    "text_native": "손끼임주의",
    "reason_ko": "엘리베이터 문이 닫히는 틈새를 클로즈업하는 구도에서 한국 엘리베이터 문에 실제로 흔히 붙어 있는 안전 경고 표식을 재현하여 현실감을 높입니다."
   }
  ]
 },
 "era_assess::188031d5dd513d46": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 아파트 및 건물 엘리베이터 내부와 문",
    "search_terms_native": [
     "한국 엘리베이터 내부",
     "2000년대 엘리베이터",
     "승강기 내부 버튼",
     "엘리베이터 문 주의 스티커"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 작성해야 하며 영어나 다른 언어로 번역하여 검색을 진행하지 마십시오.",
    "reason_ko": "2000년대~2010년대 한국 엘리베이터의 문에 붙은 특유의 노란색 손끼임 주의 스티커, 층수 안내판, 버튼 디자인 및 금속 마감재 스타일은 서구권이나 현대식 승강기와 확연히 구분되는 고유의 디테일입니다."
   }
  ]
 },
 "era_ref::be54d722c52bd8eb": {
  "subject": "2000년대~2010년대 한국의 아파트 및 건물 엘리베이터 내부와 문",
  "terms": [
   "한국 엘리베이터 내부",
   "2000년대 엘리베이터",
   "승강기 내부 버튼",
   "엘리베이터 문 주의 스티커"
  ],
  "queries": [
   [
    "한국 엘리베이터 내부 2000년대 승강기 내부 버튼",
    "2000년대 한국 엘리베이터 문 주의 스티커"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://t1.daumcdn.net/tvpot/thumb/vcdd4AaCDLB7D96F6JB9BRS/thumb.png",
  "picked_reason_ko": "3번은 2000~2010년대 한국의 일상적인 건물용 엘리베이터 승강장 문 전체와 문틀, 층 표시기, 금속 마감 및 안전 표지를 가장 명확하게 보여준다.",
  "sha256": "1a6cf82162d77c15033b69b1b79b0a2c2ca12fa081e628d98b53ab91466bcd16",
  "file": "eraref_be54d722c52bd8eb.png"
 },
 "S62sh9::bgfirst_bg": {
  "input_fingerprint": "a594e37ef71c71ef",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서로를 응시하는 두 사람을 사이에 두고 반쯤 닫힌 엘리베이터 문 틈새 클로즈업\n\nLOCATION (lock): At the narrowing gap of the elevator's sliding doors, looking into the compact car and its tightly fixed interior fittings.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Immediately outside the elevator at shoulder height, the camera completes its push toward the off-center seam as the half-closed panels create a narrow vertical viewing channel. 차장검사's face remains visible on the left and 장원섭's on the right, each in opposing three-quarter angles, with the door edges dividing their faces but leaving their reciprocal eyeline unbroken.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, background, looks toward 장원섭 across the closing seam; 장원섭 in the middle-right of the frame, background, looks toward 차장검사 across the closing seam; half-closed elevator door seam in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: elevator door panels (half closed) — Their front faces converge toward an off-center vertical seam facing the camera; used as The panels form foreground shutters that physically divide the two faces while preserving the eyeline through the central gap.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination remains subdued and neutral-to-cool, with moderate contrast across the faces and door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 아파트 및 건물 엘리베이터 내부와 문: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서로를 응시하는 두 사람을 사이에 두고 반쯤 닫힌 엘리베이터 문 틈새 클로즈업\n\nLOCATION (lock): At the narrowing gap of the elevator's sliding doors, looking into the compact car and its tightly fixed interior fittings.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Immediately outside the elevator at shoulder height, the camera completes its push toward the off-center seam as the half-closed panels create a narrow vertical viewing channel. 차장검사's face remains visible on the left and 장원섭's on the right, each in opposing three-quarter angles, with the door edges dividing their faces but leaving their reciprocal eyeline unbroken.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, background, looks toward 장원섭 across the closing seam; 장원섭 in the middle-right of the frame, background, looks toward 차장검사 across the closing seam; half-closed elevator door seam in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: elevator door panels (half closed) — Their front faces converge toward an off-center vertical seam facing the camera; used as The panels form foreground shutters that physically divide the two faces while preserving the eyeline through the central gap.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination remains subdued and neutral-to-cool, with moderate contrast across the faces and door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 아파트 및 건물 엘리베이터 내부와 문: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh9__bgfirst_bg.png",
  "asset_id": "6cfcce0f-bc01-4484-917f-2270a03b7a91",
  "input_asset_ids": [
   "d516a7ff-9aaa-4cc8-95cf-d07f6a717fcd",
   "3d62c13f-094e-4df1-92c2-bab2967d17b9"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국의 아파트 및 건물 엘리베이터 내부와 문",
   "queries": [
    [
     "한국 엘리베이터 내부 2000년대 승강기 내부 버튼",
     "2000년대 한국 엘리베이터 문 주의 스티커"
    ]
   ],
   "picked_url": "https://t1.daumcdn.net/tvpot/thumb/vcdd4AaCDLB7D96F6JB9BRS/thumb.png",
   "sha256": "1a6cf82162d77c15033b69b1b79b0a2c2ca12fa081e628d98b53ab91466bcd16",
   "file": "eraref_be54d722c52bd8eb.png"
  }
 },
 "S62sh9": {
  "input_fingerprint": "abbf20c5a634e36c",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로를 응시하는 두 사람을 사이에 두고 반쯤 닫힌 엘리베이터 문 틈새 클로즈업\n\nLOCATION (lock): At the narrowing gap of the elevator's sliding doors, looking into the compact car and its tightly fixed interior fittings. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Immediately outside the elevator at shoulder height, the camera completes its push toward the off-center seam as the half-closed panels create a narrow vertical viewing channel. 차장검사's face remains visible on the left and 장원섭's on the right, each in opposing three-quarter angles, with the door edges dividing their faces but leaving their reciprocal eyeline unbroken.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, background, looks toward 장원섭 across the closing seam; 장원섭 in the middle-right of the frame, background, looks toward 차장검사 across the closing seam; half-closed elevator door seam in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: elevator door panels (half closed) — Their front faces converge toward an off-center vertical seam facing the camera; used as The panels form foreground shutters that physically divide the two faces while preserving the eyeline through the central gap.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination remains subdued and neutral-to-cool, with moderate contrast across the faces and door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭, 차장검사 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same elevator interior, closing metal doors, lighting, and both men's fixed positions. Exclude their full bodies and focus on the narrowing gap between the doors.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The elevator doors are closing again around the two men, narrowing the visible gap between them.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 엘리베이터 문 안전 경고 스티커: \"손끼임주의\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective, every panel, control, seat, mirror, window and fixture stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: how many steering wheels or control surfaces, how many seats and which way each faces, where each mirror sits and what it could reflect from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second rim, sliding a seat, turning a mirror or growing a new panel is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: which seat the body occupies, which side of the wheel the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로를 응시하는 두 사람을 사이에 두고 반쯤 닫힌 엘리베이터 문 틈새 클로즈업\n\nLOCATION (lock): At the narrowing gap of the elevator's sliding doors, looking into the compact car and its tightly fixed interior fittings. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Immediately outside the elevator at shoulder height, the camera completes its push toward the off-center seam as the half-closed panels create a narrow vertical viewing channel. 차장검사's face remains visible on the left and 장원섭's on the right, each in opposing three-quarter angles, with the door edges dividing their faces but leaving their reciprocal eyeline unbroken.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, background, looks toward 장원섭 across the closing seam; 장원섭 in the middle-right of the frame, background, looks toward 차장검사 across the closing seam; half-closed elevator door seam in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: elevator door panels (half closed) — Their front faces converge toward an off-center vertical seam facing the camera; used as The panels form foreground shutters that physically divide the two faces while preserving the eyeline through the central gap.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination remains subdued and neutral-to-cool, with moderate contrast across the faces and door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The elevator doors are closing again around the two men, narrowing the visible gap between them.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서로를 응시하는 두 사람을 사이에 두고 반쯤 닫힌 엘리베이터 문 틈새 클로즈업\n\nLOCATION (lock): At the narrowing gap of the elevator's sliding doors, looking into the compact car and its tightly fixed interior fittings. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Immediately outside the elevator at shoulder height, the camera completes its push toward the off-center seam as the half-closed panels create a narrow vertical viewing channel. 차장검사's face remains visible on the left and 장원섭's on the right, each in opposing three-quarter angles, with the door edges dividing their faces but leaving their reciprocal eyeline unbroken.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, background, looks toward 장원섭 across the closing seam; 장원섭 in the middle-right of the frame, background, looks toward 차장검사 across the closing seam; half-closed elevator door seam in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: elevator door panels (half closed) — Their front faces converge toward an off-center vertical seam facing the camera; used as The panels form foreground shutters that physically divide the two faces while preserving the eyeline through the central gap.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime institutional illumination remains subdued and neutral-to-cool, with moderate contrast across the faces and door edges.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭, 차장검사 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same elevator interior, closing metal doors, lighting, and both men's fixed positions. Exclude their full bodies and focus on the narrowing gap between the doors.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The elevator doors are closing again around the two men, narrowing the visible gap between them.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리); 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 엘리베이터 문 안전 경고 스티커: \"손끼임주의\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh9__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched panels, seats, controls and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S62sh9.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L57B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 583,
      "verdict_ko": "지정된 인물의 좌우 배치와 복장, '손끼임주의' 텍스트를 완벽히 구현했으나 배경 공간(조작반과 외부 창문 혼합) 묘사에 오류가 있습니다.  ★위반: [gemini-pro] 엘리베이터 내부 조작반과 외부 복도 창문이 같은 벽에 합성된 물리적으로 불가능한 건축 구조 / [gemini-pro] 지정된 카메라 방향(내부를 바라봄)과 어긋나게 외부 복도가 배경으로 렌더링됨 / [openrouter:x-ai/grok-4.6] 약정된 카메라(엘리베이터 바로 밖에서 차 안을 봄)가 복도 창과 랜딩 패널이 보이는 반대 시점으로 렌더됨"
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "배경과 카메라 방향은 올바르나, 인물의 좌우 배치가 반대로 뒤바뀌었고 의상 오류 및 필수 텍스트가 누락되었습니다.  ★위반: [gemini-pro] 지정된 두 인물의 프레임 내 좌우 배치 지시를 반대로 렌더링함"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.25,
      "B": 0.583
     },
     "violations": {
      "A": [
       "[gemini-pro] 지정된 두 인물의 프레임 내 좌우 배치 지시를 반대로 렌더링함"
      ],
      "B": [
       "[gemini-pro] 엘리베이터 내부 조작반과 외부 복도 창문이 같은 벽에 합성된 물리적으로 불가능한 건축 구조",
       "[gemini-pro] 지정된 카메라 방향(내부를 바라봄)과 어긋나게 외부 복도가 배경으로 렌더링됨",
       "[openrouter:x-ai/grok-4.6] 약정된 카메라(엘리베이터 바로 밖에서 차 안을 봄)가 복도 창과 랜딩 패널이 보이는 반대 시점으로 렌더됨"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 583,
      "verdict_ko": "지정된 인물의 좌우 배치와 복장, '손끼임주의' 텍스트를 완벽히 구현했으나 배경 공간(조작반과 외부 창문 혼합) 묘사에 오류가 있습니다.  ★위반: [gemini-pro] 엘리베이터 내부 조작반과 외부 복도 창문이 같은 벽에 합성된 물리적으로 불가능한 건축 구조 / [gemini-pro] 지정된 카메라 방향(내부를 바라봄)과 어긋나게 외부 복도가 배경으로 렌더링됨 / [openrouter:x-ai/grok-4.6] 약정된 카메라(엘리베이터 바로 밖에서 차 안을 봄)가 복도 창과 랜딩 패널이 보이는 반대 시점으로 렌더됨"
     },
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "배경과 카메라 방향은 올바르나, 인물의 좌우 배치가 반대로 뒤바뀌었고 의상 오류 및 필수 텍스트가 누락되었습니다.  ★위반: [gemini-pro] 지정된 두 인물의 프레임 내 좌우 배치 지시를 반대로 렌더링함"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "지정된 인물 배치와 의상, 배경의 창문 및 '손끼임주의' 텍스트를 모두 정확하게 구현하여 프롬프트 충실도가 높습니다.  ★위반: [openrouter:x-ai/grok-4.6] 약정된 카메라 위치(엘리베이터 바로 밖에서 칸 안을 봄)를 어기고 승강장 창이 배경인 내부→외부 시점으로 렌더링됨"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "인물 간 의상이 뒤바뀌었고, 기준 위치의 창문과 요구된 경고문 텍스트가 누락되어 지시사항을 다수 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.5
     },
     "adjusted": {
      "A": 1.25,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 약정된 카메라 위치(엘리베이터 바로 밖에서 칸 안을 봄)를 어기고 승강장 창이 배경인 내부→외부 시점으로 렌더링됨"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1250,
      "verdict_ko": "지정된 인물 배치와 의상, 배경의 창문 및 '손끼임주의' 텍스트를 모두 정확하게 구현하여 프롬프트 충실도가 높습니다.  ★위반: [openrouter:x-ai/grok-4.6] 약정된 카메라 위치(엘리베이터 바로 밖에서 칸 안을 봄)를 어기고 승강장 창이 배경인 내부→외부 시점으로 렌더링됨"
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "인물 간 의상이 뒤바뀌었고, 기준 위치의 창문과 요구된 경고문 텍스트가 누락되어 지시사항을 다수 위반했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2750,
     "B": 1833
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 2750,
   "B": 1833
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 583,
    "verdict_ko": "지정된 인물의 좌우 배치와 복장, '손끼임주의' 텍스트를 완벽히 구현했으나 배경 공간(조작반과 외부 창문 혼합) 묘사에 오류가 있습니다.  ★위반: [gemini-pro] 엘리베이터 내부 조작반과 외부 복도 창문이 같은 벽에 합성된 물리적으로 불가능한 건축 구조 / [gemini-pro] 지정된 카메라 방향(내부를 바라봄)과 어긋나게 외부 복도가 배경으로 렌더링됨 / [openrouter:x-ai/grok-4.6] 약정된 카메라(엘리베이터 바로 밖에서 차 안을 봄)가 복도 창과 랜딩 패널이 보이는 반대 시점으로 렌더됨"
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "배경과 카메라 방향은 올바르나, 인물의 좌우 배치가 반대로 뒤바뀌었고 의상 오류 및 필수 텍스트가 누락되었습니다.  ★위반: [gemini-pro] 지정된 두 인물의 프레임 내 좌우 배치 지시를 반대로 렌더링함"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L57B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   },
   {
    "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869633>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "왼쪽 인물(차장검사)과 오른쪽 인물(장원섭)의 정장과 넥타이 의상이 서로 뒤바뀌어 렌더링되었습니다.",
     "fix_en": "Change the left man's suit to navy pinstripe double-breasted with a burgundy tie and tie bar, and the right man's suit to dark grey single-breasted with a navy patterned tie. Preserve the men's faces, poses, the foreground doors, the background, and the lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "배경 레퍼런스의 정면에 있던 창문 구조가 완전히 사라지고, 임의의 가로 손잡이가 있는 막힌 금속 벽면으로 대체되었습니다.",
     "fix_en": "Redraw the background space visible between the men to show a bright window with a drawn blind set into tiled walls, replacing the metal wall and handrail. Preserve the men, their clothing, the foreground doors, and the lighting.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전경 문 패널이 참조의 점박이 석재 질감이 아니라 브러시드 스테인리스 금속이다",
     "fix_en": "Change the material of the foreground elevator door panels to a dark speckled stone texture. Preserve the men, the background, and the lighting.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "왼쪽 인물(차장검사)과 오른쪽 인물(장원섭)의 정장과 넥타이 의상이 서로 뒤바뀌어 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경 레퍼런스의 정면에 있던 창문 구조가 완전히 사라지고, 임의의 가로 손잡이가 있는 막힌 금속 벽면으로 대체되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "오른쪽 인물이 장원섭(40대 초반·갸름한 얼굴·회색 수트)이 아니라 차장검사와 같은 중년 각진 얼굴과 남색 핀스트라이프 더블브레스트 수트다",
     "severity": "critical"
    },
    {
     "issue_ko": "문 틈 너머 내부가 참조 배경의 창문·블라인드·석재 벽·타일 바닥이 아니라 금속 엘리베이터 벽으로 바뀌었다",
     "severity": "major"
    },
    {
     "issue_ko": "전경 문 패널이 참조의 점박이 석재 질감이 아니라 브러시드 스테인리스 금속이다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 차장검사의 넥타이가 참조의 버건디 무늬 타이·타이바가 아니라 어두운 남색 넥타이다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "전경 문 패널이 참조의 점박이 석재 질감이 아니라 브러시드 스테인리스 금속이다",
    "fix_en": "Change the material of the foreground elevator door panels to a dark speckled stone texture. Preserve the men, the background, and the lighting.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the left man's suit to navy pinstripe double-breasted with a burgundy tie and tie bar, and the right man's suit to dark grey single-breasted with a navy patterned tie. Preserve the men's faces, poses, the foreground doors, the background, and the lighting.\n- Redraw the background space visible between the men to show a bright window with a drawn blind set into tiled walls, replacing the metal wall and handrail. Preserve the men, their clothing, the foreground doors, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "배경의 창문 공간을 엘리베이터 내부로 왜곡하고 두 인물의 의상과 위치가 반대로 배치되는 심각한 오류가 있으나, 엘리베이터 슬라이딩 도어의 기본적인 형태와 금속 재질은 유지했습니다.  ★위반: [gemini-pro] 배경 공간 완전 변형 (레퍼런스의 창문과 복도를 무시하고 엘리베이터 내부로 렌더링) / [gemini-pro] 인물 위치 및 의상 배치 오류 (좌우 캐릭터의 설정이 뒤바뀜) / [openrouter:x-ai/grok-4.6] 장원섭 신원 상실·차장검사형 인물 중복 / [openrouter:x-ai/grok-4.6] 잠긴 배경을 엘리베이터 캐빈으로 전면 재설계 / [openrouter:x-ai/grok-4.6] 참조 도어 재질·창문·바닥 픽스처 미유지"
     },
     {
      "label": "B",
      "score": 667,
      "verdict_ko": "의상 배치는 프롬프트를 따랐으나, 엘리베이터 문을 손잡이가 달린 일반 나무 여닫이문으로 완전히 잘못 묘사하였고 두 인물의 얼굴이 복제된 것처럼 유사하게 렌더링되어 심각하게 훼손되었습니다.  ★위반: [gemini-pro] 필수 구조물 완전 변형 (엘리베이터 슬라이딩 도어가 손잡이 달린 나무 문으로 변경됨) / [gemini-pro] 인물 외모 동일화 (두 캐릭터의 얼굴이 복제된 것처럼 비슷하게 렌더링됨) / [openrouter:x-ai/grok-4.6] 잠긴 엘리베이터 문틈 로케이션을 목재 방문·타일 실내로 교체 / [openrouter:x-ai/grok-4.6] 참조 배경의 기하·재질·픽스처 미유지"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.25,
      "B": 1.667
     },
     "adjusted": {
      "A": 0.0,
      "B": 0.667
     },
     "violations": {
      "A": [
       "[gemini-pro] 배경 공간 완전 변형 (레퍼런스의 창문과 복도를 무시하고 엘리베이터 내부로 렌더링)",
       "[gemini-pro] 인물 위치 및 의상 배치 오류 (좌우 캐릭터의 설정이 뒤바뀜)",
       "[openrouter:x-ai/grok-4.6] 장원섭 신원 상실·차장검사형 인물 중복",
       "[openrouter:x-ai/grok-4.6] 잠긴 배경을 엘리베이터 캐빈으로 전면 재설계",
       "[openrouter:x-ai/grok-4.6] 참조 도어 재질·창문·바닥 픽스처 미유지"
      ],
      "B": [
       "[gemini-pro] 필수 구조물 완전 변형 (엘리베이터 슬라이딩 도어가 손잡이 달린 나무 문으로 변경됨)",
       "[gemini-pro] 인물 외모 동일화 (두 캐릭터의 얼굴이 복제된 것처럼 비슷하게 렌더링됨)",
       "[openrouter:x-ai/grok-4.6] 잠긴 엘리베이터 문틈 로케이션을 목재 방문·타일 실내로 교체",
       "[openrouter:x-ai/grok-4.6] 참조 배경의 기하·재질·픽스처 미유지"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.75,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "배경의 창문 공간을 엘리베이터 내부로 왜곡하고 두 인물의 의상과 위치가 반대로 배치되는 심각한 오류가 있으나, 엘리베이터 슬라이딩 도어의 기본적인 형태와 금속 재질은 유지했습니다.  ★위반: [gemini-pro] 배경 공간 완전 변형 (레퍼런스의 창문과 복도를 무시하고 엘리베이터 내부로 렌더링) / [gemini-pro] 인물 위치 및 의상 배치 오류 (좌우 캐릭터의 설정이 뒤바뀜) / [openrouter:x-ai/grok-4.6] 장원섭 신원 상실·차장검사형 인물 중복 / [openrouter:x-ai/grok-4.6] 잠긴 배경을 엘리베이터 캐빈으로 전면 재설계 / [openrouter:x-ai/grok-4.6] 참조 도어 재질·창문·바닥 픽스처 미유지"
     },
     {
      "label": "B",
      "score": 667,
      "verdict_ko": "의상 배치는 프롬프트를 따랐으나, 엘리베이터 문을 손잡이가 달린 일반 나무 여닫이문으로 완전히 잘못 묘사하였고 두 인물의 얼굴이 복제된 것처럼 유사하게 렌더링되어 심각하게 훼손되었습니다.  ★위반: [gemini-pro] 필수 구조물 완전 변형 (엘리베이터 슬라이딩 도어가 손잡이 달린 나무 문으로 변경됨) / [gemini-pro] 인물 외모 동일화 (두 캐릭터의 얼굴이 복제된 것처럼 비슷하게 렌더링됨) / [openrouter:x-ai/grok-4.6] 잠긴 엘리베이터 문틈 로케이션을 목재 방문·타일 실내로 교체 / [openrouter:x-ai/grok-4.6] 참조 배경의 기하·재질·픽스처 미유지"
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 583,
      "verdict_ko": "엘리베이터 문 틈새의 클로즈업 구도는 어느 정도 반영했으나, 레퍼런스의 배경(창문과 타일)을 무시하고 임의의 금속 배경으로 대체했으며, 두 인물의 의상이 서로 뒤바뀌는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 지정된 장소 배경(내부 창문 및 구조) 임의 변형 및 대체 / [openrouter:x-ai/grok-4.6] 잠금된 샷 배경을 스테인리스 엘리베이터 카로 교체함 / [openrouter:x-ai/grok-4.6] 두 인물의 얼굴·헤어·의상이 캐릭터 참조와 불일치해 다른 사람이 됨"
     },
     {
      "label": "A",
      "score": 250,
      "verdict_ko": "지정된 엘리베이터 슬라이딩 도어를 손잡이가 달린 나무 여닫이문으로 완전히 변형시켰고, 왼쪽 인물의 얼굴이 오른쪽 인물과 동일하게 복제되어 신원(Identity) 규칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 지정된 장소(엘리베이터 구조 및 문) 완전 변형 (나무 여닫이문 생성) / [gemini-pro] 인물 신원 상실 (왼쪽 인물이 오른쪽 인물의 얼굴과 동일하게 복제됨) / [openrouter:x-ai/grok-4.6] 잠금된 샷 배경을 나무 여닫이문·타일 실내·다른 창으로 교체함"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.0,
      "B": 1.333
     },
     "adjusted": {
      "A": 0.25,
      "B": 0.583
     },
     "violations": {
      "A": [
       "[gemini-pro] 지정된 장소(엘리베이터 구조 및 문) 완전 변형 (나무 여닫이문 생성)",
       "[gemini-pro] 인물 신원 상실 (왼쪽 인물이 오른쪽 인물의 얼굴과 동일하게 복제됨)",
       "[openrouter:x-ai/grok-4.6] 잠금된 샷 배경을 나무 여닫이문·타일 실내·다른 창으로 교체함"
      ],
      "B": [
       "[gemini-pro] 지정된 장소 배경(내부 창문 및 구조) 임의 변형 및 대체",
       "[openrouter:x-ai/grok-4.6] 잠금된 샷 배경을 스테인리스 엘리베이터 카로 교체함",
       "[openrouter:x-ai/grok-4.6] 두 인물의 얼굴·헤어·의상이 캐릭터 참조와 불일치해 다른 사람이 됨"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 583,
      "verdict_ko": "엘리베이터 문 틈새의 클로즈업 구도는 어느 정도 반영했으나, 레퍼런스의 배경(창문과 타일)을 무시하고 임의의 금속 배경으로 대체했으며, 두 인물의 의상이 서로 뒤바뀌는 치명적인 오류가 있습니다.  ★위반: [gemini-pro] 지정된 장소 배경(내부 창문 및 구조) 임의 변형 및 대체 / [openrouter:x-ai/grok-4.6] 잠금된 샷 배경을 스테인리스 엘리베이터 카로 교체함 / [openrouter:x-ai/grok-4.6] 두 인물의 얼굴·헤어·의상이 캐릭터 참조와 불일치해 다른 사람이 됨"
     },
     {
      "label": "B",
      "score": 250,
      "verdict_ko": "지정된 엘리베이터 슬라이딩 도어를 손잡이가 달린 나무 여닫이문으로 완전히 변형시켰고, 왼쪽 인물의 얼굴이 오른쪽 인물과 동일하게 복제되어 신원(Identity) 규칙을 심각하게 위반했습니다.  ★위반: [gemini-pro] 지정된 장소(엘리베이터 구조 및 문) 완전 변형 (나무 여닫이문 생성) / [gemini-pro] 인물 신원 상실 (왼쪽 인물이 오른쪽 인물의 얼굴과 동일하게 복제됨) / [openrouter:x-ai/grok-4.6] 잠금된 샷 배경을 나무 여닫이문·타일 실내·다른 창으로 교체함"
     }
    ],
    "all_candidates_fail": true
   },
   "combined": {
    "totals": {
     "A": 583,
     "B": 917
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S62sh9__bgfirst_bg.png",
   "bg_asset_id": "6cfcce0f-bc01-4484-917f-2270a03b7a91",
   "bg_record_key": "S62sh9::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S62sh6"
  }
 },
 "S62sh9::cine": {
  "applied": true,
  "fingerprint": "d5b12495591d729455dccaf7115983f57ecfec4d666687ebd2832c439b93f493",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S62sh9_sel.png",
  "source_sha256": "1ef899e1e6bf8a703832d48edff7bdcb80f8afc6ba412d6f3486759aa7aca13c",
  "file": "S62sh9_cine.png",
  "latency_ms": 10864
 },
 "S63sh5::signage": {
  "fp": "d0fa1d7de4ca0ea0",
  "inscriptions": []
 },
 "era_assess::35d9f1d4da3290a1": {
  "subjects": [
   {
    "subject_native": "대한민국 교도소 접견창구 및 가림막 (2001년~2017년)",
    "search_terms_native": [
     "교도소 접견실",
     "교도소 면회실",
     "교도소 접견창구",
     "교도소 접견대"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "2000년대 대한민국 교도소의 접견실 구조, 아크릴 가림막 형태, 인터폰 및 접견대 배치는 서구식 교도소의 일반적인 면회실 모습과 상이하므로 고증이 필요합니다."
   }
  ]
 },
 "era_ref::705bdd20141d4656": {
  "subject": "대한민국 교도소 접견창구 및 가림막 (2001년~2017년)",
  "terms": [
   "교도소 접견실",
   "교도소 면회실",
   "교도소 접견창구",
   "교도소 접견대"
  ],
  "queries": [
   [
    "대한민국 교도소 접견실 접견창구 가림막 2001년 2017년",
    "교도소 면회실 접견대 유리 칸막이 사진",
    "교정시설 접견실 접견창구 법무부",
    "교도소 접견실 구조 접견대 가림막"
   ],
   [
    "2001년 교도소 접견실 강화유리 마이크",
    "2005년 교도소 접견실 강화유리 접견대",
    "2013년 교도소 접견실 유리 칸막이 변호사",
    "2017년 교도소 접견실 사진 접견창구"
   ]
  ],
  "candidates": 3,
  "picked_index": 1,
  "picked_url": "https://www.corrections.go.kr/sites/corrections/images/contents/cor_new101001_9.jpg",
  "picked_reason_ko": "대한민국 교도소의 실제 접견창구와 유리 가림막이 주 피사체로서 구조·재료·구획 방식까지 가장 명확하게 보인다.",
  "sha256": "5e63c33ae2a2e16f906c9fcb669ed6549c575eea913bc7fdd88b2be66f1abcb3",
  "file": "eraref_705bdd20141d4656.png"
 },
 "S63sh5::bgfirst_bg": {
  "input_fingerprint": "5f9a1f3c270bd55d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 가림판 너머 신경희의 손바닥 위치에 맞춰 자신의 손을 포개어 얹은 지국현의 손 클로즈업\n\nLOCATION (lock): At the divided visitation counter inside the prison visiting room, across the transparent partition.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height and very close to the divider, an oblique insert places 지국현's palm prominently in the center-right while 신경희's aligned palm remains clearly legible on the opposite side. The paired hands stay below forty percent each and are framed by a visible strip of divider and surrounding surface, making the separation as readable as the attempted contact.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 지국현's hand in the middle-right of the frame, foreground, reaches for 신경희's aligned palm across the divider; 신경희's hand in the middle-left of the frame, background, reaches for 지국현's aligned palm across the divider; transparent divider between the palms in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between the visitors) — The camera sees 지국현's near-side palm against the surface and 신경희's aligned palm through the opposite side; used as Separates the two palms while allowing their alignment to remain visible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination is soft and restrained, with neutral-to-cool color and enough tonal separation to distinguish both palms through the divider.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 교도소 접견창구 및 가림막 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 가림판 너머 신경희의 손바닥 위치에 맞춰 자신의 손을 포개어 얹은 지국현의 손 클로즈업\n\nLOCATION (lock): At the divided visitation counter inside the prison visiting room, across the transparent partition.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At hand height and very close to the divider, an oblique insert places 지국현's palm prominently in the center-right while 신경희's aligned palm remains clearly legible on the opposite side. The paired hands stay below forty percent each and are framed by a visible strip of divider and surrounding surface, making the separation as readable as the attempted contact.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 지국현's hand in the middle-right of the frame, foreground, reaches for 신경희's aligned palm across the divider; 신경희's hand in the middle-left of the frame, background, reaches for 지국현's aligned palm across the divider; transparent divider between the palms in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between the visitors) — The camera sees 지국현's near-side palm against the surface and 신경희's aligned palm through the opposite side; used as Separates the two palms while allowing their alignment to remain visible.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination is soft and restrained, with neutral-to-cool color and enough tonal separation to distinguish both palms through the divider.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 교도소 접견창구 및 가림막 (2001년~2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S63sh5__bgfirst_bg.png",
  "asset_id": "e0f1a0c4-3ce1-481a-9ed6-1655e6fd9fdc",
  "input_asset_ids": [
   "a9978e5e-1bf1-4f7f-a1f5-2b51c38b1c14",
   "9f094613-15bf-4b6a-a4f1-7fd66f944936"
  ],
  "era_research": {
   "subject": "대한민국 교도소 접견창구 및 가림막 (2001년~2017년)",
   "queries": [
    [
     "대한민국 교도소 접견실 접견창구 가림막 2001년 2017년",
     "교도소 면회실 접견대 유리 칸막이 사진",
     "교정시설 접견실 접견창구 법무부",
     "교도소 접견실 구조 접견대 가림막"
    ],
    [
     "2001년 교도소 접견실 강화유리 마이크",
     "2005년 교도소 접견실 강화유리 접견대",
     "2013년 교도소 접견실 유리 칸막이 변호사",
     "2017년 교도소 접견실 사진 접견창구"
    ]
   ],
   "picked_url": "https://www.corrections.go.kr/sites/corrections/images/contents/cor_new101001_9.jpg",
   "sha256": "5e63c33ae2a2e16f906c9fcb669ed6549c575eea913bc7fdd88b2be66f1abcb3",
   "file": "eraref_705bdd20141d4656.png"
  }
 },
 "S63sh5": {
  "input_fingerprint": "18b59ee36f1f3b8d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 가림판 너머 신경희의 손바닥 위치에 맞춰 자신의 손을 포개어 얹은 지국현의 손 클로즈업\n\nLOCATION (lock): At the divided visitation counter inside the prison visiting room, across the transparent partition. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height and very close to the divider, an oblique insert places 지국현's palm prominently in the center-right while 신경희's aligned palm remains clearly legible on the opposite side. The paired hands stay below forty percent each and are framed by a visible strip of divider and surrounding surface, making the separation as readable as the attempted contact.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 지국현's hand in the middle-right of the frame, foreground, reaches for 신경희's aligned palm across the divider; 신경희's hand in the middle-left of the frame, background, reaches for 지국현's aligned palm across the divider; transparent divider between the palms in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between the visitors) — The camera sees 지국현's near-side palm against the surface and 신경희's aligned palm through the opposite side; used as Separates the two palms while allowing their alignment to remain visible.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination is soft and restrained, with neutral-to-cool color and enough tonal separation to distinguish both palms through the divider.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Gyeong-hui's palm remains against the transparent partition, with Ji Gukhyeon's hand aligned over hers on the opposite side.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 가림판 너머 신경희의 손바닥 위치에 맞춰 자신의 손을 포개어 얹은 지국현의 손 클로즈업\n\nLOCATION (lock): At the divided visitation counter inside the prison visiting room, across the transparent partition. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height and very close to the divider, an oblique insert places 지국현's palm prominently in the center-right while 신경희's aligned palm remains clearly legible on the opposite side. The paired hands stay below forty percent each and are framed by a visible strip of divider and surrounding surface, making the separation as readable as the attempted contact.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 지국현's hand in the middle-right of the frame, foreground, reaches for 신경희's aligned palm across the divider; 신경희's hand in the middle-left of the frame, background, reaches for 지국현's aligned palm across the divider; transparent divider between the palms in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between the visitors) — The camera sees 지국현's near-side palm against the surface and 신경희's aligned palm through the opposite side; used as Separates the two palms while allowing their alignment to remain visible.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination is soft and restrained, with neutral-to-cool color and enough tonal separation to distinguish both palms through the divider.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Gyeong-hui's palm remains against the transparent partition, with Ji Gukhyeon's hand aligned over hers on the opposite side.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 가림판 너머 신경희의 손바닥 위치에 맞춰 자신의 손을 포개어 얹은 지국현의 손 클로즈업\n\nLOCATION (lock): At the divided visitation counter inside the prison visiting room, across the transparent partition. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hand height and very close to the divider, an oblique insert places 지국현's palm prominently in the center-right while 신경희's aligned palm remains clearly legible on the opposite side. The paired hands stay below forty percent each and are framed by a visible strip of divider and surrounding surface, making the separation as readable as the attempted contact.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 지국현's hand in the middle-right of the frame, foreground, reaches for 신경희's aligned palm across the divider; 신경희's hand in the middle-left of the frame, background, reaches for 지국현's aligned palm across the divider; transparent divider between the palms in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between the visitors) — The camera sees 지국현's near-side palm against the surface and 신경희's aligned palm through the opposite side; used as Separates the two palms while allowing their alignment to remain visible.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination is soft and restrained, with neutral-to-cool color and enough tonal separation to distinguish both palms through the divider.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Gyeong-hui's palm remains against the transparent partition, with Ji Gukhyeon's hand aligned over hers on the opposite side.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리); 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S63sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S63sh5.png"
    },
    {
     "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1176458>"
    },
    {
     "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1236081>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L45B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1176458>"
    },
    {
     "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:1236081>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 인서트 클로즈업 프레이밍을 지키며 로케이션 사진의 스테인리스 카운터 재질을 정확히 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레이밍이 지시된 것보다 넓게 잡혔으며 로케이션의 카운터 재질(석재로 묘사)이 일치하지 않음."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 손이 가림판을 사이에 두고 맞닿음. 여성의 시선은 손을 향함.",
      "built_space": "면회실 가림판을 묘사했으나 카운터가 석재 질감으로 보임.",
      "entities": "지국현의 팔과 신경희의 얼굴/상반신이 보이며 인물 외형은 일치함.",
      "hard_violations": [
       "프레이밍 위반 (인서트 클로즈업이 아닌 상반신 포함 샷)",
       "로케이션 불일치 (스테인리스가 아닌 석재 카운터)"
      ],
      "physics": "두 사람의 손 모두 투명 가림판에 지탱됨."
     },
     {
      "label": "B",
      "direction": "두 손이 가림판을 사이에 두고 마주 닿아 있음.",
      "built_space": "스테인리스 카운터와 가림판 구조가 로케이션 레퍼런스와 일치함.",
      "entities": "지국현과 신경희의 손 및 팔이 지정된 복장과 함께 등장함.",
      "hard_violations": [],
      "physics": "두 사람의 손 모두 투명 가림판 표면에 안정적으로 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 인서트 클로즈업 프레이밍을 지키며 로케이션 사진의 스테인리스 카운터 재질을 정확히 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레이밍이 지시된 것보다 넓게 잡혔으며 로케이션의 카운터 재질(석재로 묘사)이 일치하지 않음."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "두 손이 가림판을 사이에 두고 맞닿음. 여성의 시선은 손을 향함.",
      "built_space": "면회실 가림판을 묘사했으나 카운터가 석재 질감으로 보임.",
      "entities": "지국현의 팔과 신경희의 얼굴/상반신이 보이며 인물 외형은 일치함.",
      "hard_violations": [
       "프레이밍 위반 (인서트 클로즈업이 아닌 상반신 포함 샷)",
       "로케이션 불일치 (스테인리스가 아닌 석재 카운터)"
      ],
      "physics": "두 사람의 손 모두 투명 가림판에 지탱됨."
     },
     {
      "label": "B",
      "direction": "두 손이 가림판을 사이에 두고 마주 닿아 있음.",
      "built_space": "스테인리스 카운터와 가림판 구조가 로케이션 레퍼런스와 일치함.",
      "entities": "지국현과 신경희의 손 및 팔이 지정된 복장과 함께 등장함.",
      "hard_violations": [],
      "physics": "두 사람의 손 모두 투명 가림판 표면에 안정적으로 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1250,
      "verdict_ko": "클로즈업 프레이밍은 지시사항에 부합하나, 여성 손의 치명적인 해부학적 오류(오른팔에 왼손 결합)로 인해 실격입니다.  ★위반: [gemini-pro] 해부학적으로 불가능한 구조 (여성의 오른팔에 엄지손가락 방향이 반대인 왼손이 달려 있음)"
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "요구된 인서트 샷보다 넓은 프레이밍과 다른 카운터 재질을 보여주지만, 치명적 오류 없이 인물 식별과 동작을 안정적으로 구현했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.571
     },
     "adjusted": {
      "A": 1.25,
      "B": 1.571
     },
     "violations": {
      "A": [
       "[gemini-pro] 해부학적으로 불가능한 구조 (여성의 오른팔에 엄지손가락 방향이 반대인 왼손이 달려 있음)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1250,
      "verdict_ko": "클로즈업 프레이밍은 지시사항에 부합하나, 여성 손의 치명적인 해부학적 오류(오른팔에 왼손 결합)로 인해 실격입니다.  ★위반: [gemini-pro] 해부학적으로 불가능한 구조 (여성의 오른팔에 엄지손가락 방향이 반대인 왼손이 달려 있음)"
     },
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "요구된 인서트 샷보다 넓은 프레이밍과 다른 카운터 재질을 보여주지만, 치명적 오류 없이 인물 식별과 동작을 안정적으로 구현했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1574,
     "B": 1257
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "두 손이 가림판을 사이에 두고 맞닿음. 여성의 시선은 손을 향함.",
    "built_space": "면회실 가림판을 묘사했으나 카운터가 석재 질감으로 보임.",
    "entities": "지국현의 팔과 신경희의 얼굴/상반신이 보이며 인물 외형은 일치함.",
    "hard_violations": [
     "프레이밍 위반 (인서트 클로즈업이 아닌 상반신 포함 샷)",
     "로케이션 불일치 (스테인리스가 아닌 석재 카운터)"
    ],
    "physics": "두 사람의 손 모두 투명 가림판에 지탱됨."
   },
   {
    "label": "B",
    "direction": "두 손이 가림판을 사이에 두고 마주 닿아 있음.",
    "built_space": "스테인리스 카운터와 가림판 구조가 로케이션 레퍼런스와 일치함.",
    "entities": "지국현과 신경희의 손 및 팔이 지정된 복장과 함께 등장함.",
    "hard_violations": [],
    "physics": "두 사람의 손 모두 투명 가림판 표면에 안정적으로 지탱됨."
   }
  ],
  "totals": {
   "A": 1574,
   "B": 1257
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 인서트 클로즈업 프레이밍을 지키며 로케이션 사진의 스테인리스 카운터 재질을 정확히 구현함."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "프레이밍이 지시된 것보다 넓게 잡혔으며 로케이션의 카운터 재질(석재로 묘사)이 일치하지 않음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L45B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1176458>"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1236081>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "지정된 샷 배경(스테인리스 스틸 카운터, 연녹색 벽면 등)이 그대로 유지되지 않고, 재질(석재 카운터)과 색상이 다른 새로운 배경이 렌더링되었습니다.",
     "fix_en": "Change the counter to stainless steel and the lower walls to light green, preserving the characters, poses, clothing, and the transparent partition.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "프레이밍 스케일이 두 손에 집중하는 '인서트 클로즈업' 지시와 달리 너무 넓게 잡혀 배경 인물의 얼굴과 상반신이 크게 노출되었습니다.",
     "fix_en": "Darken the outer frame to draw focus to the hands, preserving the current camera position, character scale, poses, and background.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "왼쪽 신경희가 레퍼런스의 라운드넥 상의가 아닌 블레이저를 입고 있다.",
     "fix_en": "Change Gyeong-hui's clothing to a navy round-neck top, preserving her face, hand pose, the man's suit, and the partition.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "지정된 샷 배경(스테인리스 스틸 카운터, 연녹색 벽면 등)이 그대로 유지되지 않고, 재질(석재 카운터)과 색상이 다른 새로운 배경이 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "프레이밍 스케일이 두 손에 집중하는 '인서트 클로즈업' 지시와 달리 너무 넓게 잡혀 배경 인물의 얼굴과 상반신이 크게 노출되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "제공된 샷 배경의 스테인리스 카운터·원근·고정 시설이 돌 재질 선반과 다른 각도로 바뀌어 배경이 그대로 유지되지 않았다.",
     "severity": "major"
    },
    {
     "issue_ko": "손 높이의 인서트 클로즈업이어야 하는데 왼쪽 신경희의 얼굴과 상체가 크게 들어와 명시된 프레이밍 스케일이 아니다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 신경희가 레퍼런스의 라운드넥 상의가 아닌 블레이저를 입고 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "프레이밍 스케일이 두 손에 집중하는 '인서트 클로즈업' 지시와 달리 너무 넓게 잡혀 배경 인물의 얼굴과 상반신이 크게 노출되었습니다.",
    "fix_en": "Darken the outer frame to draw focus to the hands, preserving the current camera position, character scale, poses, and background.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "왼쪽 신경희가 레퍼런스의 라운드넥 상의가 아닌 블레이저를 입고 있다.",
    "fix_en": "Change Gyeong-hui's clothing to a navy round-neck top, preserving her face, hand pose, the man's suit, and the partition.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the counter to stainless steel and the lower walls to light green, preserving the characters, poses, clothing, and the transparent partition.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트가 지시한 비스듬한 앵글의 클로즈업 프레이밍과 인물 배치, 시선을 매우 충실히 구현했으나, 카운터의 재질이 스테인리스 스틸에서 석재로 변형된 점은 아쉬움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "배경 구조와 카운터 재질은 완벽하게 유지했으나, 두 인물의 위치가 반대로 바뀌었고 카메라를 정면으로 응시하여 스케치와 연기 지시를 완전히 무시함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 지국현을 향하고 있으며, 두 사람의 손은 유리 가림판을 사이에 두고 동일한 위치를 향해 맞닿아 있음.",
      "built_space": "가림판이 비스듬한 구도로 배치되어 있으며, 카운터 표면이 배경 참조 이미지의 금속이 아닌 석재 계열로 묘사됨.",
      "entities": "배경 좌측에 신경희(짙은 남색 상의), 전경 우측에 지국현(수트 소매)이 프롬프트의 위치대로 정확히 배치되었으며 인상착의가 일치함.",
      "hard_violations": [],
      "physics": "두 사람의 손은 투명한 가림판 표면에 의해 물리적으로 지지되고 있음."
     },
     {
      "label": "B",
      "direction": "두 사람 모두 손을 맞대고 있으나 시선은 서로를 향하지 않고 카메라 렌즈를 정면으로 응시하고 있음.",
      "built_space": "참조 이미지와 동일한 수평 구도의 유리 가림판, 스테인리스 스틸 카운터, 배경의 창문과 문이 정확하게 묘사됨.",
      "entities": "지국현이 배경 좌측에, 신경희가 전경 우측에 배치되어 텍스트 및 스케치가 요구한 인물 위치와 완전히 반대로 적용됨.",
      "hard_violations": [],
      "physics": "손바닥이 가림판 유리에 닿아 지지되고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트가 지시한 비스듬한 앵글의 클로즈업 프레이밍과 인물 배치, 시선을 매우 충실히 구현했으나, 카운터의 재질이 스테인리스 스틸에서 석재로 변형된 점은 아쉬움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "배경 구조와 카운터 재질은 완벽하게 유지했으나, 두 인물의 위치가 반대로 바뀌었고 카메라를 정면으로 응시하여 스케치와 연기 지시를 완전히 무시함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 지국현을 향하고 있으며, 두 사람의 손은 유리 가림판을 사이에 두고 동일한 위치를 향해 맞닿아 있음.",
      "built_space": "가림판이 비스듬한 구도로 배치되어 있으며, 카운터 표면이 배경 참조 이미지의 금속이 아닌 석재 계열로 묘사됨.",
      "entities": "배경 좌측에 신경희(짙은 남색 상의), 전경 우측에 지국현(수트 소매)이 프롬프트의 위치대로 정확히 배치되었으며 인상착의가 일치함.",
      "hard_violations": [],
      "physics": "두 사람의 손은 투명한 가림판 표면에 의해 물리적으로 지지되고 있음."
     },
     {
      "label": "B",
      "direction": "두 사람 모두 손을 맞대고 있으나 시선은 서로를 향하지 않고 카메라 렌즈를 정면으로 응시하고 있음.",
      "built_space": "참조 이미지와 동일한 수평 구도의 유리 가림판, 스테인리스 스틸 카운터, 배경의 창문과 문이 정확하게 묘사됨.",
      "entities": "지국현이 배경 좌측에, 신경희가 전경 우측에 배치되어 텍스트 및 스케치가 요구한 인물 위치와 완전히 반대로 적용됨.",
      "hard_violations": [],
      "physics": "손바닥이 가림판 유리에 닿아 지지되고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 인서트 클로즈업 프레이밍과 레이아웃 스케치의 구도를 정확히 따랐으며, 전경과 배경의 인물 배치 및 손을 포개는 동작이 요청사항과 잘 일치합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "요구된 손 클로즈업 대신 미디엄 샷으로 연출되었고, 두 인물의 전경/배경 위치가 서로 뒤바뀌었으며, 부자연스럽게 카메라를 응시하고 있어 주요 지시사항을 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "배경의 여성은 유리창에 맞닿은 손을 향해 시선을 두고 있습니다.",
      "built_space": "교도소 면회실의 구조와 투명 가림판이 배경 레퍼런스와 일치하게 배치되어 있습니다.",
      "entities": "배경 좌측에 신경희의 얼굴과 상반신이, 전경 우측에 지국현의 팔과 손이 지시된 대로 위치하며 신원과 복장이 일치합니다.",
      "hard_violations": [],
      "physics": "두 사람의 손이 투명 가림판 표면에 의해 물리적으로 지지되고 있습니다."
     },
     {
      "label": "A",
      "direction": "두 인물 모두 정면의 카메라 렌즈를 똑바로 응시하고 있습니다.",
      "built_space": "면회실 가림판이 존재하나, 카메라 앵글과 프레이밍이 지시된 샷과 완전히 다릅니다.",
      "entities": "지국현(남성)과 신경희(여성)의 신원과 복장은 맞으나, 지국현이 배경 좌측으로, 신경희가 전경 우측으로 위치가 서로 바뀌어 있습니다.",
      "hard_violations": [
       "지시된 프레이밍 스케일 위반 (클로즈업 샷이 아닌 미디엄 샷)",
       "인물 위치 위반 (전경과 배경의 인물 배치가 스케치 및 프롬프트와 반대로 뒤바뀜)",
       "시선 처리 위반 (상황에 맞지 않게 카메라를 응시함)"
      ],
      "physics": "두 사람의 손이 가림판에 지지되어 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 인서트 클로즈업 프레이밍과 레이아웃 스케치의 구도를 정확히 따랐으며, 전경과 배경의 인물 배치 및 손을 포개는 동작이 요청사항과 잘 일치합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 손 클로즈업 대신 미디엄 샷으로 연출되었고, 두 인물의 전경/배경 위치가 서로 뒤바뀌었으며, 부자연스럽게 카메라를 응시하고 있어 주요 지시사항을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "배경의 여성은 유리창에 맞닿은 손을 향해 시선을 두고 있습니다.",
      "built_space": "교도소 면회실의 구조와 투명 가림판이 배경 레퍼런스와 일치하게 배치되어 있습니다.",
      "entities": "배경 좌측에 신경희의 얼굴과 상반신이, 전경 우측에 지국현의 팔과 손이 지시된 대로 위치하며 신원과 복장이 일치합니다.",
      "hard_violations": [],
      "physics": "두 사람의 손이 투명 가림판 표면에 의해 물리적으로 지지되고 있습니다."
     },
     {
      "label": "B",
      "direction": "두 인물 모두 정면의 카메라 렌즈를 똑바로 응시하고 있습니다.",
      "built_space": "면회실 가림판이 존재하나, 카메라 앵글과 프레이밍이 지시된 샷과 완전히 다릅니다.",
      "entities": "지국현(남성)과 신경희(여성)의 신원과 복장은 맞으나, 지국현이 배경 좌측으로, 신경희가 전경 우측으로 위치가 서로 바뀌어 있습니다.",
      "hard_violations": [
       "지시된 프레이밍 스케일 위반 (클로즈업 샷이 아닌 미디엄 샷)",
       "인물 위치 위반 (전경과 배경의 인물 배치가 스케치 및 프롬프트와 반대로 뒤바뀜)",
       "시선 처리 위반 (상황에 맞지 않게 카메라를 응시함)"
      ],
      "physics": "두 사람의 손이 가림판에 지지되어 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S63sh5__bgfirst_bg.png",
   "bg_asset_id": "e0f1a0c4-3ce1-481a-9ed6-1655e6fd9fdc",
   "bg_record_key": "S63sh5::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S63sh5::cine": {
  "applied": true,
  "fingerprint": "c4a545313beb466810ca36da056f9df3b0b20228de03d131e224aae3fd6761cb",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S63sh5_sel.png",
  "source_sha256": "480b3716aa8df4cae47c86b307f39d20f98296fe29b2a3df7cbb70ebe14aa2ce",
  "file": "S63sh5_cine.png",
  "latency_ms": 8532
 },
 "S63sh7::signage": {
  "fp": "0a5059cb25b759df",
  "inscriptions": []
 },
 "S63sh7": {
  "input_fingerprint": "22d8aa9a806b9bd9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현을 향해 씩 웃어 보이는 신경희의 편안해진 얼굴 클로즈업\n\nLOCATION (lock): On the visitor's side of the transparent partition in the prison visiting room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height just behind and beside 지국현's shoulder, the camera finishes its dolly-in through the transparent divider on a close, off-axis three-quarter view of 신경희. Her relaxed face occupies the center-left as she looks directly toward 지국현 within the scene, while his shoulder and partial head remain a narrow foreground presence at the right edge; the frame holds until her smile fully settles.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between them) — Its surface lies between the lens and 신경희, with her face visible through it and 지국현 on the near side; used as The camera looks through it, retaining a subtle physical separation between 신경희 and 지국현.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination keeps 신경희's softened expression natural, subdued, and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현을 향해 씩 웃어 보이는 신경희의 편안해진 얼굴 클로즈업\n\nLOCATION (lock): On the visitor's side of the transparent partition in the prison visiting room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height just behind and beside 지국현's shoulder, the camera finishes its dolly-in through the transparent divider on a close, off-axis three-quarter view of 신경희. Her relaxed face occupies the center-left as she looks directly toward 지국현 within the scene, while his shoulder and partial head remain a narrow foreground presence at the right edge; the frame holds until her smile fully settles.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between them) — Its surface lies between the lens and 신경희, with her face visible through it and 지국현 on the near side; used as The camera looks through it, retaining a subtle physical separation between 신경희 and 지국현.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination keeps 신경희's softened expression natural, subdued, and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현을 향해 씩 웃어 보이는 신경희의 편안해진 얼굴 클로즈업\n\nLOCATION (lock): On the visitor's side of the transparent partition in the prison visiting room. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height just behind and beside 지국현's shoulder, the camera finishes its dolly-in through the transparent divider on a close, off-axis three-quarter view of 신경희. Her relaxed face occupies the center-left as she looks directly toward 지국현 within the scene, while his shoulder and partial head remain a narrow foreground presence at the right edge; the frame holds until her smile fully settles.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: transparent plastic divider (fixed between them) — Its surface lies between the lens and 신경희, with her face visible through it and 지국현 on the near side; used as The camera looks through it, retaining a subtle physical separation between 신경희 and 지국현.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime visitation-room illumination keeps 신경희's softened expression natural, subdued, and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "신경희는 오른쪽 전경에 위치한 지국현의 얼굴 쪽을 향해 정확히 시선을 맞추며 미소 짓고 있습니다.",
    "built_space": "이전 샷 스틸과 동일하게 시야를 가리는 수직 프레임 없는 통유리 칸막이가 카메라와 신경희 사이에 위치하며, 뒷배경의 투톤 벽면, 창살 있는 창문, 닫힌 문 등 접견실 내부 구조가 정확히 배치되어 있습니다.",
    "entities": "신경희의 얼굴과 헤어스타일은 레퍼런스와 일치하나, 의상은 이전 샷 스틸(남색 상의)로 고정하라는 지시와 달리 패턴 상의를 입고 있습니다. 오른쪽 전경에 지국현의 정장 입은 어깨와 머리 일부가 올바르게 자리 잡고 있습니다.",
    "hard_violations": [],
    "physics": "인물의 앉은 자세와 머리의 기울기가 자연스럽고, 지면과 의자에 닿아 있는 몸의 무게 중심 등 물리적 법칙에 위배되는 요소 없이 안정적으로 표현되었습니다."
   },
   {
    "label": "A",
    "direction": "신경희는 화면 우측 전경의 지국현을 향해 부드러운 시선을 보내고 있습니다.",
    "built_space": "이전 샷 스틸에 존재하지 않았던 두꺼운 수직 금속 프레임이 유리 칸막이 중간에 추가되어 물리적 구조가 훼손되었고, 우측 뒷배경의 문 형태도 기존과 다르게 변형되었습니다.",
    "entities": "신경희의 외모는 레퍼런스와 일치하지만 B와 마찬가지로 이전 샷 스틸의 의상을 따르지 않았습니다. 지국현의 머리와 어깨가 프롬프트대로 오른쪽 프레임 가장자리에 배치되었습니다.",
    "hard_violations": [],
    "physics": "어색하게 떠 있는 사물이나 신체 부위 없이 인물의 자세와 옷 주름이 중력에 맞게 자연스럽게 지지되고 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "이전 샷 스틸의 공간 구조(통유리, 배경 문과 벽)를 정확히 유지하고 유리창 표면 질감과 클로즈업 앵글을 훌륭하게 구현했습니다(단, 두 후보 모두 이전 샷 의상 고정 지시는 미준수)."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "이전 샷 스틸에 없던 수직 금속 프레임이 창 한가운데 생성되어 장소 연속성이 깨졌으며, 프레이밍이 지시된 클로즈업보다 다소 멉니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 신경희 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S63sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:843594>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "신경희의 옷이 이전 샷(검은색 정장)과 다름(참조 이미지의 사복 착용)",
     "fix_en": "Change the woman's patterned blouse and scarf to a dark navy blazer over a dark navy top. Preserve the people present and their positions, the man's clothing, the prison visitation room set, the transparent divider, the lighting, and the framing.",
     "severity": "critical",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "신경희의 옷이 이전 샷(검은색 정장)과 다름(참조 이미지의 사복 착용)",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 샷에 고정된 네이비 블레이저가 아니라 무늬 블라우스와 스카프를 입은 신경희가 화면 중앙-왼쪽에 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the woman's patterned blouse and scarf to a dark navy blazer over a dark navy top. Preserve the people present and their positions, the man's clothing, the prison visitation room set, the transparent divider, the lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 샷 스틸에 지정된 의상(남색 정장)을 정확히 유지하면서, 지시된 편안한 미소와 클로즈업 앵글을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 표정 연출은 우수하나, 이전 샷에 고정되어야 할 의상 대신 캐릭터 레퍼런스의 패턴 의상을 착용하여 연속성 지침을 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "신경희의 시선은 화면 우측 전경에 있는 지국현을 향하고 있습니다.",
      "built_space": "두 인물 사이에 투명한 플라스틱 칸막이가 위치하며 표면 반사광이 보입니다. 배경에 면회실의 구조가 잘 나타납니다.",
      "entities": "신경희의 얼굴이 레퍼런스와 일치하며, 지침대로 이전 샷 스틸의 남색 정장을 입고 있습니다. 화면 우측에 지국현의 어깨와 뒷머리가 배치되어 있습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 안정적으로 자리에 앉아 있는 자세이며, 물리적 오류 없이 지지되고 있습니다."
     },
     {
      "label": "A",
      "direction": "신경희의 시선이 화면 우측 전경의 지국현을 자연스럽게 향하고 있습니다.",
      "built_space": "두 인물 사이의 투명 칸막이와 배경의 면회실 문, 창문 구조가 정상적으로 배치되어 있습니다.",
      "entities": "신경희의 얼굴은 일치하나, 이전 샷 스틸에 고정된 의상이 아닌 캐릭터 레퍼런스의 패턴 상의를 입고 있습니다. 지국현은 우측 가장자리에 잘림 형태로 나타납니다.",
      "hard_violations": [],
      "physics": "자세와 신체 비율이 자연스러우며 떠 있거나 불안정한 부분 없이 잘 지지되어 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 샷 스틸에 지정된 의상(남색 정장)을 정확히 유지하면서, 지시된 편안한 미소와 클로즈업 앵글을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 표정 연출은 우수하나, 이전 샷에 고정되어야 할 의상 대신 캐릭터 레퍼런스의 패턴 의상을 착용하여 연속성 지침을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "신경희의 시선은 화면 우측 전경에 있는 지국현을 향하고 있습니다.",
      "built_space": "두 인물 사이에 투명한 플라스틱 칸막이가 위치하며 표면 반사광이 보입니다. 배경에 면회실의 구조가 잘 나타납니다.",
      "entities": "신경희의 얼굴이 레퍼런스와 일치하며, 지침대로 이전 샷 스틸의 남색 정장을 입고 있습니다. 화면 우측에 지국현의 어깨와 뒷머리가 배치되어 있습니다.",
      "hard_violations": [],
      "physics": "두 인물 모두 안정적으로 자리에 앉아 있는 자세이며, 물리적 오류 없이 지지되고 있습니다."
     },
     {
      "label": "A",
      "direction": "신경희의 시선이 화면 우측 전경의 지국현을 자연스럽게 향하고 있습니다.",
      "built_space": "두 인물 사이의 투명 칸막이와 배경의 면회실 문, 창문 구조가 정상적으로 배치되어 있습니다.",
      "entities": "신경희의 얼굴은 일치하나, 이전 샷 스틸에 고정된 의상이 아닌 캐릭터 레퍼런스의 패턴 상의를 입고 있습니다. 지국현은 우측 가장자리에 잘림 형태로 나타납니다.",
      "hard_violations": [],
      "physics": "자세와 신체 비율이 자연스러우며 떠 있거나 불안정한 부분 없이 잘 지지되어 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 프레이밍과 표정을 훌륭하게 연출했으며, 특히 프롬프트가 강조한 '이전 샷 스틸'과의 의상 연속성(짙은 파란색 정장)을 정확히 유지하여 지침을 충실히 이행했습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "표정과 구도는 요구사항에 부합하나, '이전 샷 스틸'의 의상으로 고정하라는 명시적인 연속성 지침을 어기고 캐릭터 레퍼런스의 의상을 적용한 점이 치명적인 감점 요인입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "신경희의 시선은 프레임 오른쪽 전경에 위치한 지국현의 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "교도소 면회실 내부. 두 사람 사이에 투명한 플라스틱 칸막이가 있으며, 표면에 희미한 반사가 맺혀 있어 위치와 질감이 올바르게 읽힘.",
      "entities": "신경희의 얼굴과 헤어스타일이 레퍼런스와 일치하며, 의상은 '이전 샷 스틸'에 맞춰 짙은 파란색 정장을 입고 있음. 오른쪽 가장자리에 지국현의 어깨와 머리 일부가 올바른 비율로 존재함.",
      "hard_violations": [],
      "physics": "두 인물 모두 화면 밖의 의자나 바닥에 안정적으로 지지된 자세를 취하고 있으며 물리적 오류 없음."
     },
     {
      "label": "B",
      "direction": "신경희의 시선은 프레임 오른쪽 전경에 위치한 지국현을 향하고 있음.",
      "built_space": "교도소 면회실 내부. 두 사람 사이에 투명한 칸막이가 존재하며 반사 디테일이 보임.",
      "entities": "신경희의 신원과 외모는 일치하나, 의상이 '이전 샷 스틸'의 파란색 정장이 아닌 '캐릭터 레퍼런스'의 무늬 있는 셔츠와 스카프로 변경되어 연속성 지침을 위반함. 오른쪽 가장자리에 지국현의 뒷모습이 일부 보임.",
      "hard_violations": [],
      "physics": "두 인물 모두 물리적으로 올바른 자세를 유지하며 안착해 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 프레이밍과 표정을 훌륭하게 연출했으며, 특히 프롬프트가 강조한 '이전 샷 스틸'과의 의상 연속성(짙은 파란색 정장)을 정확히 유지하여 지침을 충실히 이행했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "표정과 구도는 요구사항에 부합하나, '이전 샷 스틸'의 의상으로 고정하라는 명시적인 연속성 지침을 어기고 캐릭터 레퍼런스의 의상을 적용한 점이 치명적인 감점 요인입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "신경희의 시선은 프레임 오른쪽 전경에 위치한 지국현의 얼굴 쪽을 정확히 향하고 있음.",
      "built_space": "교도소 면회실 내부. 두 사람 사이에 투명한 플라스틱 칸막이가 있으며, 표면에 희미한 반사가 맺혀 있어 위치와 질감이 올바르게 읽힘.",
      "entities": "신경희의 얼굴과 헤어스타일이 레퍼런스와 일치하며, 의상은 '이전 샷 스틸'에 맞춰 짙은 파란색 정장을 입고 있음. 오른쪽 가장자리에 지국현의 어깨와 머리 일부가 올바른 비율로 존재함.",
      "hard_violations": [],
      "physics": "두 인물 모두 화면 밖의 의자나 바닥에 안정적으로 지지된 자세를 취하고 있으며 물리적 오류 없음."
     },
     {
      "label": "A",
      "direction": "신경희의 시선은 프레임 오른쪽 전경에 위치한 지국현을 향하고 있음.",
      "built_space": "교도소 면회실 내부. 두 사람 사이에 투명한 칸막이가 존재하며 반사 디테일이 보임.",
      "entities": "신경희의 신원과 외모는 일치하나, 의상이 '이전 샷 스틸'의 파란색 정장이 아닌 '캐릭터 레퍼런스'의 무늬 있는 셔츠와 스카프로 변경되어 연속성 지침을 위반함. 오른쪽 가장자리에 지국현의 뒷모습이 일부 보임.",
      "hard_violations": [],
      "physics": "두 인물 모두 물리적으로 올바른 자세를 유지하며 안착해 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S63sh5"
  }
 },
 "S63sh7::cine": {
  "applied": true,
  "fingerprint": "fc219afa3e02f39524f1bfe9da5868bdf4ca9d274f05fd8fef3d54d4cb2eeec6",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S63sh7_sel.png",
  "source_sha256": "8d2b327197c053f456ad039a32e7eebfca1d0552aa92cf4949f5f7e6d0ba961c",
  "file": "S63sh7_cine.png",
  "latency_ms": 9663
 },
 "S64sh2::signage": {
  "fp": "08fa971636b413d3",
  "inscriptions": [
   {
    "surface_native": "펼쳐진 성경 소책자 페이지",
    "text_native": "수고하고 무거운 짐 진 자들아 다 내게로 오라 내가 너희를 쉬게 하리라",
    "reason_ko": "독방에 갇힌 인물이 펼쳐 든 성경 소책자의 내용을 사실적으로 묘사하기 위해 한국어 성경 구절이 필요함."
   }
  ]
 },
 "S64sh2": {
  "input_fingerprint": "041cade948383d31",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 양손으로 낡은 성경 소책자를 펼쳐 든 지국현의 상체\n\nLOCATION (lock): Inside the sparse prison solitary cell, beside the sleeping platform. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From 지국현's three-quarter side at upper-torso distance, the camera sits slightly above his hands and tilts down from his expression toward the open Bible booklet. His upper body occupies the left half while both hands support the booklet across the lower center; his attention remains lowered to the inserted memo without changing the established side axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: worn Bible booklet (open in both hands) — The open page faces upward and partly toward the camera, showing printed content and the inserted memo; used as Primary object revealed beneath 지국현's lowered gaze, with its open spread kept in the lower center; parole-review memo (inserted between pages and densely written) — Its densely written face is turned upward within the open booklet; used as Visible between the booklet pages as the next focal destination.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime cell illumination is restrained and neutral-to-cool, preserving a natural tonal range across 지국현 and the open pages.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon holds the worn Bible booklet open, with the parole-review memorandum still inserted between its pages.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 지국현 right now — 지국현's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 지국현 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 펼쳐진 성경 소책자 페이지: \"수고하고 무거운 짐 진 자들아 다 내게로 오라 내가 너희를 쉬게 하리라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 양손으로 낡은 성경 소책자를 펼쳐 든 지국현의 상체\n\nLOCATION (lock): Inside the sparse prison solitary cell, beside the sleeping platform. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From 지국현's three-quarter side at upper-torso distance, the camera sits slightly above his hands and tilts down from his expression toward the open Bible booklet. His upper body occupies the left half while both hands support the booklet across the lower center; his attention remains lowered to the inserted memo without changing the established side axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: worn Bible booklet (open in both hands) — The open page faces upward and partly toward the camera, showing printed content and the inserted memo; used as Primary object revealed beneath 지국현's lowered gaze, with its open spread kept in the lower center; parole-review memo (inserted between pages and densely written) — Its densely written face is turned upward within the open booklet; used as Visible between the booklet pages as the next focal destination.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime cell illumination is restrained and neutral-to-cool, preserving a natural tonal range across 지국현 and the open pages.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon holds the worn Bible booklet open, with the parole-review memorandum still inserted between its pages.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 지국현 right now — 지국현's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 지국현 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 펼쳐진 성경 소책자 페이지: \"수고하고 무거운 짐 진 자들아 다 내게로 오라 내가 너희를 쉬게 하리라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 양손으로 낡은 성경 소책자를 펼쳐 든 지국현의 상체\n\nLOCATION (lock): Inside the sparse prison solitary cell, beside the sleeping platform. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From 지국현's three-quarter side at upper-torso distance, the camera sits slightly above his hands and tilts down from his expression toward the open Bible booklet. His upper body occupies the left half while both hands support the booklet across the lower center; his attention remains lowered to the inserted memo without changing the established side axis.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: worn Bible booklet (open in both hands) — The open page faces upward and partly toward the camera, showing printed content and the inserted memo; used as Primary object revealed beneath 지국현's lowered gaze, with its open spread kept in the lower center; parole-review memo (inserted between pages and densely written) — Its densely written face is turned upward within the open booklet; used as Visible between the booklet pages as the next focal destination.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime cell illumination is restrained and neutral-to-cool, preserving a natural tonal range across 지국현 and the open pages.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Ji Gukhyeon holds the worn Bible booklet open, with the parole-review memorandum still inserted between its pages.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 지국현 right now — 지국현's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 지국현 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 펼쳐진 성경 소책자 페이지: \"수고하고 무거운 짐 진 자들아 다 내게로 오라 내가 너희를 쉬게 하리라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.667
   },
   "violations": {
    "A": [
     "[gemini-pro] 양손으로 책을 들고 있어야 하나 왼손이 누락되었으며, 책의 오른쪽 절반이 아무런 지지 없이 허공에 수평으로 떠 있습니다."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1667,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "이전 샷의 배경과 조명을 잘 유지했으며 양손으로 책을 받치는 지시를 따랐으나, 삽입된 메모 묘사가 누락되었습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "요구된 텍스트는 정확히 반영되었으나, 한 손만으로 책을 들고 있어 나머지 반쪽이 허공에 떠 있는 물리적 오류가 발생했습니다.  ★위반: [gemini-pro] 양손으로 책을 들고 있어야 하나 왼손이 누락되었으며, 책의 오른쪽 절반이 아무런 지지 없이 허공에 수평으로 떠 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S58sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   },
   {
    "label": "PROP REFERENCE — 낡은 성경 소책자: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:910177>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "지국현이 양손으로 성경을 들고 있어야 한다는 지시와 달리 한 손으로만 책을 받치고 있음 (나머지 한 손은 보이지 않음).",
     "fix_en": "Draw a right hand holding the right edge of the book so both hands support it. Maintain the character's face, left hand, uniform, book, and background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "펼쳐진 성경책 사이에 끼워져 있어야 할 '가석방 심사 메모(parole-review memo)'가 화면에 존재하지 않음.",
     "fix_en": "Insert a densely written paper memo between the book pages. Maintain the character, hands, uniform, book cover, and background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "성경 소책자의 펼쳐진 면이 일반적인 인쇄 내용 없이 백지로 되어 있으며, 지시된 문구만 페이지 중간에 부자연스럽게 배치됨.",
     "fix_en": "Fill the blank book pages with dense, illegible printed text lines around the main verse. Maintain the character, hands, uniform, book position, and background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "가슴 패치가 캐릭터 참조의 '교도소 1234'가 아니라 '4021'이다",
     "fix_en": "Blur the text on the chest patch into illegible strokes. Maintain the character, hands, uniform, book, and background.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "지국현이 양손으로 성경을 들고 있어야 한다는 지시와 달리 한 손으로만 책을 받치고 있음 (나머지 한 손은 보이지 않음).",
     "severity": "critical"
    },
    {
     "issue_ko": "펼쳐진 성경책 사이에 끼워져 있어야 할 '가석방 심사 메모(parole-review memo)'가 화면에 존재하지 않음.",
     "severity": "critical"
    },
    {
     "issue_ko": "성경 소책자의 펼쳐진 면이 일반적인 인쇄 내용 없이 백지로 되어 있으며, 지시된 문구만 페이지 중간에 부자연스럽게 배치됨.",
     "severity": "major"
    },
    {
     "issue_ko": "펼친 책 페이지 사이에 위로 향한 빽빽한 가석방 심사 메모가 없다",
     "severity": "critical"
    },
    {
     "issue_ko": "손에 든 책이 참조의 낡은 성경 소책자가 아니라 거의 빈 페이지의 다른 양장본이다",
     "severity": "major"
    },
    {
     "issue_ko": "펼친 면의 성경 구절이 실제 활자가 아니라 빈 지면 위 큰 오버레이 글씨처럼 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "가슴 패치가 캐릭터 참조의 '교도소 1234'가 아니라 '4021'이다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "성경 소책자의 펼쳐진 면이 일반적인 인쇄 내용 없이 백지로 되어 있으며, 지시된 문구만 페이지 중간에 부자연스럽게 배치됨.",
    "fix_en": "Fill the blank book pages with dense, illegible printed text lines around the main verse. Maintain the character, hands, uniform, book position, and background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "가슴 패치가 캐릭터 참조의 '교도소 1234'가 아니라 '4021'이다",
    "fix_en": "Blur the text on the chest patch into illegible strokes. Maintain the character, hands, uniform, book, and background.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Draw a right hand holding the right edge of the book so both hands support it. Maintain the character's face, left hand, uniform, book, and background.\n- Insert a densely written paper memo between the book pages. Maintain the character, hands, uniform, book cover, and background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 한국어 텍스트를 성경 페이지에 정확히 구현하고 소품을 책으로 교체했으나, 삽입된 가석방 심사 메모가 누락됨."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "이전 샷의 사진첩을 그대로 들고 있으며, 요구된 성경책 소품과 텍스트를 전혀 반영하지 못한 치명적인 실패작."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 양손으로 든 펼쳐진 책(성경)을 향해 아래로 고정됨.",
      "built_space": "캐비닛, 바인더, 열린 복도 문, 좌측의 취침용 평상 등 이전 샷의 감방 구조를 동일하게 유지함.",
      "entities": "지국현(참조 이미지 일치), 죄수복. 양손에 펼쳐진 책이 있으며 오른쪽 페이지에 요구된 텍스트('수고하고 무거운 짐 진 자들아...')가 정확히 적혀 있으나 밀도 있게 쓰인 메모지는 보이지 않음.",
      "hard_violations": [],
      "physics": "양손이 무거운 책의 하단을 안정적으로 받치고 있음."
     },
     {
      "label": "B",
      "direction": "시선은 양손으로 든 펼쳐진 사진첩을 향해 아래로 고정됨.",
      "built_space": "캐비닛과 열린 복도 문 등 이전 샷의 감방 구조가 유지됨.",
      "entities": "지국현(참조 이미지 일치), 죄수복. 성경책 대신 이전 샷의 사진첩을 그대로 들고 있으며, 사진 위에 알아볼 수 없는 낙서가 된 종이가 덧대어져 있음. 지정된 텍스트 누락.",
      "hard_violations": [],
      "physics": "양손이 사진첩 하단을 자연스럽게 받치고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 한국어 텍스트를 성경 페이지에 정확히 구현하고 소품을 책으로 교체했으나, 삽입된 가석방 심사 메모가 누락됨."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "이전 샷의 사진첩을 그대로 들고 있으며, 요구된 성경책 소품과 텍스트를 전혀 반영하지 못한 치명적인 실패작."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 양손으로 든 펼쳐진 책(성경)을 향해 아래로 고정됨.",
      "built_space": "캐비닛, 바인더, 열린 복도 문, 좌측의 취침용 평상 등 이전 샷의 감방 구조를 동일하게 유지함.",
      "entities": "지국현(참조 이미지 일치), 죄수복. 양손에 펼쳐진 책이 있으며 오른쪽 페이지에 요구된 텍스트('수고하고 무거운 짐 진 자들아...')가 정확히 적혀 있으나 밀도 있게 쓰인 메모지는 보이지 않음.",
      "hard_violations": [],
      "physics": "양손이 무거운 책의 하단을 안정적으로 받치고 있음."
     },
     {
      "label": "B",
      "direction": "시선은 양손으로 든 펼쳐진 사진첩을 향해 아래로 고정됨.",
      "built_space": "캐비닛과 열린 복도 문 등 이전 샷의 감방 구조가 유지됨.",
      "entities": "지국현(참조 이미지 일치), 죄수복. 성경책 대신 이전 샷의 사진첩을 그대로 들고 있으며, 사진 위에 알아볼 수 없는 낙서가 된 종이가 덧대어져 있음. 지정된 텍스트 누락.",
      "hard_violations": [],
      "physics": "양손이 사진첩 하단을 자연스럽게 받치고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "이전 컷의 사진첩을 성경책으로 교체하지 못한 채 정체불명의 종이만 덧붙여졌으며, 필수 텍스트 렌더링에 완전히 실패했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "사진첩을 레퍼런스와 유사한 낡은 성경책으로 성공적으로 교체하였고, 페이지 내 지정된 한국어 문구를 정확하게 구현하여 프롬프트 지시를 훌륭히 충족했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 손에 들린 책자(사진첩)를 향하고 있습니다.",
      "built_space": "교도소 독방 내부, 뒷배경의 관물대와 열린 문 너머의 복도가 이전 컷과 동일하게 위치합니다.",
      "entities": "지국현의 외모와 복장은 일치하나, 손에 든 물건이 낡은 성경책이 아닌 이전 컷의 사진첩에 형태를 알 수 없는 종이가 합쳐진 모습이며, 지정된 문구가 없습니다.",
      "hard_violations": [],
      "physics": "양손으로 책자를 안정적으로 받치고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 손에 펼쳐든 낡은 책자를 향하고 있습니다.",
      "built_space": "교도소 독방 내부, 관물대와 복도 등 배경 요소가 레퍼런스와 일치합니다.",
      "entities": "지국현의 외모와 복장 일치. 손에 든 물건이 낡은 성경책으로 올바르게 묘사되었으며, 펼쳐진 페이지에 '수고하고 무거운 짐 진 자들아...' 문구가 정확히 적혀 있습니다.",
      "hard_violations": [],
      "physics": "양손으로 성경책의 양쪽을 안정적으로 쥐고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "이전 컷의 사진첩을 성경책으로 교체하지 못한 채 정체불명의 종이만 덧붙여졌으며, 필수 텍스트 렌더링에 완전히 실패했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "사진첩을 레퍼런스와 유사한 낡은 성경책으로 성공적으로 교체하였고, 페이지 내 지정된 한국어 문구를 정확하게 구현하여 프롬프트 지시를 훌륭히 충족했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물의 시선이 손에 들린 책자(사진첩)를 향하고 있습니다.",
      "built_space": "교도소 독방 내부, 뒷배경의 관물대와 열린 문 너머의 복도가 이전 컷과 동일하게 위치합니다.",
      "entities": "지국현의 외모와 복장은 일치하나, 손에 든 물건이 낡은 성경책이 아닌 이전 컷의 사진첩에 형태를 알 수 없는 종이가 합쳐진 모습이며, 지정된 문구가 없습니다.",
      "hard_violations": [],
      "physics": "양손으로 책자를 안정적으로 받치고 있습니다."
     },
     {
      "label": "A",
      "direction": "인물의 시선이 손에 펼쳐든 낡은 책자를 향하고 있습니다.",
      "built_space": "교도소 독방 내부, 관물대와 복도 등 배경 요소가 레퍼런스와 일치합니다.",
      "entities": "지국현의 외모와 복장 일치. 손에 든 물건이 낡은 성경책으로 올바르게 묘사되었으며, 펼쳐진 페이지에 '수고하고 무거운 짐 진 자들아...' 문구가 정확히 적혀 있습니다.",
      "hard_violations": [],
      "physics": "양손으로 성경책의 양쪽을 안정적으로 쥐고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S58sh4"
  }
 },
 "S64sh2::cine": {
  "applied": true,
  "fingerprint": "34a112862f63833f94aa9f8bc108dbbba22635a76eab0b5d161e47e5176dc4eb",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S64sh2_sel.png",
  "source_sha256": "e003a83329c9b7fef3fab25b6bc17043323377c84b1ade2ac57a964cb6343083",
  "file": "S64sh2_cine.png",
  "latency_ms": 10898
 },
 "S64sh3::signage": {
  "fp": "0fa1d6c42a906364",
  "inscriptions": [
   {
    "surface_native": "낡은 종이 쪽지",
    "text_native": "두려워하지 말라",
    "reason_ko": "독방 안 성경책 사이에 끼워진 메모지에 적힌 손글씨로, 수감자가 스스로를 위로하기 위해 적어둔 성경 구절의 일부를 표현합니다."
   }
  ]
 },
 "S64sh3": {
  "input_fingerprint": "371b8ba64cb43a22",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성경책 사이에 끼워진 낡은 메모 쪽지 표면 클로즈업\n\nLOCATION (lock): Inside the solitary cell, over the opened Bible booklet held near the sleeping platform. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Directly above and slightly offset from the open booklet, the push-in ends on a steep insert of the worn memo resting between the pages. The memo occupies roughly the central lower third rather than overwhelming the frame, while surrounding printed pages and booklet edges preserve scale and context before the hard cut.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: worn memo note (inserted between the pages and densely written) — The densely written front face is fully turned toward the overhead camera, showing parole timing, review conditions, and model-prisoner records; used as Primary focal detail, framed closely enough for the dense parole-review notes to register while remaining bounded by the booklet; open Bible booklet pages (open) — The printed page faces upward around the memo, with both page margins visible; used as Surrounds the memo as contextual scale and identifies where the note was concealed.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime cell illumination remains even and restrained, with neutral-to-cool color and sufficient tonal clarity for the written surface.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old memorandum remains tucked inside the Bible and is densely covered with parole dates, review conditions, and model-prisoner records.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 낡은 종이 쪽지: \"두려워하지 말라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성경책 사이에 끼워진 낡은 메모 쪽지 표면 클로즈업\n\nLOCATION (lock): Inside the solitary cell, over the opened Bible booklet held near the sleeping platform. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Directly above and slightly offset from the open booklet, the push-in ends on a steep insert of the worn memo resting between the pages. The memo occupies roughly the central lower third rather than overwhelming the frame, while surrounding printed pages and booklet edges preserve scale and context before the hard cut.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: worn memo note (inserted between the pages and densely written) — The densely written front face is fully turned toward the overhead camera, showing parole timing, review conditions, and model-prisoner records; used as Primary focal detail, framed closely enough for the dense parole-review notes to register while remaining bounded by the booklet; open Bible booklet pages (open) — The printed page faces upward around the memo, with both page margins visible; used as Surrounds the memo as contextual scale and identifies where the note was concealed.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime cell illumination remains even and restrained, with neutral-to-cool color and sufficient tonal clarity for the written surface.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old memorandum remains tucked inside the Bible and is densely covered with parole dates, review conditions, and model-prisoner records.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 낡은 종이 쪽지: \"두려워하지 말라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성경책 사이에 끼워진 낡은 메모 쪽지 표면 클로즈업\n\nLOCATION (lock): Inside the solitary cell, over the opened Bible booklet held near the sleeping platform. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Directly above and slightly offset from the open booklet, the push-in ends on a steep insert of the worn memo resting between the pages. The memo occupies roughly the central lower third rather than overwhelming the frame, while surrounding printed pages and booklet edges preserve scale and context before the hard cut.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: worn memo note (inserted between the pages and densely written) — The densely written front face is fully turned toward the overhead camera, showing parole timing, review conditions, and model-prisoner records; used as Primary focal detail, framed closely enough for the dense parole-review notes to register while remaining bounded by the booklet; open Bible booklet pages (open) — The printed page faces upward around the memo, with both page margins visible; used as Surrounds the memo as contextual scale and identifies where the note was concealed.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient daytime cell illumination remains even and restrained, with neutral-to-cool color and sufficient tonal clarity for the written surface.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The old memorandum remains tucked inside the Bible and is densely covered with parole dates, review conditions, and model-prisoner records.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 낡은 종이 쪽지: \"두려워하지 말라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.444
   },
   "adjusted": {
    "A": 1.667,
    "B": 1.444
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1667,
   "B": 1444
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1667,
    "verdict_ko": "지정된 텍스트('두려워하지 말라')만 선명하게 표시하고 나머지 부분은 블러(얕은 심도) 처리하여, '다른 읽을 수 있는 텍스트를 추가하지 말 것'이라는 프롬프트의 엄격한 지시를 완벽하게 준수했습니다."
   },
   {
    "label": "B",
    "score": 1444,
    "verdict_ko": "지정된 텍스트는 포함되어 있으나, 메모지와 성경책 페이지 전반에 걸쳐 읽을 수 있는 형태의 의미 없는 가짜 한글 텍스트가 다수 생성되어 '추가적인 텍스트 금지' 지시를 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S64sh2_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "메모지 표면 자체는 초점이 선명하게 맞은 상태이나 그 위에 적힌 세부 내용들만 인위적인 디지털 블러(모자이크) 효과로 가려져 있어, 카메라의 얕은 심도 표현이 아닌 후보정 그래픽 오버레이처럼 보입니다.",
     "fix_en": "Replace digital blur with natural illegible handwritten text. Preserve the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "메모가 하단 중앙 1/3이 아니라 화면 중앙을 크게 차지해 프레임을 압도함",
     "fix_en": "Scale down memo to lower third. Preserve the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "쪽지가 성경 페이지 사이에 끼워지지 않고 오른쪽 펼친 면 위에 올려져 있음",
     "fix_en": "Tuck memo into the spine between pages. Preserve the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "성경 페이지 인쇄 글자가 한국어가 아닌 왜곡된 서양 문자처럼 보임",
     "fix_en": "Change Bible text to blurred Korean characters. Preserve the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "책이 놓인 거친 나무 표면이 이전 샷의 독방 탁자·침상과 재질이 다름",
     "fix_en": "Change wood background to smooth gray surface. Preserve the set, the light, and the framing.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "메모지 표면 자체는 초점이 선명하게 맞은 상태이나 그 위에 적힌 세부 내용들만 인위적인 디지털 블러(모자이크) 효과로 가려져 있어, 카메라의 얕은 심도 표현이 아닌 후보정 그래픽 오버레이처럼 보입니다.",
     "severity": "major"
    },
    {
     "issue_ko": "메모가 하단 중앙 1/3이 아니라 화면 중앙을 크게 차지해 프레임을 압도함",
     "severity": "major"
    },
    {
     "issue_ko": "쪽지가 성경 페이지 사이에 끼워지지 않고 오른쪽 펼친 면 위에 올려져 있음",
     "severity": "major"
    },
    {
     "issue_ko": "성경 페이지 인쇄 글자가 한국어가 아닌 왜곡된 서양 문자처럼 보임",
     "severity": "major"
    },
    {
     "issue_ko": "책이 놓인 거친 나무 표면이 이전 샷의 독방 탁자·침상과 재질이 다름",
     "severity": "major"
    },
    {
     "issue_ko": "쪽지 본문의 가석방·심사 기록이 식별되지 않고 전체가 흐릿한 얼룩임",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "메모지 표면 자체는 초점이 선명하게 맞은 상태이나 그 위에 적힌 세부 내용들만 인위적인 디지털 블러(모자이크) 효과로 가려져 있어, 카메라의 얕은 심도 표현이 아닌 후보정 그래픽 오버레이처럼 보입니다.",
    "fix_en": "Replace digital blur with natural illegible handwritten text. Preserve the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "메모가 하단 중앙 1/3이 아니라 화면 중앙을 크게 차지해 프레임을 압도함",
    "fix_en": "Scale down memo to lower third. Preserve the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "쪽지가 성경 페이지 사이에 끼워지지 않고 오른쪽 펼친 면 위에 올려져 있음",
    "fix_en": "Tuck memo into the spine between pages. Preserve the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "성경 페이지 인쇄 글자가 한국어가 아닌 왜곡된 서양 문자처럼 보임",
    "fix_en": "Change Bible text to blurred Korean characters. Preserve the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "책이 놓인 거친 나무 표면이 이전 샷의 독방 탁자·침상과 재질이 다름",
    "fix_en": "Change wood background to smooth gray surface. Preserve the set, the light, and the framing.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S64sh2"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S64sh3::cine": {
  "applied": true,
  "fingerprint": "205d4fd1eec4f6024c258054481af59f17589fb62b6abc7c93d687b46848faf1",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S64sh3_sel.png",
  "source_sha256": "9ff9ff4eff5b03a53ce06c9acbe6e716d8bce56705da58794742a083e0b503e0",
  "file": "S64sh3_cine.png",
  "latency_ms": 11226
 },
 "S65sh2::signage": {
  "fp": "de7c775748b7d133",
  "inscriptions": [
   {
    "surface_native": "시위 피켓",
    "text_native": "진실을 규명하라!",
    "reason_ko": "검찰청 정문 앞에서 피켓 시위를 하는 인물들의 절박한 요구를 화면에 명확히 전달하기 위해 피켓에 적힌 한글 슬로건이 필요합니다."
   },
   {
    "surface_native": "시위 피켓",
    "text_native": "공정수사 촉구!",
    "reason_ko": "두 사람이 각각 피켓을 들고 있는 구도이므로, 시위의 사실감을 더하기 위해 다른 피켓에도 공정 수사를 요구하는 한글 메시지를 배치합니다."
   }
  ]
 },
 "S65sh2": {
  "input_fingerprint": "43a01539a7e69dcc",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): noon.\n\nSHOT TEXT (authoritative, Korean): 정문 앞에서 피켓을 높이 치켜든 채 굳은 표정으로 선 유경자와 신경희의 상체\n\nLOCATION (lock): On the sunlit pavement directly outside the prosecution office's main gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest-to-shoulder height, the nearly completed dolly-in holds 유경자 on the left and 신경희 on the right in a close three-quarter side view, with their raised pickets sharing the upper frame without obscuring either face. Their arms form separate vertical lines while both women remain absorbed in the protest ahead, and fragments of the other demonstrators stay loose and asynchronous behind them.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 지국현의 무혐의를 주장하는 피켓 (Held high) — The message-bearing faces are angled toward the front of the protest and remain legible to the camera at a three-quarter angle; used as Raised above the two women to preserve the protest message within the close framing; 정문 앞 시위대 (About ten church members conducting a picket protest, with varied arm heights, head turns, and weight shifts); used as Provides broken layers of bodies and additional raised signs behind the two primary subjects; 광주지검 정문 (In use as the protest location) — Seen obliquely behind the demonstrators; used as Places the protest at the entrance without competing with the close foreground faces.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Hard midday sunlight is rendered in a restrained neutral-to-cool palette with moderate contrast and natural facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecution-building entrance, concrete facade, gate, and bright daylight. Exclude the arriving vehicle and add the stationary picketers with their signs.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Yu Gyeong-ja and Sin Gyeong-hui continue holding protest placards asserting Ji Gukhyeon's innocence.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 유경자와 신경희 right now — 유경자와 신경희's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 유경자와 신경희 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리); 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 시위 피켓: \"진실을 규명하라!\"\n- 시위 피켓: \"공정수사 촉구!\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): noon.\n\nSHOT TEXT (authoritative, Korean): 정문 앞에서 피켓을 높이 치켜든 채 굳은 표정으로 선 유경자와 신경희의 상체\n\nLOCATION (lock): On the sunlit pavement directly outside the prosecution office's main gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest-to-shoulder height, the nearly completed dolly-in holds 유경자 on the left and 신경희 on the right in a close three-quarter side view, with their raised pickets sharing the upper frame without obscuring either face. Their arms form separate vertical lines while both women remain absorbed in the protest ahead, and fragments of the other demonstrators stay loose and asynchronous behind them.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 지국현의 무혐의를 주장하는 피켓 (Held high) — The message-bearing faces are angled toward the front of the protest and remain legible to the camera at a three-quarter angle; used as Raised above the two women to preserve the protest message within the close framing; 정문 앞 시위대 (About ten church members conducting a picket protest, with varied arm heights, head turns, and weight shifts); used as Provides broken layers of bodies and additional raised signs behind the two primary subjects; 광주지검 정문 (In use as the protest location) — Seen obliquely behind the demonstrators; used as Places the protest at the entrance without competing with the close foreground faces.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Hard midday sunlight is rendered in a restrained neutral-to-cool palette with moderate contrast and natural facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecution-building entrance, concrete facade, gate, and bright daylight. Exclude the arriving vehicle and add the stationary picketers with their signs.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Yu Gyeong-ja and Sin Gyeong-hui continue holding protest placards asserting Ji Gukhyeon's innocence.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 유경자와 신경희 right now — 유경자와 신경희's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 유경자와 신경희 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리); 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 시위 피켓: \"진실을 규명하라!\"\n- 시위 피켓: \"공정수사 촉구!\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): noon.\n\nSHOT TEXT (authoritative, Korean): 정문 앞에서 피켓을 높이 치켜든 채 굳은 표정으로 선 유경자와 신경희의 상체\n\nLOCATION (lock): On the sunlit pavement directly outside the prosecution office's main gate. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest-to-shoulder height, the nearly completed dolly-in holds 유경자 on the left and 신경희 on the right in a close three-quarter side view, with their raised pickets sharing the upper frame without obscuring either face. Their arms form separate vertical lines while both women remain absorbed in the protest ahead, and fragments of the other demonstrators stay loose and asynchronous behind them.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 지국현의 무혐의를 주장하는 피켓 (Held high) — The message-bearing faces are angled toward the front of the protest and remain legible to the camera at a three-quarter angle; used as Raised above the two women to preserve the protest message within the close framing; 정문 앞 시위대 (About ten church members conducting a picket protest, with varied arm heights, head turns, and weight shifts); used as Provides broken layers of bodies and additional raised signs behind the two primary subjects; 광주지검 정문 (In use as the protest location) — Seen obliquely behind the demonstrators; used as Places the protest at the entrance without competing with the close foreground faces.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Hard midday sunlight is rendered in a restrained neutral-to-cool palette with moderate contrast and natural facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same prosecution-building entrance, concrete facade, gate, and bright daylight. Exclude the arriving vehicle and add the stationary picketers with their signs.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Yu Gyeong-ja and Sin Gyeong-hui continue holding protest placards asserting Ji Gukhyeon's innocence.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 유경자와 신경희 right now — 유경자와 신경희's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 유경자와 신경희 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리); 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 시위 피켓: \"진실을 규명하라!\"\n- 시위 피켓: \"공정수사 촉구!\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "두 사람 모두 프레임 오른쪽 앞을 응시하며, 피켓은 시위 방향을 향합니다.",
    "built_space": "배경에 검찰청 건물 외벽과 창문이 보입니다.",
    "entities": "유경자와 신경희의 인물 및 의상은 레퍼런스와 일치하나, 배경 피켓에 의미를 알 수 없는 오탈자들이 존재합니다.",
    "hard_violations": [
     "오른쪽 피켓 막대를 쥐고 있는 손이 없어 피켓이 허공에 떠 있음 (신경희가 팔을 내리고 있음)"
    ],
    "physics": "유경자는 피켓을 정상적으로 쥐고 있으나, 신경희는 팔을 내리고 있어 오른쪽 피켓을 지지하는 물리적 요소가 없습니다."
   },
   {
    "label": "B",
    "direction": "두 사람 모두 결연한 표정으로 오른쪽 앞을 응시하며, 피켓은 카메라와 시위 앞쪽을 향합니다.",
    "built_space": "이전 샷의 검찰청 건물이 배경에 적절한 스케일로 배치되어 있습니다.",
    "entities": "두 인물의 외모와 의상이 완벽히 일치하며, 지정된 두 개의 피켓 문구(\"진실을 규명하라!\", \"공정수사 촉구!\")가 정확히 출력되었습니다.",
    "hard_violations": [],
    "physics": "두 여성 모두 오른팔을 위로 뻗어 피켓의 나무 손잡이를 단단히 쥐고 있어 지지 상태가 자연스럽습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "인물의 외모, 수직으로 뻗은 팔의 자세, 피켓의 정확한 텍스트 및 프레이밍까지 프롬프트의 요구사항을 훌륭하게 충족했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "신경희가 피켓을 들고 있지 않으며 오른쪽 피켓이 공중에 떠 있어 행동 및 지지대 관련 치명적인 오류가 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S34sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:843594>"
   },
   {
    "label": "CHARACTER REFERENCE — 유경자: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1007208>"
   },
   {
    "label": "PROP REFERENCE — 드들강 사건 재수사 촉구 피켓과 플래카드: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:961450>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트에 '3/4 측면 구도(three-quarter side view)'로 명시되었으나, 두 인물이 카메라 렌즈를 향해 정면으로 서서 포즈를 취한 단체 사진처럼 연출되었습니다.",
     "fix_en": "Turn the bodies and faces of the two women so they are seen in a three-quarter profile facing off-screen, maintaining their current clothing and the background.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경 시위대가 들고 있는 피켓들에 '재수수사', '진정수사 출국하라' 등 철자가 틀리거나 문맥에 맞지 않는 텍스트가 생성되었습니다.",
     "fix_en": "Blur the text on the background signs or obscure them with shallow focus so they are completely illegible, keeping the main foreground subjects, their specific signs, and the lighting unchanged.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "유경자(왼쪽 인물)가 피켓 막대를 쥔 손의 새끼손가락 아래쪽 손바닥 부분이 비정상적으로 부풀고 뭉개져 형태가 어색합니다.",
     "fix_en": "Refine the anatomy of the left woman's hand gripping the stick, restoring a natural curve to the palm and pinky finger, while keeping her pose, face, sleeve, and the placard unchanged.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "배경에 광주지검 정문이 사선으로 보이지 않고 건물 외관이 이전 컷과 다르다",
     "fix_en": "Replace the background architecture to show the gate and facade from the reference still at an oblique angle, keeping the foreground subjects, their signs, and lighting exactly as they are.",
     "severity": "major",
     "observation_index": 8,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트에 '3/4 측면 구도(three-quarter side view)'로 명시되었으나, 두 인물이 카메라 렌즈를 향해 정면으로 서서 포즈를 취한 단체 사진처럼 연출되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 시위대가 들고 있는 피켓들에 '재수수사', '진정수사 출국하라' 등 철자가 틀리거나 문맥에 맞지 않는 텍스트가 생성되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "유경자(왼쪽 인물)가 피켓 막대를 쥔 손의 새끼손가락 아래쪽 손바닥 부분이 비정상적으로 부풀고 뭉개져 형태가 어색합니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "유경자와 신경희가 시위 전방이 아니라 상체를 카메라에 거의 정면으로 두고 렌즈를 향해 굳은 포즈로 서 있다",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 가슴 높이 클로즈 3/4 측면이 아니라 두 사람의 정면 상반신 투샷이다",
     "severity": "major"
    },
    {
     "issue_ko": "두 주 피켓 문구면이 3/4 각이 아니라 카메라에 완전 정면을 향한다",
     "severity": "major"
    },
    {
     "issue_ko": "배경 시위대가 느슨한 파편이 아니라 카메라를 향해 줄지어 정렬되어 있다",
     "severity": "major"
    },
    {
     "issue_ko": "배경 피켓에 지정되지 않은 ‘진정수사 출국하라!’ ‘재수사 규명하라!’ 등 문구가 또렷이 읽힌다",
     "severity": "major"
    },
    {
     "issue_ko": "배경에 광주지검 정문이 사선으로 보이지 않고 건물 외관이 이전 컷과 다르다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 6
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트에 '3/4 측면 구도(three-quarter side view)'로 명시되었으나, 두 인물이 카메라 렌즈를 향해 정면으로 서서 포즈를 취한 단체 사진처럼 연출되었습니다.",
    "fix_en": "Turn the bodies and faces of the two women so they are seen in a three-quarter profile facing off-screen, maintaining their current clothing and the background.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경 시위대가 들고 있는 피켓들에 '재수수사', '진정수사 출국하라' 등 철자가 틀리거나 문맥에 맞지 않는 텍스트가 생성되었습니다.",
    "fix_en": "Blur the text on the background signs or obscure them with shallow focus so they are completely illegible, keeping the main foreground subjects, their specific signs, and the lighting unchanged.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "유경자(왼쪽 인물)가 피켓 막대를 쥔 손의 새끼손가락 아래쪽 손바닥 부분이 비정상적으로 부풀고 뭉개져 형태가 어색합니다.",
    "fix_en": "Refine the anatomy of the left woman's hand gripping the stick, restoring a natural curve to the palm and pinky finger, while keeping her pose, face, sleeve, and the placard unchanged.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "배경에 광주지검 정문이 사선으로 보이지 않고 건물 외관이 이전 컷과 다르다",
    "fix_en": "Replace the background architecture to show the gate and facade from the reference still at an oblique angle, keeping the foreground subjects, their signs, and lighting exactly as they are.",
    "severity": "major",
    "observation_index": 8,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S34sh1"
  }
 },
 "S65sh2::cine": {
  "applied": true,
  "fingerprint": "a1dfe74dd54892caceb68c02413eec095fbbc135201db50f47980b74aed336b3",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S65sh2_sel.png",
  "source_sha256": "71fe2ce7eb91cad4f67f6651bcc3a267f9ac2a416c24864eb91fd3565d011a72",
  "file": "S65sh2_cine.png",
  "latency_ms": 13423
 },
 "S66sh3::signage": {
  "fp": "e309971ddd656f41",
  "inscriptions": [
   {
    "surface_native": "단상 뒤 현수막",
    "text_native": "시민과 함께하는 검찰 간담회",
    "reason_ko": "검찰청 세미나실에서 시민들을 대상으로 발표하는 상황이므로 단상 뒤편에 배치될 행사 현수막 문구가 필요합니다."
   }
  ]
 },
 "era_assess::3746474a9efdadd0": {
  "subjects": [
   {
    "subject_native": "대한민국 검찰청 세미나실 (2000년대-2010년대)",
    "search_terms_native": [
     "검찰청 회의실",
     "검찰청 세미나",
     "지검 대회의실",
     "검찰청 강당"
    ],
    "language_lock_native": "이 검색어들은 오직 한국어로만 작성되어야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 검찰청 내부 회의실 및 세미나실은 특유의 검찰 휘장(로고), 목재 벽면 마감, 그리고 특유의 관공서용 가구 배열이 있어 일반적인 서구식 강의실과 명확히 구분됩니다."
   }
  ]
 },
 "era_ref::76bc9d6a6ad2d7f1": {
  "subject": "대한민국 검찰청 세미나실 (2000년대-2010년대)",
  "terms": [
   "검찰청 회의실",
   "검찰청 세미나",
   "지검 대회의실",
   "검찰청 강당"
  ],
  "queries": [
   [
    "검찰청 회의실",
    "검찰청 세미나",
    "지검 대회의실",
    "검찰청 강당"
   ],
   [
    "검찰청 회의실",
    "검찰청 강당"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://www.chosun.com/resizer/v2/XUMVMD74HNMLDB3IMPZGGKCVSQ.jpg?auth=d074263bea2c7e493dd04819be48f3e0af9c9c89194a4b7db18167218d531cc2&width=2560",
  "picked_reason_ko": "1번은 2000~2010년대 대한민국 검찰청의 평범한 세미나·회의실을 명확히 보여 주며, 벽 마감과 창호, 이동식 회의 탁자, 좌석, 마이크, 연단과 국기 등 공간의 구조와 설비를 읽기 좋다.",
  "sha256": "a0f3d8c8d7318501b3f852b94f55057a19627e0bb02ecf5ae38e0c6071464205",
  "file": "eraref_76bc9d6a6ad2d7f1.png"
 },
 "S66sh3::bgfirst_bg": {
  "input_fingerprint": "840d84cf09cc2337",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 단상에 선 채 객석을 향해 확신에 찬 눈빛으로 입을 크게 벌린 장원섭의 상체\n\nLOCATION (lock): At the podium in the prosecution office seminar room, facing the seated citizens and presentation screen.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From near the front row at 장원섭's chest height, the dolly-in settles into a subtly low, off-axis three-quarter close view of his upper body as he speaks with his mouth open and conviction held in his eyes. He occupies the middle-right of the frame and directs his attention beyond the lens toward the seated citizen committee, with only restrained seminar-room context remaining at the edges.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 프레젠테이션 스크린 (Displaying the case presentation) — Only part of the displayed front face is visible behind 장원섭, carrying the prosecution citizens' committee case title; used as A partial background edge ties the close speech shot to the stated case presentation; 시민위원 객석 (Twenty men and women of varied ages seated and listening); used as Soft peripheral silhouettes establish the listeners toward whom 장원섭 is speaking.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime seminar-room illumination is kept naturalistic and neutral-to-cool, with moderate contrast across the speaker's assured expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 검찰청 세미나실 (2000년대-2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 단상에 선 채 객석을 향해 확신에 찬 눈빛으로 입을 크게 벌린 장원섭의 상체\n\nLOCATION (lock): At the podium in the prosecution office seminar room, facing the seated citizens and presentation screen.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From near the front row at 장원섭's chest height, the dolly-in settles into a subtly low, off-axis three-quarter close view of his upper body as he speaks with his mouth open and conviction held in his eyes. He occupies the middle-right of the frame and directs his attention beyond the lens toward the seated citizen committee, with only restrained seminar-room context remaining at the edges.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 프레젠테이션 스크린 (Displaying the case presentation) — Only part of the displayed front face is visible behind 장원섭, carrying the prosecution citizens' committee case title; used as A partial background edge ties the close speech shot to the stated case presentation; 시민위원 객석 (Twenty men and women of varied ages seated and listening); used as Soft peripheral silhouettes establish the listeners toward whom 장원섭 is speaking.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime seminar-room illumination is kept naturalistic and neutral-to-cool, with moderate contrast across the speaker's assured expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 검찰청 세미나실 (2000년대-2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S66sh3__bgfirst_bg.png",
  "asset_id": "db243723-b7ec-455c-ad10-9c25f451a35e",
  "input_asset_ids": [
   "9b37a808-2a88-484d-a0d8-4570bccc77ad",
   "42bd8477-4de9-4549-9e26-e3dc46fec241"
  ],
  "era_research": {
   "subject": "대한민국 검찰청 세미나실 (2000년대-2010년대)",
   "queries": [
    [
     "검찰청 회의실",
     "검찰청 세미나",
     "지검 대회의실",
     "검찰청 강당"
    ],
    [
     "검찰청 회의실",
     "검찰청 강당"
    ]
   ],
   "picked_url": "https://www.chosun.com/resizer/v2/XUMVMD74HNMLDB3IMPZGGKCVSQ.jpg?auth=d074263bea2c7e493dd04819be48f3e0af9c9c89194a4b7db18167218d531cc2&width=2560",
   "sha256": "a0f3d8c8d7318501b3f852b94f55057a19627e0bb02ecf5ae38e0c6071464205",
   "file": "eraref_76bc9d6a6ad2d7f1.png"
  }
 },
 "S66sh3": {
  "input_fingerprint": "87e4e5887375eed2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단상에 선 채 객석을 향해 확신에 찬 눈빛으로 입을 크게 벌린 장원섭의 상체\n\nLOCATION (lock): At the podium in the prosecution office seminar room, facing the seated citizens and presentation screen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near the front row at 장원섭's chest height, the dolly-in settles into a subtly low, off-axis three-quarter close view of his upper body as he speaks with his mouth open and conviction held in his eyes. He occupies the middle-right of the frame and directs his attention beyond the lens toward the seated citizen committee, with only restrained seminar-room context remaining at the edges.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 프레젠테이션 스크린 (Displaying the case presentation) — Only part of the displayed front face is visible behind 장원섭, carrying the prosecution citizens' committee case title; used as A partial background edge ties the close speech shot to the stated case presentation; 시민위원 객석 (Twenty men and women of varied ages seated and listening); used as Soft peripheral silhouettes establish the listeners toward whom 장원섭 is speaking.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime seminar-room illumination is kept naturalistic and neutral-to-cool, with moderate contrast across the speaker's assured expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 단상 뒤 현수막: \"시민과 함께하는 검찰 간담회\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단상에 선 채 객석을 향해 확신에 찬 눈빛으로 입을 크게 벌린 장원섭의 상체\n\nLOCATION (lock): At the podium in the prosecution office seminar room, facing the seated citizens and presentation screen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near the front row at 장원섭's chest height, the dolly-in settles into a subtly low, off-axis three-quarter close view of his upper body as he speaks with his mouth open and conviction held in his eyes. He occupies the middle-right of the frame and directs his attention beyond the lens toward the seated citizen committee, with only restrained seminar-room context remaining at the edges.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 프레젠테이션 스크린 (Displaying the case presentation) — Only part of the displayed front face is visible behind 장원섭, carrying the prosecution citizens' committee case title; used as A partial background edge ties the close speech shot to the stated case presentation; 시민위원 객석 (Twenty men and women of varied ages seated and listening); used as Soft peripheral silhouettes establish the listeners toward whom 장원섭 is speaking.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime seminar-room illumination is kept naturalistic and neutral-to-cool, with moderate contrast across the speaker's assured expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 단상 뒤 현수막: \"시민과 함께하는 검찰 간담회\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단상에 선 채 객석을 향해 확신에 찬 눈빛으로 입을 크게 벌린 장원섭의 상체\n\nLOCATION (lock): At the podium in the prosecution office seminar room, facing the seated citizens and presentation screen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near the front row at 장원섭's chest height, the dolly-in settles into a subtly low, off-axis three-quarter close view of his upper body as he speaks with his mouth open and conviction held in his eyes. He occupies the middle-right of the frame and directs his attention beyond the lens toward the seated citizen committee, with only restrained seminar-room context remaining at the edges.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 프레젠테이션 스크린 (Displaying the case presentation) — Only part of the displayed front face is visible behind 장원섭, carrying the prosecution citizens' committee case title; used as A partial background edge ties the close speech shot to the stated case presentation; 시민위원 객석 (Twenty men and women of varied ages seated and listening); used as Soft peripheral silhouettes establish the listeners toward whom 장원섭 is speaking.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime seminar-room illumination is kept naturalistic and neutral-to-cool, with moderate contrast across the speaker's assured expression.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 단상 뒤 현수막: \"시민과 함께하는 검찰 간담회\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S66sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S66sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L58B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:872599>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 앵글과 프레임 우측 중앙 배치를 충실히 따랐으며, 레퍼런스의 공간 구조와 요구된 텍스트를 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "청중석을 단상 뒤편에 배치하는 치명적인 공간 구조 오류를 범했으며, 피사체를 중앙에 배치하여 프레이밍 지시를 어김."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선과 열린 입이 화면 밖 왼쪽의 청중을 향하고 있음.",
      "built_space": "단상이 앞쪽에 위치하고 그 뒤로 스크린과 현수막이 있으며, 왼쪽 전경에 청중의 뒷모습이 배치되어 기준 사진의 공간 구조와 일치함.",
      "entities": "장원섭의 인상착의가 레퍼런스와 일치하며, 현수막의 '시민과 함께하는 검찰 간담회' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "단상 뒤에 서서 양손을 단상 위에 자연스럽게 올린 채 자세를 유지하고 있음."
     },
     {
      "label": "B",
      "direction": "장원섭의 시선과 입이 화면 밖 왼쪽을 향해 있음.",
      "built_space": "단상과 전면 스크린이 위치한 곳 뒤편 배경에 책상과 청중석이 배치되어 있어, 주어진 세미나실 구조와 완전히 모순됨.",
      "entities": "장원섭의 외모 및 복장이 일치하고 지정된 현수막 텍스트가 명확하게 보임.",
      "hard_violations": [
       "단상 뒤쪽 배경에 청중석을 배치하여 기준 사진의 공간 구조를 심각하게 왜곡함(물리적으로 불가능한 스테이징)."
      ],
      "physics": "단상 뒤에 서서 손을 올린 채 바닥에 지지되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 앵글과 프레임 우측 중앙 배치를 충실히 따랐으며, 레퍼런스의 공간 구조와 요구된 텍스트를 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "청중석을 단상 뒤편에 배치하는 치명적인 공간 구조 오류를 범했으며, 피사체를 중앙에 배치하여 프레이밍 지시를 어김."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선과 열린 입이 화면 밖 왼쪽의 청중을 향하고 있음.",
      "built_space": "단상이 앞쪽에 위치하고 그 뒤로 스크린과 현수막이 있으며, 왼쪽 전경에 청중의 뒷모습이 배치되어 기준 사진의 공간 구조와 일치함.",
      "entities": "장원섭의 인상착의가 레퍼런스와 일치하며, 현수막의 '시민과 함께하는 검찰 간담회' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "단상 뒤에 서서 양손을 단상 위에 자연스럽게 올린 채 자세를 유지하고 있음."
     },
     {
      "label": "B",
      "direction": "장원섭의 시선과 입이 화면 밖 왼쪽을 향해 있음.",
      "built_space": "단상과 전면 스크린이 위치한 곳 뒤편 배경에 책상과 청중석이 배치되어 있어, 주어진 세미나실 구조와 완전히 모순됨.",
      "entities": "장원섭의 외모 및 복장이 일치하고 지정된 현수막 텍스트가 명확하게 보임.",
      "hard_violations": [
       "단상 뒤쪽 배경에 청중석을 배치하여 기준 사진의 공간 구조를 심각하게 왜곡함(물리적으로 불가능한 스테이징)."
      ],
      "physics": "단상 뒤에 서서 손을 올린 채 바닥에 지지되어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "전경에 청중의 뒷모습을 실루엣으로 배치하여 지정된 3/4 클로즈업 앵글과 세미나실의 공간 구조를 완벽하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "연설자 뒤편의 배경에 청중이 마주 보고 앉아 있는 물리적으로 불가능한 공간 구조를 렌더링하여 심각한 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭은 화면 좌측을 향해 시선을 던지며, 배경에 앉은 청중들은 우측 전경의 장원섭을 향해 시선을 고정하고 있습니다.",
      "built_space": "공간 구조가 심각하게 왜곡되었습니다. 장원섭의 뒤편으로 프레젠테이션 스크린과 현수막이 있는 벽면이 이어져야 하나, 스크린은 장원섭 바로 뒤에, 현수막은 배경의 청중 뒤편에 위치하며 청중이 장원섭을 마주 보고 앉아 있는 모순된 구조입니다.",
      "entities": "장원섭의 외모, 복장(회색 정장, 파란 셔츠, 도트 넥타이)은 레퍼런스와 일치하며, 현수막과 스크린의 텍스트도 지시사항에 맞게 출력되었습니다.",
      "hard_violations": [
       " physically impossible staging (강연자 뒤편에 청중이 마주 보고 앉아 있는 불가능한 공간 및 좌석 배치)"
      ],
      "physics": "인물들과 사물들은 정상적으로 지면과 책상에 지지되어 있습니다."
     },
     {
      "label": "B",
      "direction": "장원섭은 화면 좌측 앞쪽을 향해 확신에 찬 시선을 던지고 있으며, 전경의 청중들은 장원섭을 향해 시선을 두고 있습니다.",
      "built_space": "카메라가 앞줄 근처에서 연단 위의 장원섭을 향하고, 프레젠테이션 스크린과 현수막이 그 뒤쪽 벽면에 올바르게 위치해 있으며, 청중은 전경에 자연스럽게 배치되었습니다.",
      "entities": "장원섭의 인상착의가 레퍼런스와 완벽히 일치하며, 지시된 현수막 텍스트와 스크린 내용 모두 정확하게 나타납니다.",
      "hard_violations": [],
      "physics": "장원섭의 손은 단상 위에 자연스럽게 놓여 지지받고 있으며, 전체적인 물리적 상태가 안정적입니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "전경에 청중의 뒷모습을 실루엣으로 배치하여 지정된 3/4 클로즈업 앵글과 세미나실의 공간 구조를 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "연설자 뒤편의 배경에 청중이 마주 보고 앉아 있는 물리적으로 불가능한 공간 구조를 렌더링하여 심각한 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "장원섭은 화면 좌측을 향해 시선을 던지며, 배경에 앉은 청중들은 우측 전경의 장원섭을 향해 시선을 고정하고 있습니다.",
      "built_space": "공간 구조가 심각하게 왜곡되었습니다. 장원섭의 뒤편으로 프레젠테이션 스크린과 현수막이 있는 벽면이 이어져야 하나, 스크린은 장원섭 바로 뒤에, 현수막은 배경의 청중 뒤편에 위치하며 청중이 장원섭을 마주 보고 앉아 있는 모순된 구조입니다.",
      "entities": "장원섭의 외모, 복장(회색 정장, 파란 셔츠, 도트 넥타이)은 레퍼런스와 일치하며, 현수막과 스크린의 텍스트도 지시사항에 맞게 출력되었습니다.",
      "hard_violations": [
       " physically impossible staging (강연자 뒤편에 청중이 마주 보고 앉아 있는 불가능한 공간 및 좌석 배치)"
      ],
      "physics": "인물들과 사물들은 정상적으로 지면과 책상에 지지되어 있습니다."
     },
     {
      "label": "A",
      "direction": "장원섭은 화면 좌측 앞쪽을 향해 확신에 찬 시선을 던지고 있으며, 전경의 청중들은 장원섭을 향해 시선을 두고 있습니다.",
      "built_space": "카메라가 앞줄 근처에서 연단 위의 장원섭을 향하고, 프레젠테이션 스크린과 현수막이 그 뒤쪽 벽면에 올바르게 위치해 있으며, 청중은 전경에 자연스럽게 배치되었습니다.",
      "entities": "장원섭의 인상착의가 레퍼런스와 완벽히 일치하며, 지시된 현수막 텍스트와 스크린 내용 모두 정확하게 나타납니다.",
      "hard_violations": [],
      "physics": "장원섭의 손은 단상 위에 자연스럽게 놓여 지지받고 있으며, 전체적인 물리적 상태가 안정적입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 15,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "장원섭의 시선과 열린 입이 화면 밖 왼쪽의 청중을 향하고 있음.",
    "built_space": "단상이 앞쪽에 위치하고 그 뒤로 스크린과 현수막이 있으며, 왼쪽 전경에 청중의 뒷모습이 배치되어 기준 사진의 공간 구조와 일치함.",
    "entities": "장원섭의 인상착의가 레퍼런스와 일치하며, 현수막의 '시민과 함께하는 검찰 간담회' 텍스트가 정확히 표기됨.",
    "hard_violations": [],
    "physics": "단상 뒤에 서서 양손을 단상 위에 자연스럽게 올린 채 자세를 유지하고 있음."
   },
   {
    "label": "B",
    "direction": "장원섭의 시선과 입이 화면 밖 왼쪽을 향해 있음.",
    "built_space": "단상과 전면 스크린이 위치한 곳 뒤편 배경에 책상과 청중석이 배치되어 있어, 주어진 세미나실 구조와 완전히 모순됨.",
    "entities": "장원섭의 외모 및 복장이 일치하고 지정된 현수막 텍스트가 명확하게 보임.",
    "hard_violations": [
     "단상 뒤쪽 배경에 청중석을 배치하여 기준 사진의 공간 구조를 심각하게 왜곡함(물리적으로 불가능한 스테이징)."
    ],
    "physics": "단상 뒤에 서서 손을 올린 채 바닥에 지지되어 있음."
   }
  ],
  "totals": {
   "A": 15,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지정된 앵글과 프레임 우측 중앙 배치를 충실히 따랐으며, 레퍼런스의 공간 구조와 요구된 텍스트를 정확하게 구현함."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "청중석을 단상 뒤편에 배치하는 치명적인 공간 구조 오류를 범했으며, 피사체를 중앙에 배치하여 프레이밍 지시를 어김."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L58B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "스크린 위의 텍스트가 표면 원근감을 무시한 채 수평으로 합성된 평면 그래픽 오버레이처럼 보입니다.",
     "fix_en": "Map the text directly onto the projector screen surface so it perfectly follows the screen's perspective angle and lighting, removing the flat horizontal graphic overlay. Maintain Jang Won-seop's exact face, expression, pose, grey suit, the podium, the audience silhouettes, and the room's lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "우측 상단의 현수막이 조명이나 공간감이 전혀 없이 화면에 평평하게 붙인 디지털 스티커처럼 렌더링되었습니다.",
     "fix_en": "Render the top-right banner as a physical material attached to the wall with appropriate room lighting, texture, and subtle shadows, removing the flat digital sticker look. Maintain Jang Won-seop's exact face, expression, pose, grey suit, the podium, the projector screen, and the room's lighting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "샷 텍스트에 없는 객석 인물들이 왼쪽 하단에 여러 명 선명하게 들어가 있다",
     "fix_en": "Remove the distinctly detailed people in the bottom-left foreground and replace them with out-of-focus, very soft and indistinct dark peripheral silhouettes. Maintain Jang Won-seop's exact face, expression, pose, grey suit, the podium, the projector screen, and the background signage.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "잠긴 배경 사진의 카메라·원근·벽과 스크린 배치가 전혀 유지되지 않았다",
     "fix_en": "Adjust the background elements slightly to hint at the original locked reference geometry. Maintain the speaker, his pose, the foreground, and current lighting.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "상체 클로즈업이 아니라 단상·객석·스크린이 넓게 잡힌 미디엄 샷이다",
     "fix_en": "Crop the framing tighter to an upper-body close-up. Maintain the speaker's face, suit, and the podium.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "스크린 위의 텍스트가 표면 원근감을 무시한 채 수평으로 합성된 평면 그래픽 오버레이처럼 보입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "우측 상단의 현수막이 조명이나 공간감이 전혀 없이 화면에 평평하게 붙인 디지털 스티커처럼 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "다양한 연령대의 남녀가 있어야 할 객석에 비슷한 짧은 머리와 정장 차림의 남성들만 묘사되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "샷 텍스트에 없는 객석 인물들이 왼쪽 하단에 여러 명 선명하게 들어가 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "잠긴 배경 사진의 카메라·원근·벽과 스크린 배치가 전혀 유지되지 않았다",
     "severity": "major"
    },
    {
     "issue_ko": "상체 클로즈업이 아니라 단상·객석·스크린이 넓게 잡힌 미디엄 샷이다",
     "severity": "major"
    },
    {
     "issue_ko": "스크린이 인물 뒤 일부 가장자리가 아니라 화면 왼쪽을 거의 가득 채운다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "잠긴 배경 사진의 카메라·원근·벽과 스크린 배치가 전혀 유지되지 않았다",
    "fix_en": "Adjust the background elements slightly to hint at the original locked reference geometry. Maintain the speaker, his pose, the foreground, and current lighting.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "상체 클로즈업이 아니라 단상·객석·스크린이 넓게 잡힌 미디엄 샷이다",
    "fix_en": "Crop the framing tighter to an upper-body close-up. Maintain the speaker's face, suit, and the podium.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Map the text directly onto the projector screen surface so it perfectly follows the screen's perspective angle and lighting, removing the flat horizontal graphic overlay. Maintain Jang Won-seop's exact face, expression, pose, grey suit, the podium, the audience silhouettes, and the room's lighting.\n- Render the top-right banner as a physical material attached to the wall with appropriate room lighting, texture, and subtle shadows, removing the flat digital sticker look. Maintain Jang Won-seop's exact face, expression, pose, grey suit, the podium, the projector screen, and the room's lighting.\n- Remove the distinctly detailed people in the bottom-left foreground and replace them with out-of-focus, very soft and indistinct dark peripheral silhouettes. Maintain Jang Won-seop's exact face, expression, pose, grey suit, the podium, the projector screen, and the background signage.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "요청된 클로즈업 프레이밍과 크게 벌린 입의 액션을 완벽히 구현했으며, 현수막의 텍스트까지 정확하게 렌더링하여 지시사항을 매우 훌륭하게 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레이밍이 지정된 클로즈업이 아닌 와이드 샷이며, 액션(크게 벌린 입)이 부족하고 현수막에 '배너'라는 잘못된 텍스트를 출력하여 우선순위 조건들을 다수 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선과 크게 열린 입이 화면 왼쪽 하단의 객석 청중을 명확히 향하고 있음.",
      "built_space": "단상 뒤에 화자가 위치하며, 왼쪽 배경에 프레젠테이션 스크린 일부가, 오른쪽 벽면에 현수막이 배치됨. 전경 왼쪽에 청중이 앉아 있음.",
      "entities": "장원섭의 얼굴, 헤어스타일, 수트 착장이 레퍼런스와 정확히 일치함. 스크린과 현수막의 텍스트가 프롬프트대로 완벽히 생성됨.",
      "hard_violations": [],
      "physics": "화자는 단상 뒤에 체중을 싣고 안정적으로 서 있으며, 전경의 청중들은 의자에 정상적으로 착석해 있음."
     },
     {
      "label": "B",
      "direction": "장원섭의 시선은 왼쪽 하단을 향하고 있으나, 입은 다물어져 있거나 아주 살짝만 열려 있음.",
      "built_space": "세미나실의 넓은 전경이 보이며, 중앙의 단상 뒤에 화자가 섰고 오른쪽 문 위로 종이가 붙어 있음. 전경에 청중이 위치함.",
      "entities": "장원섭의 외모와 복장은 일치하나, 요청된 현수막 텍스트 대신 '배너'라는 엉뚱한 글자가 생성됨.",
      "hard_violations": [],
      "physics": "화자는 단상 뒤 바닥에 서서 한 손으로 가벼운 제스처를 취하고 있으며, 물리적으로 어색한 부분은 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "요청된 클로즈업 프레이밍과 크게 벌린 입의 액션을 완벽히 구현했으며, 현수막의 텍스트까지 정확하게 렌더링하여 지시사항을 매우 훌륭하게 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레이밍이 지정된 클로즈업이 아닌 와이드 샷이며, 액션(크게 벌린 입)이 부족하고 현수막에 '배너'라는 잘못된 텍스트를 출력하여 우선순위 조건들을 다수 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선과 크게 열린 입이 화면 왼쪽 하단의 객석 청중을 명확히 향하고 있음.",
      "built_space": "단상 뒤에 화자가 위치하며, 왼쪽 배경에 프레젠테이션 스크린 일부가, 오른쪽 벽면에 현수막이 배치됨. 전경 왼쪽에 청중이 앉아 있음.",
      "entities": "장원섭의 얼굴, 헤어스타일, 수트 착장이 레퍼런스와 정확히 일치함. 스크린과 현수막의 텍스트가 프롬프트대로 완벽히 생성됨.",
      "hard_violations": [],
      "physics": "화자는 단상 뒤에 체중을 싣고 안정적으로 서 있으며, 전경의 청중들은 의자에 정상적으로 착석해 있음."
     },
     {
      "label": "B",
      "direction": "장원섭의 시선은 왼쪽 하단을 향하고 있으나, 입은 다물어져 있거나 아주 살짝만 열려 있음.",
      "built_space": "세미나실의 넓은 전경이 보이며, 중앙의 단상 뒤에 화자가 섰고 오른쪽 문 위로 종이가 붙어 있음. 전경에 청중이 위치함.",
      "entities": "장원섭의 외모와 복장은 일치하나, 요청된 현수막 텍스트 대신 '배너'라는 엉뚱한 글자가 생성됨.",
      "hard_violations": [],
      "physics": "화자는 단상 뒤 바닥에 서서 한 손으로 가벼운 제스처를 취하고 있으며, 물리적으로 어색한 부분은 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "지시된 클로즈업 프레이밍, 입을 크게 벌린 역동적인 표정, 정확한 현수막 텍스트를 모두 훌륭하게 구현한 최적의 결과물입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 너무 넓게 잡았으며, 입을 벌린 표정이 누락되었고 현수막 텍스트가 엉뚱하게 출력되어 활용할 수 없습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "남자의 시선과 얼굴 방향이 화면 왼쪽 아래에 앉아있는 청중들을 향하고 있음.",
      "built_space": "카메라가 피사체에 가깝게 다가가 배경의 스크린과 단상 일부, 지정된 현수막이 앵글 안에 배치됨.",
      "entities": "장원섭의 얼굴, 헤어, 정장이 레퍼런스와 일치하며 입을 크게 벌린 표정이 정확히 묘사됨. 현수막에 '시민과 함께하는 검찰 간담회' 텍스트가 정확히 출력됨.",
      "hard_violations": [],
      "physics": "남자는 단상 뒤에 안정적으로 서서 손을 단상 위에 얹고 있음."
     },
     {
      "label": "A",
      "direction": "남자의 시선이 화면 왼쪽 아래의 청중을 향하고 있음.",
      "built_space": "요구된 클로즈업이 아닌 넓은 화각으로 배경의 전체적인 윤곽(스크린, 오른쪽 문 등)이 모두 노출됨.",
      "entities": "장원섭의 외모는 일치하나 입을 크게 벌리지 않음. 뒤쪽 현수막에 요구된 텍스트 대신 '배 너'라는 잘못된 글자가 적혀 있음.",
      "hard_violations": [
       "프레이밍 위반 (클로즈업이 아닌 미디엄 샷 이상으로 렌더링됨)",
       "텍스트 위반 (요구된 문구 대신 '배 너' 출력)"
      ],
      "physics": "단상 뒤에 서서 오른손을 가볍게 들고 있는 안정적인 자세."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지시된 클로즈업 프레이밍, 입을 크게 벌린 역동적인 표정, 정확한 현수막 텍스트를 모두 훌륭하게 구현한 최적의 결과물입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 지시를 무시하고 너무 넓게 잡았으며, 입을 벌린 표정이 누락되었고 현수막 텍스트가 엉뚱하게 출력되어 활용할 수 없습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "남자의 시선과 얼굴 방향이 화면 왼쪽 아래에 앉아있는 청중들을 향하고 있음.",
      "built_space": "카메라가 피사체에 가깝게 다가가 배경의 스크린과 단상 일부, 지정된 현수막이 앵글 안에 배치됨.",
      "entities": "장원섭의 얼굴, 헤어, 정장이 레퍼런스와 일치하며 입을 크게 벌린 표정이 정확히 묘사됨. 현수막에 '시민과 함께하는 검찰 간담회' 텍스트가 정확히 출력됨.",
      "hard_violations": [],
      "physics": "남자는 단상 뒤에 안정적으로 서서 손을 단상 위에 얹고 있음."
     },
     {
      "label": "B",
      "direction": "남자의 시선이 화면 왼쪽 아래의 청중을 향하고 있음.",
      "built_space": "요구된 클로즈업이 아닌 넓은 화각으로 배경의 전체적인 윤곽(스크린, 오른쪽 문 등)이 모두 노출됨.",
      "entities": "장원섭의 외모는 일치하나 입을 크게 벌리지 않음. 뒤쪽 현수막에 요구된 텍스트 대신 '배 너'라는 잘못된 글자가 적혀 있음.",
      "hard_violations": [
       "프레이밍 위반 (클로즈업이 아닌 미디엄 샷 이상으로 렌더링됨)",
       "텍스트 위반 (요구된 문구 대신 '배 너' 출력)"
      ],
      "physics": "단상 뒤에 서서 오른손을 가볍게 들고 있는 안정적인 자세."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 18,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S66sh3__bgfirst_bg.png",
   "bg_asset_id": "db243723-b7ec-455c-ad10-9c25f451a35e",
   "bg_record_key": "S66sh3::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S66sh3::cine": {
  "applied": true,
  "fingerprint": "f0fe5b5187475169eee76c25c71714143f194a2e4e21fa48b406e7844f58f095",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S66sh3_sel.png",
  "source_sha256": "7c13efee6857c1c018b59e7416e1b8f7f5681a45039e1a45172c5c6d61f09bb5",
  "file": "S66sh3_cine.png",
  "latency_ms": 14339
 },
 "S67sh1::signage": {
  "fp": "f9e5f488ba207529",
  "inscriptions": [
   {
    "surface_native": "세미나실 문 위의 아크릴 표지판",
    "text_native": "세미나실",
    "reason_ko": "검찰청 복도에서 해당 공간이 세미나실임을 명확히 인지할 수 있도록 문에 부착된 한국어 표지판이 필요합니다."
   }
  ]
 },
 "era_assess::52ce2982b67bed4c": {
  "subjects": [
   {
    "subject_native": "2000년대 한국 대학 세미나실 문과 복도",
    "search_terms_native": [
     "대학 강의실 복도 문",
     "세미나실 문 유리창",
     "대학교 강의실 문 디자인"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어로 번역하거나 다른 언어를 혼용해서는 안 됩니다.",
    "reason_ko": "2000년대 한국 대학 강의실이나 세미나실의 문(특유의 세로형 유리창이 있는 목제/철제 문)과 복도 마감재, 바닥재 등은 서구식 대학 복도 디자인과 뚜렷이 구분되므로 현지 고증이 필요합니다."
   }
  ]
 },
 "era_ref::ff3f44c4228edc71": {
  "subject": "2000년대 한국 대학 세미나실 문과 복도",
  "terms": [
   "대학 강의실 복도 문",
   "세미나실 문 유리창",
   "대학교 강의실 문 디자인"
  ],
  "queries": [
   [
    "2000년대 한국 대학교 강의실 복도 문 세미나실 문 유리창 디자인",
    "한국 대학 인문대학 강의실 복도 나무문 세로 유리창"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://media.khu.ac.kr/data/editor/2503/thumb-604e7d1cf00ca8dece6f69fe2451bf16_1743383976_9615_853x640.jpg",
  "picked_reason_ko": "3번은 한국 대학에서 흔히 볼 수 있는 세미나실 출입문과 소박한 복도를 함께 명확히 보여 주며, 문틀·유리문·바닥·천장 조명 등 실제 시공 요소도 잘 읽힌다.",
  "sha256": "c2a6ba01f3dd1b3cf1c356c7e3a5eef12538dfc8421caedd4e007bf247e9df22",
  "file": "eraref_ff3f44c4228edc71.png"
 },
 "S67sh1::bgfirst_bg": {
  "input_fingerprint": "61bcff90c5bc43be",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 굳게 닫힌 세미나실 문 중앙의 작은 유리창 쪽으로 얼굴을 바짝 들이민 차장검사의 상체\n\nLOCATION (lock): In the corridor directly outside the seminar room, at the small window in its closed door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside and slightly behind 차장검사, the continuing dolly-in frames his upper body diagonally against the closed door as he brings his face close to the small window. His near shoulder occupies the lower-left edge, the window sits ahead of his eyeline near the upper center, and his attention remains fixed on the seminar room beyond it.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, foreground, looks toward small door window; small door window in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 세미나실 문의 작은 유리창 (Set into the closed door) — The window is seen obliquely, with the seminar-room interior lying beyond it; used as The camera looks across 차장검사 toward this constrained viewing aperture; 세미나실 문 (Closed) — Its corridor-facing side is seen at an oblique angle; used as Creates a firm planar barrier around the small window and compresses the observer against it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime corridor ambience is treated with subdued neutral-to-cool tones and naturalistic contrast, without adding a specified practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 대학 세미나실 문과 복도: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 굳게 닫힌 세미나실 문 중앙의 작은 유리창 쪽으로 얼굴을 바짝 들이민 차장검사의 상체\n\nLOCATION (lock): In the corridor directly outside the seminar room, at the small window in its closed door.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At shoulder height beside and slightly behind 차장검사, the continuing dolly-in frames his upper body diagonally against the closed door as he brings his face close to the small window. His near shoulder occupies the lower-left edge, the window sits ahead of his eyeline near the upper center, and his attention remains fixed on the seminar room beyond it.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, foreground, looks toward small door window; small door window in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 세미나실 문의 작은 유리창 (Set into the closed door) — The window is seen obliquely, with the seminar-room interior lying beyond it; used as The camera looks across 차장검사 toward this constrained viewing aperture; 세미나실 문 (Closed) — Its corridor-facing side is seen at an oblique angle; used as Creates a firm planar barrier around the small window and compresses the observer against it.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime corridor ambience is treated with subdued neutral-to-cool tones and naturalistic contrast, without adding a specified practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국 대학 세미나실 문과 복도: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S67sh1__bgfirst_bg.png",
  "asset_id": "44d9f8df-7e42-4a49-806d-5b5e4aa30ae8",
  "input_asset_ids": [
   "4f5a4f8a-eb97-48b6-93e1-be5731a94ae7",
   "2c7cc51a-b502-4fd9-a742-048b4ff57f07"
  ],
  "era_research": {
   "subject": "2000년대 한국 대학 세미나실 문과 복도",
   "queries": [
    [
     "2000년대 한국 대학교 강의실 복도 문 세미나실 문 유리창 디자인",
     "한국 대학 인문대학 강의실 복도 나무문 세로 유리창"
    ]
   ],
   "picked_url": "https://media.khu.ac.kr/data/editor/2503/thumb-604e7d1cf00ca8dece6f69fe2451bf16_1743383976_9615_853x640.jpg",
   "sha256": "c2a6ba01f3dd1b3cf1c356c7e3a5eef12538dfc8421caedd4e007bf247e9df22",
   "file": "eraref_ff3f44c4228edc71.png"
  }
 },
 "S67sh1": {
  "input_fingerprint": "49fe9402c092f9bd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 굳게 닫힌 세미나실 문 중앙의 작은 유리창 쪽으로 얼굴을 바짝 들이민 차장검사의 상체\n\nLOCATION (lock): In the corridor directly outside the seminar room, at the small window in its closed door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height beside and slightly behind 차장검사, the continuing dolly-in frames his upper body diagonally against the closed door as he brings his face close to the small window. His near shoulder occupies the lower-left edge, the window sits ahead of his eyeline near the upper center, and his attention remains fixed on the seminar room beyond it.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, foreground, looks toward small door window; small door window in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 세미나실 문의 작은 유리창 (Set into the closed door) — The window is seen obliquely, with the seminar-room interior lying beyond it; used as The camera looks across 차장검사 toward this constrained viewing aperture; 세미나실 문 (Closed) — Its corridor-facing side is seen at an oblique angle; used as Creates a firm planar barrier around the small window and compresses the observer against it.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime corridor ambience is treated with subdued neutral-to-cool tones and naturalistic contrast, without adding a specified practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 세미나실 문 위의 아크릴 표지판: \"세미나실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 굳게 닫힌 세미나실 문 중앙의 작은 유리창 쪽으로 얼굴을 바짝 들이민 차장검사의 상체\n\nLOCATION (lock): In the corridor directly outside the seminar room, at the small window in its closed door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height beside and slightly behind 차장검사, the continuing dolly-in frames his upper body diagonally against the closed door as he brings his face close to the small window. His near shoulder occupies the lower-left edge, the window sits ahead of his eyeline near the upper center, and his attention remains fixed on the seminar room beyond it.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, foreground, looks toward small door window; small door window in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 세미나실 문의 작은 유리창 (Set into the closed door) — The window is seen obliquely, with the seminar-room interior lying beyond it; used as The camera looks across 차장검사 toward this constrained viewing aperture; 세미나실 문 (Closed) — Its corridor-facing side is seen at an oblique angle; used as Creates a firm planar barrier around the small window and compresses the observer against it.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime corridor ambience is treated with subdued neutral-to-cool tones and naturalistic contrast, without adding a specified practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 세미나실 문 위의 아크릴 표지판: \"세미나실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 굳게 닫힌 세미나실 문 중앙의 작은 유리창 쪽으로 얼굴을 바짝 들이민 차장검사의 상체\n\nLOCATION (lock): In the corridor directly outside the seminar room, at the small window in its closed door. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At shoulder height beside and slightly behind 차장검사, the continuing dolly-in frames his upper body diagonally against the closed door as he brings his face close to the small window. His near shoulder occupies the lower-left edge, the window sits ahead of his eyeline near the upper center, and his attention remains fixed on the seminar room beyond it.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 차장검사 in the middle-left of the frame, foreground, looks toward small door window; small door window in the upper-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 세미나실 문의 작은 유리창 (Set into the closed door) — The window is seen obliquely, with the seminar-room interior lying beyond it; used as The camera looks across 차장검사 toward this constrained viewing aperture; 세미나실 문 (Closed) — Its corridor-facing side is seen at an oblique angle; used as Creates a firm planar barrier around the small window and compresses the observer against it.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime corridor ambience is treated with subdued neutral-to-cool tones and naturalistic contrast, without adding a specified practical source.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 세미나실 문 위의 아크릴 표지판: \"세미나실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S67sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S67sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L59B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "표지판이 문 위가 아닌 옆 벽면에 위치한 점은 아쉬우나, 참조된 인물의 외모와 복장을 완벽하게 구현했으며 해부학적 오류 없이 프레이밍 및 시선 지시를 정확히 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "표지판의 위치와 전반적인 구도는 좋으나, 화면 우측 하단의 오른팔에 왼손이 결합되어 있는 치명적인 해부학적 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "검사의 시선 및 상체가 닫힌 문의 중앙에 있는 작은 유리창을 정확히 향하고 있음.",
      "built_space": "복도에 닫힌 문과 유리창이 배치되어 있으며, 문의 우측 벽면에 '세미나실' 텍스트가 적힌 아크릴 표지판이 부착됨.",
      "entities": "지시된 차장검사의 얼굴, 헤어스타일, 줄무늬 정장 및 넥타이가 참조 이미지와 정확히 일치함. 표지판의 텍스트도 올바름.",
      "hard_violations": [],
      "physics": "화면 밖 바닥에 발을 딛고 상체를 기울여 유리창 쪽으로 다가가는 자연스러운 자세로 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "검사의 시선이 문 중앙의 작은 유리창을 향하고 있음.",
      "built_space": "복도와 유리창이 달린 문이 있으며, 문 상단에 '세미나실' 표지판이 부착됨. 창문 너머로 내부가 보임.",
      "entities": "차장검사의 외모와 복장이 참조와 일치하며, 표지판의 문자도 올바르게 생성됨.",
      "hard_violations": [
       "화면 우측 하단에서 문을 짚고 있는 오른팔에 엄지가 위로 향하는 왼손이 달려 있는 물리적으로 불가능한 해부학적 오류"
      ],
      "physics": "양손으로 문과 문틀을 짚고 체중을 지탱하고 있으나, 우측 손의 구조가 해부학적으로 성립할 수 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "표지판이 문 위가 아닌 옆 벽면에 위치한 점은 아쉬우나, 참조된 인물의 외모와 복장을 완벽하게 구현했으며 해부학적 오류 없이 프레이밍 및 시선 지시를 정확히 따랐습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "표지판의 위치와 전반적인 구도는 좋으나, 화면 우측 하단의 오른팔에 왼손이 결합되어 있는 치명적인 해부학적 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "검사의 시선 및 상체가 닫힌 문의 중앙에 있는 작은 유리창을 정확히 향하고 있음.",
      "built_space": "복도에 닫힌 문과 유리창이 배치되어 있으며, 문의 우측 벽면에 '세미나실' 텍스트가 적힌 아크릴 표지판이 부착됨.",
      "entities": "지시된 차장검사의 얼굴, 헤어스타일, 줄무늬 정장 및 넥타이가 참조 이미지와 정확히 일치함. 표지판의 텍스트도 올바름.",
      "hard_violations": [],
      "physics": "화면 밖 바닥에 발을 딛고 상체를 기울여 유리창 쪽으로 다가가는 자연스러운 자세로 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "검사의 시선이 문 중앙의 작은 유리창을 향하고 있음.",
      "built_space": "복도와 유리창이 달린 문이 있으며, 문 상단에 '세미나실' 표지판이 부착됨. 창문 너머로 내부가 보임.",
      "entities": "차장검사의 외모와 복장이 참조와 일치하며, 표지판의 문자도 올바르게 생성됨.",
      "hard_violations": [
       "화면 우측 하단에서 문을 짚고 있는 오른팔에 엄지가 위로 향하는 왼손이 달려 있는 물리적으로 불가능한 해부학적 오류"
      ],
      "physics": "양손으로 문과 문틀을 짚고 체중을 지탱하고 있으나, 우측 손의 구조가 해부학적으로 성립할 수 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1429,
      "verdict_ko": "지시된 구도, 인물, 텍스트를 모두 정확하게 구현했으나, 표지판이 문 위가 아닌 벽면에 위치함."
     },
     {
      "label": "A",
      "score": 929,
      "verdict_ko": "구도와 텍스트는 양호하나, 가슴 부근에 연결되지 않은 손이 나타나는 치명적인 신체 구조 오류가 있음.  ★위반: [gemini-pro] 가슴 부근의 문틀을 잡고 있는 왼손이 어깨와 정상적으로 연결되지 않은 해부학적 불가능 구조 / [gemini-pro] 우측 하단에 위치한 손의 형태 왜곡"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.429,
      "B": 1.429
     },
     "adjusted": {
      "A": 0.929,
      "B": 1.429
     },
     "violations": {
      "A": [
       "[gemini-pro] 가슴 부근의 문틀을 잡고 있는 왼손이 어깨와 정상적으로 연결되지 않은 해부학적 불가능 구조",
       "[gemini-pro] 우측 하단에 위치한 손의 형태 왜곡"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.571,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1429,
      "verdict_ko": "지시된 구도, 인물, 텍스트를 모두 정확하게 구현했으나, 표지판이 문 위가 아닌 벽면에 위치함."
     },
     {
      "label": "B",
      "score": 929,
      "verdict_ko": "구도와 텍스트는 양호하나, 가슴 부근에 연결되지 않은 손이 나타나는 치명적인 신체 구조 오류가 있음.  ★위반: [gemini-pro] 가슴 부근의 문틀을 잡고 있는 왼손이 어깨와 정상적으로 연결되지 않은 해부학적 불가능 구조 / [gemini-pro] 우측 하단에 위치한 손의 형태 왜곡"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1436,
     "B": 932
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "검사의 시선 및 상체가 닫힌 문의 중앙에 있는 작은 유리창을 정확히 향하고 있음.",
    "built_space": "복도에 닫힌 문과 유리창이 배치되어 있으며, 문의 우측 벽면에 '세미나실' 텍스트가 적힌 아크릴 표지판이 부착됨.",
    "entities": "지시된 차장검사의 얼굴, 헤어스타일, 줄무늬 정장 및 넥타이가 참조 이미지와 정확히 일치함. 표지판의 텍스트도 올바름.",
    "hard_violations": [],
    "physics": "화면 밖 바닥에 발을 딛고 상체를 기울여 유리창 쪽으로 다가가는 자연스러운 자세로 지지되어 있음."
   },
   {
    "label": "B",
    "direction": "검사의 시선이 문 중앙의 작은 유리창을 향하고 있음.",
    "built_space": "복도와 유리창이 달린 문이 있으며, 문 상단에 '세미나실' 표지판이 부착됨. 창문 너머로 내부가 보임.",
    "entities": "차장검사의 외모와 복장이 참조와 일치하며, 표지판의 문자도 올바르게 생성됨.",
    "hard_violations": [
     "화면 우측 하단에서 문을 짚고 있는 오른팔에 엄지가 위로 향하는 왼손이 달려 있는 물리적으로 불가능한 해부학적 오류"
    ],
    "physics": "양손으로 문과 문틀을 짚고 체중을 지탱하고 있으나, 우측 손의 구조가 해부학적으로 성립할 수 없음."
   }
  ],
  "totals": {
   "A": 1436,
   "B": 932
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "표지판이 문 위가 아닌 옆 벽면에 위치한 점은 아쉬우나, 참조된 인물의 외모와 복장을 완벽하게 구현했으며 해부학적 오류 없이 프레이밍 및 시선 지시를 정확히 따랐습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "표지판의 위치와 전반적인 구도는 좋으나, 화면 우측 하단의 오른팔에 왼손이 결합되어 있는 치명적인 해부학적 오류가 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L59B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869633>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 '세미나실' 표지판을 문 위에 배치하라고 지시했으나, 문 오른쪽 벽면에 잘못 배치됨.",
     "fix_en": "Would move the sign from the wall to the door.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "원본 배경 이미지에 명확히 존재하는 방문 손잡이가 생성된 이미지에서는 지워져 보이지 않음.",
     "fix_en": "Would restore the missing door handle.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "차장검사가 세미나실 문 유리창에 얼굴을 바짝 들이밀지 않고 떨어져 서서 바라보고 있다.",
     "fix_en": "Redraw the man's head and upper torso leaning forward to place his face extremely close to the window. Keep his suit, the door, the corridor, the sign, the lighting, and the overall framing unchanged.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "인물이 스케치처럼 등·뒤에서 창으로 기울인 포즈가 아니라 옆모습으로 서 있다.",
     "fix_en": "Would redraw the figure from a slightly behind angle.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "잠긴 배경 사진·스케치와 다른 카메라 각도와 크롭으로 배경이 그대로 유지되지 않았다.",
     "fix_en": "Would restore the original camera angle and background perspective.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 '세미나실' 표지판을 문 위에 배치하라고 지시했으나, 문 오른쪽 벽면에 잘못 배치됨.",
     "severity": "major"
    },
    {
     "issue_ko": "원본 배경 이미지에 명확히 존재하는 방문 손잡이가 생성된 이미지에서는 지워져 보이지 않음.",
     "severity": "major"
    },
    {
     "issue_ko": "차장검사가 세미나실 문 유리창에 얼굴을 바짝 들이밀지 않고 떨어져 서서 바라보고 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "인물이 스케치처럼 등·뒤에서 창으로 기울인 포즈가 아니라 옆모습으로 서 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "세미나실 표지판이 문 위가 아니라 문 오른쪽 벽에 붙어 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "잠긴 배경 사진·스케치와 다른 카메라 각도와 크롭으로 배경이 그대로 유지되지 않았다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 '세미나실' 표지판을 문 위에 배치하라고 지시했으나, 문 오른쪽 벽면에 잘못 배치됨.",
    "fix_en": "Would move the sign from the wall to the door.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "원본 배경 이미지에 명확히 존재하는 방문 손잡이가 생성된 이미지에서는 지워져 보이지 않음.",
    "fix_en": "Would restore the missing door handle.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "인물이 스케치처럼 등·뒤에서 창으로 기울인 포즈가 아니라 옆모습으로 서 있다.",
    "fix_en": "Would redraw the figure from a slightly behind angle.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "잠긴 배경 사진·스케치와 다른 카메라 각도와 크롭으로 배경이 그대로 유지되지 않았다.",
    "fix_en": "Would restore the original camera angle and background perspective.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the man's head and upper torso leaning forward to place his face extremely close to the window. Keep his suit, the door, the corridor, the sign, the lighting, and the overall framing unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "지정된 텍스트 표지판을 포함하여 프레이밍과 인물의 디테일을 잘 구현했으나, 표지판의 위치가 문 위가 아닌 우측 벽면인 점이 약간 아쉽습니다."
     },
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "요청된 표지판이 완전히 누락되었으며, 인물의 얼굴이 창틀 위로 물리적으로 불가능하게 겹쳐지는 치명적인 렌더링 오류가 발생했습니다.  ★위반: [gemini-pro] physically impossible staging (인물의 얼굴 우측면이 안쪽으로 파인 문의 금속 창틀 위로 물리적으로 불가능하게 겹쳐져 렌더링됨)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.429
     },
     "adjusted": {
      "A": 1.571,
      "B": 1.179
     },
     "violations": {
      "B": [
       "[gemini-pro] physically impossible staging (인물의 얼굴 우측면이 안쪽으로 파인 문의 금속 창틀 위로 물리적으로 불가능하게 겹쳐져 렌더링됨)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "지정된 텍스트 표지판을 포함하여 프레이밍과 인물의 디테일을 잘 구현했으나, 표지판의 위치가 문 위가 아닌 우측 벽면인 점이 약간 아쉽습니다."
     },
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "요청된 표지판이 완전히 누락되었으며, 인물의 얼굴이 창틀 위로 물리적으로 불가능하게 겹쳐지는 치명적인 렌더링 오류가 발생했습니다.  ★위반: [gemini-pro] physically impossible staging (인물의 얼굴 우측면이 안쪽으로 파인 문의 금속 창틀 위로 물리적으로 불가능하게 겹쳐져 렌더링됨)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "요구된 미디엄 샷보다 프레이밍이 넓어 레이아웃 스케치의 구도를 벗어났으며, 필수 지정된 '세미나실' 표지판이 완전히 누락되었습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 프레이밍과 레이아웃 스케치의 크기/위치를 정확히 구현했으며, '세미나실' 텍스트가 포함된 표지판을 성공적으로 렌더링했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 닫힌 문의 작은 유리창을 향하고 있습니다.",
      "built_space": "복도와 문의 형태는 배경 원본과 일치하나, 요구된 표지판 구조물이 없습니다.",
      "entities": "차장검사의 외모와 의상(핀스트라이프 정장, 넥타이)이 레퍼런스와 일치합니다.",
      "hard_violations": [],
      "physics": "바닥을 딛고 서서 상체를 앞으로 기울인 자세가 자연스럽게 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 문의 유리창 너머를 정확히 향하고 있습니다.",
      "built_space": "복도 배경이 유지된 가운데, 문 우측 벽면에 아크릴 표지판이 배치되어 있습니다.",
      "entities": "차장검사의 외양과 복장이 레퍼런스와 일치하며, 표지판에 '세미나실' 텍스트가 정확히 쓰여 있습니다.",
      "hard_violations": [],
      "physics": "상체를 굽혀 문 쪽으로 다가간 자세가 하체에 의해 안정적으로 지탱됩니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "요구된 미디엄 샷보다 프레이밍이 넓어 레이아웃 스케치의 구도를 벗어났으며, 필수 지정된 '세미나실' 표지판이 완전히 누락되었습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 프레이밍과 레이아웃 스케치의 크기/위치를 정확히 구현했으며, '세미나실' 텍스트가 포함된 표지판을 성공적으로 렌더링했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물의 시선이 닫힌 문의 작은 유리창을 향하고 있습니다.",
      "built_space": "복도와 문의 형태는 배경 원본과 일치하나, 요구된 표지판 구조물이 없습니다.",
      "entities": "차장검사의 외모와 의상(핀스트라이프 정장, 넥타이)이 레퍼런스와 일치합니다.",
      "hard_violations": [],
      "physics": "바닥을 딛고 서서 상체를 앞으로 기울인 자세가 자연스럽게 지탱되고 있습니다."
     },
     {
      "label": "A",
      "direction": "인물의 시선이 문의 유리창 너머를 정확히 향하고 있습니다.",
      "built_space": "복도 배경이 유지된 가운데, 문 우측 벽면에 아크릴 표지판이 배치되어 있습니다.",
      "entities": "차장검사의 외양과 복장이 레퍼런스와 일치하며, 표지판에 '세미나실' 텍스트가 정확히 쓰여 있습니다.",
      "hard_violations": [],
      "physics": "상체를 굽혀 문 쪽으로 다가간 자세가 하체에 의해 안정적으로 지탱됩니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1578,
     "B": 1183
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S67sh1__bgfirst_bg.png",
   "bg_asset_id": "44d9f8df-7e42-4a49-806d-5b5e4aa30ae8",
   "bg_record_key": "S67sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S67sh1::cine": {
  "applied": true,
  "fingerprint": "50fc0cbde9f7bc0d9a5ad6a2aa5e8d00df1727962729d7b38a0fc3ffdc7ceb0f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S67sh1_sel.png",
  "source_sha256": "639f506d9f4f8855278c3dcf8f5f6685aa532afe66effc6e15db03f84e904de9",
  "file": "S67sh1_cine.png",
  "latency_ms": 10356
 },
 "S68sh1::signage": {
  "fp": "e0d77624bb409e4c",
  "inscriptions": [
   {
    "surface_native": "발표용 슬라이드 화면",
    "text_native": "실종자 정보",
    "reason_ko": "어두운 세미나실의 발표용 스크린에 띄워진 여학생 사진 옆에 '실종자 정보'라는 텍스트를 표시하여 세미나의 주제와 사건의 맥락을 명확히 전달합니다."
   }
  ]
 },
 "S68sh1": {
  "input_fingerprint": "0dcd7e28c74d9d32",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 세미나실 안, 교복 입은 소녀의 흑백 사진이 띄워진 스크린 화면 클로즈업\n\nLOCATION (lock): Inside the darkened seminar room, focused on the illuminated presentation screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The static camera sits close to the screen, slightly below its center and offset to one side, holding the projected black-and-white photograph at an oblique angle as the principal image. The school-uniformed 김선영 remains centered within the projection while a narrow screen boundary preserves the fact that the audience is viewing a displayed photograph rather than a live presence.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: projection screen showing the victim's photograph in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 세미나실 스크린 (Displaying the victim's lifetime photograph) — Its front face is visible obliquely and carries a black-and-white photograph of 김선영 alive in her school uniform; used as Serves as the image-bearing surface and primary visual subject; 스크린 경계 (Stationary) — One oblique side edge remains visible around the projected image; used as A thin visible edge distinguishes the projected display from direct reality.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The seminar room is dark, with the displayed black-and-white image held in a subdued, somber tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's school-uniform photograph remains projected on the dark seminar-room screen.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 발표용 슬라이드 화면: \"실종자 정보\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 세미나실 안, 교복 입은 소녀의 흑백 사진이 띄워진 스크린 화면 클로즈업\n\nLOCATION (lock): Inside the darkened seminar room, focused on the illuminated presentation screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The static camera sits close to the screen, slightly below its center and offset to one side, holding the projected black-and-white photograph at an oblique angle as the principal image. The school-uniformed 김선영 remains centered within the projection while a narrow screen boundary preserves the fact that the audience is viewing a displayed photograph rather than a live presence.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: projection screen showing the victim's photograph in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 세미나실 스크린 (Displaying the victim's lifetime photograph) — Its front face is visible obliquely and carries a black-and-white photograph of 김선영 alive in her school uniform; used as Serves as the image-bearing surface and primary visual subject; 스크린 경계 (Stationary) — One oblique side edge remains visible around the projected image; used as A thin visible edge distinguishes the projected display from direct reality.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The seminar room is dark, with the displayed black-and-white image held in a subdued, somber tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's school-uniform photograph remains projected on the dark seminar-room screen.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 발표용 슬라이드 화면: \"실종자 정보\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 어두운 세미나실 안, 교복 입은 소녀의 흑백 사진이 띄워진 스크린 화면 클로즈업\n\nLOCATION (lock): Inside the darkened seminar room, focused on the illuminated presentation screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The static camera sits close to the screen, slightly below its center and offset to one side, holding the projected black-and-white photograph at an oblique angle as the principal image. The school-uniformed 김선영 remains centered within the projection while a narrow screen boundary preserves the fact that the audience is viewing a displayed photograph rather than a live presence.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: projection screen showing the victim's photograph in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 세미나실 스크린 (Displaying the victim's lifetime photograph) — Its front face is visible obliquely and carries a black-and-white photograph of 김선영 alive in her school uniform; used as Serves as the image-bearing surface and primary visual subject; 스크린 경계 (Stationary) — One oblique side edge remains visible around the projected image; used as A thin visible edge distinguishes the projected display from direct reality.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The seminar room is dark, with the displayed black-and-white image held in a subdued, somber tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's school-uniform photograph remains projected on the dark seminar-room screen.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 발표용 슬라이드 화면: \"실종자 정보\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "관객과 남성이 스크린을 향하며, 마이크는 관객석 방향임.",
    "built_space": "남성이 단상 뒤가 아닌 앞쪽에 서서 스크린을 보는 모순된 위치에 있음.",
    "entities": "스크린에 지정된 요소가 표시되나, 지문에 없는 관객과 남성이 등장함.",
    "hard_violations": [
     "지문에 없는 인물(관객 및 남성) 임의 추가",
     "남성의 단상 위치 모순 (잘못된 장소에 배치됨)"
    ],
    "physics": "모든 객체가 지지면에 안정적으로 닿아 있음."
   },
   {
    "label": "B",
    "direction": "관객은 스크린을, 마이크는 관객석을 향함.",
    "built_space": "우측 하단에 빈 단상이 있으며, 스크린은 약간 비스듬히 배치됨.",
    "entities": "스크린에 지정된 문구와 사진이 있으나, 지문에 없는 관객이 전면에 다수 포함됨.",
    "hard_violations": [
     "지문에 없는 인물(관객) 임의 추가"
    ],
    "physics": "모든 객체가 안정적으로 지지됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 3,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "지문에 없는 관객을 추가하고 클로즈업 구도를 위반하여 심각한 오류가 있으나, A에 비해 공간적 모순은 없습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "지문에 없는 인물을 추가하고 구도를 위반했으며, 남성이 단상 앞쪽에 서 있는 구조적 오류까지 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S66sh3_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트에 카메라가 화면의 중심에서 벗어나 스크린을 비스듬한 각도(oblique angle)로 잡아야 한다고 명시되어 있으나, 스크린이 렌즈와 평행하게 정면으로 렌더링되었습니다.",
     "fix_en": "Skew the contents of the projection screen to simulate a slight oblique viewing angle. Preserve the screen content, the podium, the audience members, and the overall room lighting.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "연속성 오류: 레퍼런스 이미지에서 스크린 우측 단상에 서 있던 인물(발표자)이 프레임 내 동일한 공간이 보임에도 불구하고 완전히 사라졌습니다.",
     "fix_en": "Render a male speaker in a dark suit standing at the wooden podium on the right. Preserve the projection screen content, the seated audience, and the room lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "스크린 화면 클로즈업이 아니라 객석 뒤에서 스크린·청중·연단이 함께 보이는 중원경이다",
     "fix_en": "Expand the projection screen to fill most of the frame, removing the surrounding room context. Preserve the black-and-white portrait and the Korean text on the screen.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "샷 텍스트에 없는 청중 여러 명의 뒷모습이 화면 하단과 우측에 들어가 있다",
     "fix_en": "Remove the silhouetted audience heads in the foreground, filling the space with the continued dark lower section of the screen and wall. Preserve the projection screen content, the podium, and the room lighting.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "이전 샷에 있던 우측 벽면 현수막이 같은 장소의 벽에서 빠져 있다",
     "fix_en": "Add a white banner with dark text to the wall section directly above the podium. Preserve the projection screen content, the podium, the audience members, and the room lighting.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트에 카메라가 화면의 중심에서 벗어나 스크린을 비스듬한 각도(oblique angle)로 잡아야 한다고 명시되어 있으나, 스크린이 렌즈와 평행하게 정면으로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "연속성 오류: 레퍼런스 이미지에서 스크린 우측 단상에 서 있던 인물(발표자)이 프레임 내 동일한 공간이 보임에도 불구하고 완전히 사라졌습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "스크린 화면 클로즈업이 아니라 객석 뒤에서 스크린·청중·연단이 함께 보이는 중원경이다",
     "severity": "major"
    },
    {
     "issue_ko": "샷 텍스트에 없는 청중 여러 명의 뒷모습이 화면 하단과 우측에 들어가 있다",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷에 있던 우측 벽면 현수막이 같은 장소의 벽에서 빠져 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트에 카메라가 화면의 중심에서 벗어나 스크린을 비스듬한 각도(oblique angle)로 잡아야 한다고 명시되어 있으나, 스크린이 렌즈와 평행하게 정면으로 렌더링되었습니다.",
    "fix_en": "Skew the contents of the projection screen to simulate a slight oblique viewing angle. Preserve the screen content, the podium, the audience members, and the overall room lighting.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "연속성 오류: 레퍼런스 이미지에서 스크린 우측 단상에 서 있던 인물(발표자)이 프레임 내 동일한 공간이 보임에도 불구하고 완전히 사라졌습니다.",
    "fix_en": "Render a male speaker in a dark suit standing at the wooden podium on the right. Preserve the projection screen content, the seated audience, and the room lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "스크린 화면 클로즈업이 아니라 객석 뒤에서 스크린·청중·연단이 함께 보이는 중원경이다",
    "fix_en": "Expand the projection screen to fill most of the frame, removing the surrounding room context. Preserve the black-and-white portrait and the Korean text on the screen.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "샷 텍스트에 없는 청중 여러 명의 뒷모습이 화면 하단과 우측에 들어가 있다",
    "fix_en": "Remove the silhouetted audience heads in the foreground, filling the space with the continued dark lower section of the screen and wall. Preserve the projection screen content, the podium, and the room lighting.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "이전 샷에 있던 우측 벽면 현수막이 같은 장소의 벽에서 빠져 있다",
    "fix_en": "Add a white banner with dark text to the wall section directly above the podium. Preserve the projection screen content, the podium, the audience members, and the room lighting.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "needs_reshoot": true,
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S66sh3"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S68sh1::cine": {
  "applied": true,
  "fingerprint": "26354fb1ec1a5f7a9624014bec71317886e0dd1361fe19af84c6c76c7972ea0b",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S68sh1_sel.png",
  "source_sha256": "afcf3cf9f284dacce88fa0f3104d8f57a4dc131269298839b62a5c0cdba89d52",
  "file": "S68sh1_cine.png",
  "latency_ms": 10263
 },
 "S68sh2::signage": {
  "fp": "7aae8c388b57e6e5",
  "inscriptions": [
   {
    "surface_native": "프레젠테이션 화면",
    "text_native": "21세기 인문학의 흐름과 전망",
    "reason_ko": "세미나실 앞쪽 스크린에 띄워진 프레젠테이션 제목으로, 발표 중인 장원섭의 학술적 상황을 사실적으로 묘사하기 위해 필요합니다."
   }
  ]
 },
 "S68sh2": {
  "input_fingerprint": "233610d81f3a1b10",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 객석의 사람들을 향해 진지한 눈빛으로 시선을 맞춘 채 입술을 벌린 장원섭의 상체\n\nLOCATION (lock): At the front podium of the darkened seminar room, facing the audience and lit screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 장원섭's audience-side three-quarter position, the final inward movement settles slightly below his face in an upper-body composition, preserving the off-axis view rather than converting it into a frontal address. He occupies the center-left while blurred edges of seated listeners remain low in frame, and his serious gaze moves across their faces beyond the camera as he continues speaking.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 시민위원 객석 (Listeners seated with individually varied head angles and attentive postures); used as Soft shoulder and head fragments at the lower frame establish the recipients of 장원섭's address without becoming a uniform crowd; 세미나실 스크린 (Displaying the photograph) — The front display face remains partially visible and shows the victim's lifetime photograph; used as A restrained background portion carries the victim photograph behind the ongoing explanation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark seminar room is rendered with subdued neutral-to-cool tonality and controlled contrast around 장원섭's tense, serious face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same seminar room, screen glow, seated committee, and prosecutor's suit. Exclude the earlier bright presentation slide and retain the darker room with the life photograph displayed.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's photograph remains projected behind Wonseop as he addresses the citizen committee.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 프레젠테이션 화면: \"21세기 인문학의 흐름과 전망\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 객석의 사람들을 향해 진지한 눈빛으로 시선을 맞춘 채 입술을 벌린 장원섭의 상체\n\nLOCATION (lock): At the front podium of the darkened seminar room, facing the audience and lit screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 장원섭's audience-side three-quarter position, the final inward movement settles slightly below his face in an upper-body composition, preserving the off-axis view rather than converting it into a frontal address. He occupies the center-left while blurred edges of seated listeners remain low in frame, and his serious gaze moves across their faces beyond the camera as he continues speaking.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 시민위원 객석 (Listeners seated with individually varied head angles and attentive postures); used as Soft shoulder and head fragments at the lower frame establish the recipients of 장원섭's address without becoming a uniform crowd; 세미나실 스크린 (Displaying the photograph) — The front display face remains partially visible and shows the victim's lifetime photograph; used as A restrained background portion carries the victim photograph behind the ongoing explanation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark seminar room is rendered with subdued neutral-to-cool tonality and controlled contrast around 장원섭's tense, serious face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same seminar room, screen glow, seated committee, and prosecutor's suit. Exclude the earlier bright presentation slide and retain the darker room with the life photograph displayed.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's photograph remains projected behind Wonseop as he addresses the citizen committee.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 프레젠테이션 화면: \"21세기 인문학의 흐름과 전망\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 객석의 사람들을 향해 진지한 눈빛으로 시선을 맞춘 채 입술을 벌린 장원섭의 상체\n\nLOCATION (lock): At the front podium of the darkened seminar room, facing the audience and lit screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 장원섭's audience-side three-quarter position, the final inward movement settles slightly below his face in an upper-body composition, preserving the off-axis view rather than converting it into a frontal address. He occupies the center-left while blurred edges of seated listeners remain low in frame, and his serious gaze moves across their faces beyond the camera as he continues speaking.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 시민위원 객석 (Listeners seated with individually varied head angles and attentive postures); used as Soft shoulder and head fragments at the lower frame establish the recipients of 장원섭's address without becoming a uniform crowd; 세미나실 스크린 (Displaying the photograph) — The front display face remains partially visible and shows the victim's lifetime photograph; used as A restrained background portion carries the victim photograph behind the ongoing explanation.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The dark seminar room is rendered with subdued neutral-to-cool tonality and controlled contrast around 장원섭's tense, serious face.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same seminar room, screen glow, seated committee, and prosecutor's suit. Exclude the earlier bright presentation slide and retain the darker room with the life photograph displayed.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's photograph remains projected behind Wonseop as he addresses the citizen committee.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 프레젠테이션 화면: \"21세기 인문학의 흐름과 전망\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "장원섭의 시선과 몸의 방향이 우측 오프스크린의 청중을 향하며 입술을 벌리고 있음.",
    "built_space": "세미나실 전면 단상에 마이크가 있으며, 배경 스크린에 지정된 텍스트와 사진이 올바르게 배치됨.",
    "entities": "장원섭(레퍼런스와 일치하는 외모 및 정장), 화면 하단의 청중들(머리와 어깨), 지정된 텍스트와 흑백 사진이 정확히 묘사됨.",
    "hard_violations": [],
    "physics": "오른손을 단상에 올린 채 자연스럽게 체중을 지지하며 서 있음."
   },
   {
    "label": "B",
    "direction": "시선과 상체가 거의 정면(카메라 방향)을 향하고 있어 오프액시스 지시에서 벗어남.",
    "built_space": "세미나실 전면 단상과 배경 스크린이 보이나 이전 샷에 있던 마이크가 누락됨.",
    "entities": "장원섭(레퍼런스와 일치), 전경의 청중들, 스크린 속 사진과 텍스트가 묘사됨.",
    "hard_violations": [],
    "physics": "양손을 단상에 올리고 안정적으로 서 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지정된 3/4 측면 앵글과 오프액시스 시선을 정확히 구현했으며, 프레젠테이션 텍스트와 이전 샷의 디테일(마이크 등)을 훌륭하게 유지함."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "정면 응시로 변환하지 말라는 카메라 지시를 어기고 피사체가 정면을 향하고 있으며, 단상의 마이크 디테일이 누락됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S68sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 숏 레퍼런스 이미지에서는 스크린 우측에 있던 단상의 위치가 생성된 이미지에서는 스크린 좌측으로 이동하여 공간 배치의 일관성이 어긋납니다.",
     "fix_en": "Rebuild room geometry to place the podium on the right side of the screen. Maintain the speaker, screen text, and lighting.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "하단 청중 머리·어깨가 흐릿한 단편이 아니라 선명하고 크게 보인다",
     "fix_en": "Apply strong blur to the foreground audience heads. Maintain the speaker, background screen, and lighting.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "스크린이 부분적으로만 보이는 절제된 배경이 아니라 프레임 오른쪽을 거의 가득 채운다",
     "fix_en": "Adjust camera framing to reduce the screen's share of the frame. Maintain the speaker, audience, and lighting.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "객석 뒷모습이 제각각 다른 머리 각도가 아니라 비슷한 정면 주시로 나란하다",
     "fix_en": "Vary the head angles of the foreground audience. Maintain the speaker, background, and framing.",
     "severity": "minor",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이전 숏 레퍼런스 이미지에서는 스크린 우측에 있던 단상의 위치가 생성된 이미지에서는 스크린 좌측으로 이동하여 공간 배치의 일관성이 어긋납니다.",
     "severity": "major"
    },
    {
     "issue_ko": "장원섭이 관객 측 3/4 오프액시스가 아니라 카메라에 거의 정면으로 서서 연설하는 구도이다",
     "severity": "major"
    },
    {
     "issue_ko": "하단 청중 머리·어깨가 흐릿한 단편이 아니라 선명하고 크게 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "스크린이 부분적으로만 보이는 절제된 배경이 아니라 프레임 오른쪽을 거의 가득 채운다",
     "severity": "major"
    },
    {
     "issue_ko": "객석 뒷모습이 제각각 다른 머리 각도가 아니라 비슷한 정면 주시로 나란하다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 숏 레퍼런스 이미지에서는 스크린 우측에 있던 단상의 위치가 생성된 이미지에서는 스크린 좌측으로 이동하여 공간 배치의 일관성이 어긋납니다.",
    "fix_en": "Rebuild room geometry to place the podium on the right side of the screen. Maintain the speaker, screen text, and lighting.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "하단 청중 머리·어깨가 흐릿한 단편이 아니라 선명하고 크게 보인다",
    "fix_en": "Apply strong blur to the foreground audience heads. Maintain the speaker, background screen, and lighting.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "스크린이 부분적으로만 보이는 절제된 배경이 아니라 프레임 오른쪽을 거의 가득 채운다",
    "fix_en": "Adjust camera framing to reduce the screen's share of the frame. Maintain the speaker, audience, and lighting.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "객석 뒷모습이 제각각 다른 머리 각도가 아니라 비슷한 정면 주시로 나란하다",
    "fix_en": "Vary the head angles of the foreground audience. Maintain the speaker, background, and framing.",
    "severity": "minor",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S68sh1"
  }
 },
 "S68sh2::cine": {
  "applied": true,
  "fingerprint": "a85c572c4b315c0895a5eda9f9cfe8c55347597e2744d165b68bb3c26502cecb",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S68sh2_sel.png",
  "source_sha256": "f371425643995ccfd6d808ef2ae9f74a8589a2cee57874b5feb7af6c2a1d634e",
  "file": "S68sh2_cine.png",
  "latency_ms": 12899
 },
 "S69sh1::signage": {
  "fp": "fd3660a7d93ee575",
  "inscriptions": [
   {
    "surface_native": "연단 전면",
    "text_native": "검찰",
    "reason_ko": "대한민국 검찰청의 브리핑룸 연단임을 보여주는 공식 표식입니다."
   },
   {
    "surface_native": "보도자료 문서",
    "text_native": "보도자료",
    "reason_ko": "차장검사가 들고 발표하는 문서의 성격을 명확히 드러내기 위해 필요한 표기입니다."
   }
  ]
 },
 "era_assess::491c7cd51a839974": {
  "subjects": [
   {
    "subject_native": "대한민국 기자회견장 포디움 (2000년대-2010년대)",
    "search_terms_native": [
     "기자회견 포디움",
     "기자회견 브리핑룸",
     "정부 브리핑 마이크",
     "기자회견 백드롭"
    ],
    "language_lock_native": "한국어 이외의 다른 언어로 번역하거나 검색어를 추가하지 마십시오.",
    "reason_ko": "한국의 기자회견장은 정부나 기업 브리핑 시 특정 마이크 배치, 방송사 로고 큐브, 독특한 휘장 및 백드롭 디자인 등 서구권과 차별화되는 고유한 시각적 구성을 가지고 있습니다."
   }
  ]
 },
 "era_ref::012496322dad0b9f": {
  "subject": "대한민국 기자회견장 포디움 (2000년대-2010년대)",
  "terms": [
   "기자회견 포디움",
   "기자회견 브리핑룸",
   "정부 브리핑 마이크",
   "기자회견 백드롭"
  ],
  "queries": [
   [
    "기자회견 포디움 기자회견 브리핑룸",
    "정부 브리핑 마이크 기자회견 백드롭"
   ],
   [
    "대한민국 기자회견장 포디움 2000년대 2010년대",
    "기자회견 포디움 기자회견 브리핑룸 정부 브리핑 마이크 기자회견 백드롭"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://thumbnews.nateimg.co.kr/view610/news.nateimg.co.kr/orgImg/ni/2025/05/09/250509121029978_w.jpg",
  "picked_reason_ko": "사진 3은 2016년 이후 대한민국정부 기자회견장에서 쓰인 포디움의 경사 상판, 각진 몸체, 목재 지지부, 표찰과 마이크 구성을 정면에서 가장 분명하게 보여준다.",
  "sha256": "30ac9f737d5261fca1d92e6ac27c57f318137f228d27a937188243ca0b8128e0",
  "file": "eraref_012496322dad0b9f.png"
 },
 "S69sh1::bgfirst_bg": {
  "input_fingerprint": "719a2841b237a78e",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수많은 카메라 플래시 불빛이 번쩍이는 가운데 단상에 서서 보도자료를 쥔 차장검사의 전경\n\nLOCATION (lock): At the press-room podium, facing rows of reporters and flashing cameras under indoor lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the front-left side of the briefing room and slightly below 차장검사's eyeline, a static medium-wide three-quarter composition places him and the podium in the left-center while reporter rows recede along the right. He braces behind the podium with the press release in both hands, reading downward as camera flashes intermittently strike his face and the seated reporters continue typing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 브리핑 단상과 연단 (Occupied by 차장검사) — The front and left side of the podium are visible from the camera's three-quarter position; used as Anchors 차장검사 at the front of the briefing room and separates him from the press rows; 보도자료 (Being read) — The printed face is angled upward toward 차장검사, leaving the camera a foreshortened view without requiring readable copy; used as Held in the speaker's hands as the textual basis of the briefing; 기자석 (Reporters seated and typing on laptops at differing moments and postures); used as Creates depth through staggered seated figures and open computers; 기자들의 노트북 (Open and in use) — The display faces are directed toward their individual users, while the camera sees varied side and rear angles; used as Supports the reporters' ongoing typing action in the receding rows.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime briefing-room ambience is punctuated by repeated white camera flashes across 차장검사's face, retaining restrained neutral-to-cool color between bursts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 기자회견장 포디움 (2000년대-2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 수많은 카메라 플래시 불빛이 번쩍이는 가운데 단상에 서서 보도자료를 쥔 차장검사의 전경\n\nLOCATION (lock): At the press-room podium, facing rows of reporters and flashing cameras under indoor lighting.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the front-left side of the briefing room and slightly below 차장검사's eyeline, a static medium-wide three-quarter composition places him and the podium in the left-center while reporter rows recede along the right. He braces behind the podium with the press release in both hands, reading downward as camera flashes intermittently strike his face and the seated reporters continue typing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 브리핑 단상과 연단 (Occupied by 차장검사) — The front and left side of the podium are visible from the camera's three-quarter position; used as Anchors 차장검사 at the front of the briefing room and separates him from the press rows; 보도자료 (Being read) — The printed face is angled upward toward 차장검사, leaving the camera a foreshortened view without requiring readable copy; used as Held in the speaker's hands as the textual basis of the briefing; 기자석 (Reporters seated and typing on laptops at differing moments and postures); used as Creates depth through staggered seated figures and open computers; 기자들의 노트북 (Open and in use) — The display faces are directed toward their individual users, while the camera sees varied side and rear angles; used as Supports the reporters' ongoing typing action in the receding rows.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime briefing-room ambience is punctuated by repeated white camera flashes across 차장검사's face, retaining restrained neutral-to-cool color between bursts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 기자회견장 포디움 (2000년대-2010년대): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S69sh1__bgfirst_bg.png",
  "asset_id": "70801f0b-27a0-4796-b9bb-8308d956248a",
  "input_asset_ids": [
   "9e277d6e-d99d-44b2-b2c5-3c425cdcff5d",
   "8104a422-e9a4-4e82-affa-9bcdc9faa2ee"
  ],
  "era_research": {
   "subject": "대한민국 기자회견장 포디움 (2000년대-2010년대)",
   "queries": [
    [
     "기자회견 포디움 기자회견 브리핑룸",
     "정부 브리핑 마이크 기자회견 백드롭"
    ],
    [
     "대한민국 기자회견장 포디움 2000년대 2010년대",
     "기자회견 포디움 기자회견 브리핑룸 정부 브리핑 마이크 기자회견 백드롭"
    ]
   ],
   "picked_url": "https://thumbnews.nateimg.co.kr/view610/news.nateimg.co.kr/orgImg/ni/2025/05/09/250509121029978_w.jpg",
   "sha256": "30ac9f737d5261fca1d92e6ac27c57f318137f228d27a937188243ca0b8128e0",
   "file": "eraref_012496322dad0b9f.png"
  }
 },
 "S69sh1": {
  "input_fingerprint": "10180759eb39e967",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수많은 카메라 플래시 불빛이 번쩍이는 가운데 단상에 서서 보도자료를 쥔 차장검사의 전경\n\nLOCATION (lock): At the press-room podium, facing rows of reporters and flashing cameras under indoor lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the front-left side of the briefing room and slightly below 차장검사's eyeline, a static medium-wide three-quarter composition places him and the podium in the left-center while reporter rows recede along the right. He braces behind the podium with the press release in both hands, reading downward as camera flashes intermittently strike his face and the seated reporters continue typing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 브리핑 단상과 연단 (Occupied by 차장검사) — The front and left side of the podium are visible from the camera's three-quarter position; used as Anchors 차장검사 at the front of the briefing room and separates him from the press rows; 보도자료 (Being read) — The printed face is angled upward toward 차장검사, leaving the camera a foreshortened view without requiring readable copy; used as Held in the speaker's hands as the textual basis of the briefing; 기자석 (Reporters seated and typing on laptops at differing moments and postures); used as Creates depth through staggered seated figures and open computers; 기자들의 노트북 (Open and in use) — The display faces are directed toward their individual users, while the camera sees varied side and rear angles; used as Supports the reporters' ongoing typing action in the receding rows.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime briefing-room ambience is punctuated by repeated white camera flashes across 차장검사's face, retaining restrained neutral-to-cool color between bursts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The deputy chief prosecutor continues holding and reading from the prepared press release at the podium.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 차장검사 right now — 차장검사's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 차장검사 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 연단 전면: \"검찰\"\n- 보도자료 문서: \"보도자료\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수많은 카메라 플래시 불빛이 번쩍이는 가운데 단상에 서서 보도자료를 쥔 차장검사의 전경\n\nLOCATION (lock): At the press-room podium, facing rows of reporters and flashing cameras under indoor lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the front-left side of the briefing room and slightly below 차장검사's eyeline, a static medium-wide three-quarter composition places him and the podium in the left-center while reporter rows recede along the right. He braces behind the podium with the press release in both hands, reading downward as camera flashes intermittently strike his face and the seated reporters continue typing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 브리핑 단상과 연단 (Occupied by 차장검사) — The front and left side of the podium are visible from the camera's three-quarter position; used as Anchors 차장검사 at the front of the briefing room and separates him from the press rows; 보도자료 (Being read) — The printed face is angled upward toward 차장검사, leaving the camera a foreshortened view without requiring readable copy; used as Held in the speaker's hands as the textual basis of the briefing; 기자석 (Reporters seated and typing on laptops at differing moments and postures); used as Creates depth through staggered seated figures and open computers; 기자들의 노트북 (Open and in use) — The display faces are directed toward their individual users, while the camera sees varied side and rear angles; used as Supports the reporters' ongoing typing action in the receding rows.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime briefing-room ambience is punctuated by repeated white camera flashes across 차장검사's face, retaining restrained neutral-to-cool color between bursts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The deputy chief prosecutor continues holding and reading from the prepared press release at the podium.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 차장검사 right now — 차장검사's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 차장검사 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 연단 전면: \"검찰\"\n- 보도자료 문서: \"보도자료\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 수많은 카메라 플래시 불빛이 번쩍이는 가운데 단상에 서서 보도자료를 쥔 차장검사의 전경\n\nLOCATION (lock): At the press-room podium, facing rows of reporters and flashing cameras under indoor lighting. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the front-left side of the briefing room and slightly below 차장검사's eyeline, a static medium-wide three-quarter composition places him and the podium in the left-center while reporter rows recede along the right. He braces behind the podium with the press release in both hands, reading downward as camera flashes intermittently strike his face and the seated reporters continue typing.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 브리핑 단상과 연단 (Occupied by 차장검사) — The front and left side of the podium are visible from the camera's three-quarter position; used as Anchors 차장검사 at the front of the briefing room and separates him from the press rows; 보도자료 (Being read) — The printed face is angled upward toward 차장검사, leaving the camera a foreshortened view without requiring readable copy; used as Held in the speaker's hands as the textual basis of the briefing; 기자석 (Reporters seated and typing on laptops at differing moments and postures); used as Creates depth through staggered seated figures and open computers; 기자들의 노트북 (Open and in use) — The display faces are directed toward their individual users, while the camera sees varied side and rear angles; used as Supports the reporters' ongoing typing action in the receding rows.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime briefing-room ambience is punctuated by repeated white camera flashes across 차장검사's face, retaining restrained neutral-to-cool color between bursts.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The deputy chief prosecutor continues holding and reading from the prepared press release at the podium.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 차장검사 right now — 차장검사's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 차장검사 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 차장검사 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 빗어 넘긴 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 연단 전면: \"검찰\"\n- 보도자료 문서: \"보도자료\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S69sh1__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S69sh1.png"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L60B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:869633>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 와이드 샷 프레이밍과 위치 묘사, 인물의 외모 및 텍스트 요구사항을 훌륭하게 구현했으나, 지시되지 않은 배경 텍스트와 단상 발판이 추가된 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 텍스트와 인물 묘사는 양호하나, 필수적인 와이드 샷 프레이밍을 어기고 지시문에 없는 카메라맨들을 등장시켜 공간 구조를 왜곡했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "차장검사는 양손으로 든 보도자료를 아래로 내려다보고 있으며, 우측의 기자들은 각자의 노트북 화면에 시선을 두고 있다. 좌측의 삼각대 카메라들은 연단을 향해 맞춰져 있고 플래시가 터지고 있다.",
      "built_space": "카메라가 좌측 전면에서 우측으로 깊어지는 브리핑룸을 와이드 샷으로 포착했으며 우측 문, 좌측 창문 등 기준 사진의 구조를 잘 따랐다. 다만 단상 아래에 기준 사진에 없는 회색 단(발판)이 존재한다.",
      "entities": "차장검사는 레퍼런스와 일치하는 50대 중반의 외모, 헤어스타일, 스트라이프 정장을 정확히 갖췄다. 문서의 '보도자료', 연단의 '검찰' 텍스트가 정확히 표기되었으나 배경에 지시되지 않은 '대한민국정부' 문구가 쓰여 있다.",
      "hard_violations": [
       "지시문에 없는 임의의 배경 텍스트('대한민국정부') 생성",
       "기준 공간에 존재하지 않는 발판(회색 단) 추가"
      ],
      "physics": "차장검사는 두 발로 안정적으로 서서 양손으로 문서를 쥐고 있으며, 기자들은 의자에 앉아 노트북 키보드에 손을 올린 채 지지되고 있다. 공중에 떠 있거나 구조적으로 불가능한 자세는 없다."
     },
     {
      "label": "B",
      "direction": "차장검사는 손에 든 보도자료를 향해 시선을 내리고 있으며, 우측 배경의 카메라맨들은 렌즈를 차장검사 쪽으로 겨냥하고 있다. 기자들은 각자의 노트북을 응시하고 있다.",
      "built_space": "지시된 와이드 샷 대신 미디엄 샷으로 프레이밍되었으며, 우측 벽에 기준 사진에 없는 창문이 생성되었다. 단상 좌측 뒤편에 공간 구조상 존재하지 않는 기자석이 배치되었다.",
      "entities": "차장검사의 외모와 의상은 레퍼런스와 잘 일치하며, '보도자료' 및 '검찰' 텍스트 역시 정확히 묘사되었다. 그러나 프롬프트에서 허용하지 않은 다수의 서 있는 카메라맨들이 등장한다.",
      "hard_violations": [
       "지시문에 명시되지 않은 인물(서 있는 카메라맨들) 추가",
       "공간 구조와 무관한 잘못된 위치(단상 좌측 뒤편)에 앉은 기자"
      ],
      "physics": "차장검사가 문서를 쥐고 서 있는 자세와 각도를 비롯해 기자들과 카메라맨들의 자세, 장비 파지법 등은 중력과 물리법칙에 맞게 자연스럽게 지지되어 있다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 와이드 샷 프레이밍과 위치 묘사, 인물의 외모 및 텍스트 요구사항을 훌륭하게 구현했으나, 지시되지 않은 배경 텍스트와 단상 발판이 추가된 점이 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "지정된 텍스트와 인물 묘사는 양호하나, 필수적인 와이드 샷 프레이밍을 어기고 지시문에 없는 카메라맨들을 등장시켜 공간 구조를 왜곡했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "차장검사는 양손으로 든 보도자료를 아래로 내려다보고 있으며, 우측의 기자들은 각자의 노트북 화면에 시선을 두고 있다. 좌측의 삼각대 카메라들은 연단을 향해 맞춰져 있고 플래시가 터지고 있다.",
      "built_space": "카메라가 좌측 전면에서 우측으로 깊어지는 브리핑룸을 와이드 샷으로 포착했으며 우측 문, 좌측 창문 등 기준 사진의 구조를 잘 따랐다. 다만 단상 아래에 기준 사진에 없는 회색 단(발판)이 존재한다.",
      "entities": "차장검사는 레퍼런스와 일치하는 50대 중반의 외모, 헤어스타일, 스트라이프 정장을 정확히 갖췄다. 문서의 '보도자료', 연단의 '검찰' 텍스트가 정확히 표기되었으나 배경에 지시되지 않은 '대한민국정부' 문구가 쓰여 있다.",
      "hard_violations": [
       "지시문에 없는 임의의 배경 텍스트('대한민국정부') 생성",
       "기준 공간에 존재하지 않는 발판(회색 단) 추가"
      ],
      "physics": "차장검사는 두 발로 안정적으로 서서 양손으로 문서를 쥐고 있으며, 기자들은 의자에 앉아 노트북 키보드에 손을 올린 채 지지되고 있다. 공중에 떠 있거나 구조적으로 불가능한 자세는 없다."
     },
     {
      "label": "B",
      "direction": "차장검사는 손에 든 보도자료를 향해 시선을 내리고 있으며, 우측 배경의 카메라맨들은 렌즈를 차장검사 쪽으로 겨냥하고 있다. 기자들은 각자의 노트북을 응시하고 있다.",
      "built_space": "지시된 와이드 샷 대신 미디엄 샷으로 프레이밍되었으며, 우측 벽에 기준 사진에 없는 창문이 생성되었다. 단상 좌측 뒤편에 공간 구조상 존재하지 않는 기자석이 배치되었다.",
      "entities": "차장검사의 외모와 의상은 레퍼런스와 잘 일치하며, '보도자료' 및 '검찰' 텍스트 역시 정확히 묘사되었다. 그러나 프롬프트에서 허용하지 않은 다수의 서 있는 카메라맨들이 등장한다.",
      "hard_violations": [
       "지시문에 명시되지 않은 인물(서 있는 카메라맨들) 추가",
       "공간 구조와 무관한 잘못된 위치(단상 좌측 뒤편)에 앉은 기자"
      ],
      "physics": "차장검사가 문서를 쥐고 서 있는 자세와 각도를 비롯해 기자들과 카메라맨들의 자세, 장비 파지법 등은 중력과 물리법칙에 맞게 자연스럽게 지지되어 있다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 공간과 인물의 특징을 충실히 구현했으나, 요구된 와이드 샷이 아닌 미디엄 샷으로 렌더링되었고 문서의 텍스트가 사용자가 아닌 카메라를 향해 있는 점이 아쉬움."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "와이드 샷 프레이밍은 준수했으나, 레퍼런스에 없는 플랫폼을 추가하고 공간 배열을 완전히 틀어버리는 등 치명적인 구조 왜곡이 발생함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "차장검사의 시선은 손에 쥔 보도자료를 향함. 카메라 플래시는 단상을 비추고 기자들은 노트북 화면을 바라봄.",
      "built_space": "파란색 백드롭과 나무 벽면 등 레퍼런스의 구조를 유지함. 단상은 바닥에 위치하며 측면 앵글이 적절히 적용됨.",
      "entities": "차장검사의 외모와 복장이 레퍼런스와 일치. 연단의 '검찰', 문서의 '보도자료' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "차장검사의 두 손이 문서를 쥐고 있으며, 기자들의 손은 키보드에 위치해 물리적 지지가 자연스러움."
     },
     {
      "label": "B",
      "direction": "차장검사는 아래 문서를 내려다보며, 기자들의 시선은 각자의 노트북 모니터로 향해 있음.",
      "built_space": "파란색 백드롭 앞에 있어야 할 단상이 측면 벽으로 이동했으며, 책상의 배열이 90도 틀어져 공간 구조가 붕괴됨.",
      "entities": "차장검사의 인상착의가 일치하며, 지시된 두 가지 텍스트 표기 조건도 모두 정확히 반영됨.",
      "hard_violations": [
       "레퍼런스에 없는 회색 플랫폼(단상) 구조물 임의 생성",
       "단상과 기자석의 방향 및 위치를 왜곡하여 로케이션 고정(Location lock) 조건 위반"
      ],
      "physics": "모든 인물이 바닥이나 의자에 안정적으로 지지되어 있고, 문서를 든 손의 파지도 정상적임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 공간과 인물의 특징을 충실히 구현했으나, 요구된 와이드 샷이 아닌 미디엄 샷으로 렌더링되었고 문서의 텍스트가 사용자가 아닌 카메라를 향해 있는 점이 아쉬움."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "와이드 샷 프레이밍은 준수했으나, 레퍼런스에 없는 플랫폼을 추가하고 공간 배열을 완전히 틀어버리는 등 치명적인 구조 왜곡이 발생함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "차장검사의 시선은 손에 쥔 보도자료를 향함. 카메라 플래시는 단상을 비추고 기자들은 노트북 화면을 바라봄.",
      "built_space": "파란색 백드롭과 나무 벽면 등 레퍼런스의 구조를 유지함. 단상은 바닥에 위치하며 측면 앵글이 적절히 적용됨.",
      "entities": "차장검사의 외모와 복장이 레퍼런스와 일치. 연단의 '검찰', 문서의 '보도자료' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "차장검사의 두 손이 문서를 쥐고 있으며, 기자들의 손은 키보드에 위치해 물리적 지지가 자연스러움."
     },
     {
      "label": "A",
      "direction": "차장검사는 아래 문서를 내려다보며, 기자들의 시선은 각자의 노트북 모니터로 향해 있음.",
      "built_space": "파란색 백드롭 앞에 있어야 할 단상이 측면 벽으로 이동했으며, 책상의 배열이 90도 틀어져 공간 구조가 붕괴됨.",
      "entities": "차장검사의 인상착의가 일치하며, 지시된 두 가지 텍스트 표기 조건도 모두 정확히 반영됨.",
      "hard_violations": [
       "레퍼런스에 없는 회색 플랫폼(단상) 구조물 임의 생성",
       "단상과 기자석의 방향 및 위치를 왜곡하여 로케이션 고정(Location lock) 조건 위반"
      ],
      "physics": "모든 인물이 바닥이나 의자에 안정적으로 지지되어 있고, 문서를 든 손의 파지도 정상적임."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 10
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "차장검사는 양손으로 든 보도자료를 아래로 내려다보고 있으며, 우측의 기자들은 각자의 노트북 화면에 시선을 두고 있다. 좌측의 삼각대 카메라들은 연단을 향해 맞춰져 있고 플래시가 터지고 있다.",
    "built_space": "카메라가 좌측 전면에서 우측으로 깊어지는 브리핑룸을 와이드 샷으로 포착했으며 우측 문, 좌측 창문 등 기준 사진의 구조를 잘 따랐다. 다만 단상 아래에 기준 사진에 없는 회색 단(발판)이 존재한다.",
    "entities": "차장검사는 레퍼런스와 일치하는 50대 중반의 외모, 헤어스타일, 스트라이프 정장을 정확히 갖췄다. 문서의 '보도자료', 연단의 '검찰' 텍스트가 정확히 표기되었으나 배경에 지시되지 않은 '대한민국정부' 문구가 쓰여 있다.",
    "hard_violations": [
     "지시문에 없는 임의의 배경 텍스트('대한민국정부') 생성",
     "기준 공간에 존재하지 않는 발판(회색 단) 추가"
    ],
    "physics": "차장검사는 두 발로 안정적으로 서서 양손으로 문서를 쥐고 있으며, 기자들은 의자에 앉아 노트북 키보드에 손을 올린 채 지지되고 있다. 공중에 떠 있거나 구조적으로 불가능한 자세는 없다."
   },
   {
    "label": "B",
    "direction": "차장검사는 손에 든 보도자료를 향해 시선을 내리고 있으며, 우측 배경의 카메라맨들은 렌즈를 차장검사 쪽으로 겨냥하고 있다. 기자들은 각자의 노트북을 응시하고 있다.",
    "built_space": "지시된 와이드 샷 대신 미디엄 샷으로 프레이밍되었으며, 우측 벽에 기준 사진에 없는 창문이 생성되었다. 단상 좌측 뒤편에 공간 구조상 존재하지 않는 기자석이 배치되었다.",
    "entities": "차장검사의 외모와 의상은 레퍼런스와 잘 일치하며, '보도자료' 및 '검찰' 텍스트 역시 정확히 묘사되었다. 그러나 프롬프트에서 허용하지 않은 다수의 서 있는 카메라맨들이 등장한다.",
    "hard_violations": [
     "지시문에 명시되지 않은 인물(서 있는 카메라맨들) 추가",
     "공간 구조와 무관한 잘못된 위치(단상 좌측 뒤편)에 앉은 기자"
    ],
    "physics": "차장검사가 문서를 쥐고 서 있는 자세와 각도를 비롯해 기자들과 카메라맨들의 자세, 장비 파지법 등은 중력과 물리법칙에 맞게 자연스럽게 지지되어 있다."
   }
  ],
  "totals": {
   "A": 10,
   "B": 10
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 와이드 샷 프레이밍과 위치 묘사, 인물의 외모 및 텍스트 요구사항을 훌륭하게 구현했으나, 지시되지 않은 배경 텍스트와 단상 발판이 추가된 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "지정된 텍스트와 인물 묘사는 양호하나, 필수적인 와이드 샷 프레이밍을 어기고 지시문에 없는 카메라맨들을 등장시켜 공간 구조를 왜곡했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L60B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 차장검사: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:869633>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "차장검사가 들고 있는 보도자료 문서 왼쪽 가장자리에 본인의 양손 외에 정체불명의 세 번째 손가락 무리가 생성되어 있습니다.",
     "fix_en": "Remove the extra set of fingers gripping the left edge of the paper so only the speaker's two hands hold the document, preserving the man's face, hair, suit, posture, the podium, the reporters, the room, lighting, and framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "화자가 읽고 있어야 할 '보도자료' 텍스트가 문서의 앞면이 아닌 카메라를 향한 뒷면에 인쇄되어 있어, 사물의 기능적 방향 지시를 위반했습니다.",
     "fix_en": "Erase the black text from the side of the paper facing the camera, leaving the reverse side blank white as staged, preserving the man's face, hair, suit, posture, hands, the podium, the reporters, the room, lighting, and framing.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "배경을 변경 없이 유지하라는 지시에도 불구하고, 원본 배경 사진에는 존재하지 않는 회색 단상이 연단과 차장검사 아래에 임의로 추가되었습니다.",
     "fix_en": "Remove the gray platform to place the podium and speaker directly on the original tiled floor.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "연단 전면 글자가 '검찰'이 아니라 가운데가 띄어진 '검 찰'이다",
     "fix_en": "Close the gap between the characters to read '검찰' without a space.",
     "severity": "minor",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "차장검사가 들고 있는 보도자료 문서 왼쪽 가장자리에 본인의 양손 외에 정체불명의 세 번째 손가락 무리가 생성되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화자가 읽고 있어야 할 '보도자료' 텍스트가 문서의 앞면이 아닌 카메라를 향한 뒷면에 인쇄되어 있어, 사물의 기능적 방향 지시를 위반했습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경을 변경 없이 유지하라는 지시에도 불구하고, 원본 배경 사진에는 존재하지 않는 회색 단상이 연단과 차장검사 아래에 임의로 추가되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "원본 장소에 없는 회색 단상 플랫폼이 연단 발밑을 받치고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "연단이 파란 배경벽 바로 앞이 아니라 앞으로 크게 떨어져 있다",
     "severity": "major"
    },
    {
     "issue_ko": "연단 전면 글자가 '검찰'이 아니라 가운데가 띄어진 '검 찰'이다",
     "severity": "minor"
    },
    {
     "issue_ko": "보도자료 글자가 단축되어 안 읽혀야 하는데 카메라에 정면으로 또렷이 보인다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "배경을 변경 없이 유지하라는 지시에도 불구하고, 원본 배경 사진에는 존재하지 않는 회색 단상이 연단과 차장검사 아래에 임의로 추가되었습니다.",
    "fix_en": "Remove the gray platform to place the podium and speaker directly on the original tiled floor.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "연단 전면 글자가 '검찰'이 아니라 가운데가 띄어진 '검 찰'이다",
    "fix_en": "Close the gap between the characters to read '검찰' without a space.",
    "severity": "minor",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the extra set of fingers gripping the left edge of the paper so only the speaker's two hands hold the document, preserving the man's face, hair, suit, posture, the podium, the reporters, the room, lighting, and framing.\n- Erase the black text from the side of the paper facing the camera, leaving the reverse side blank white as staged, preserving the man's face, hair, suit, posture, hands, the podium, the reporters, the room, lighting, and framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 3/4 앵글과 스케치 레이아웃을 정확히 따랐으며, 요구된 텍스트와 기자석 묘사가 충실함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "구도와 스케치를 무시하고 통로 중앙에 인물을 배치했으며, 기자와 텍스트가 모두 누락됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "차장검사는 아래의 문서를, 기자들은 노트북을 향해 시선을 둠.",
      "built_space": "좌측 연단과 우측 기자석 배열이 스케치 및 요구 구도와 일치함.",
      "entities": "인물 외형이 레퍼런스와 일치하고, 요구된 텍스트('검찰', '보도자료')와 기자들이 모두 존재함.",
      "hard_violations": [],
      "physics": "단상 뒤에서 문서를 들고 서 있으며, 기자들은 의자에 정상적으로 착석함."
     },
     {
      "label": "B",
      "direction": "인물이 정면 카메라 렌즈를 응시함.",
      "built_space": "연단을 사용하지 않고 통로 중앙에 서 있음.",
      "entities": "기자들이 전혀 없으며, 지정된 텍스트 표기도 누락됨.",
      "hard_violations": [
       "인물이 단상이 아닌 통로에 배치됨",
       "기자석 및 기자 누락",
       "지정된 구도 및 레이아웃 스케치 완전 무시"
      ],
      "physics": "두 발로 바닥에 서서 문서를 쥐고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지정된 3/4 앵글과 스케치 레이아웃을 정확히 따랐으며, 요구된 텍스트와 기자석 묘사가 충실함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "구도와 스케치를 무시하고 통로 중앙에 인물을 배치했으며, 기자와 텍스트가 모두 누락됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "차장검사는 아래의 문서를, 기자들은 노트북을 향해 시선을 둠.",
      "built_space": "좌측 연단과 우측 기자석 배열이 스케치 및 요구 구도와 일치함.",
      "entities": "인물 외형이 레퍼런스와 일치하고, 요구된 텍스트('검찰', '보도자료')와 기자들이 모두 존재함.",
      "hard_violations": [],
      "physics": "단상 뒤에서 문서를 들고 서 있으며, 기자들은 의자에 정상적으로 착석함."
     },
     {
      "label": "B",
      "direction": "인물이 정면 카메라 렌즈를 응시함.",
      "built_space": "연단을 사용하지 않고 통로 중앙에 서 있음.",
      "entities": "기자들이 전혀 없으며, 지정된 텍스트 표기도 누락됨.",
      "hard_violations": [
       "인물이 단상이 아닌 통로에 배치됨",
       "기자석 및 기자 누락",
       "지정된 구도 및 레이아웃 스케치 완전 무시"
      ],
      "physics": "두 발로 바닥에 서서 문서를 쥐고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "레이아웃 스케치의 구도, 인물 배치, 카메라 각도를 완벽하게 구현했으며 지정된 텍스트와 디테일도 훌륭하게 반영된 압도적인 결과물입니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "요구된 3/4 측면 구도와 레이아웃 스케치를 완전히 무시하고 정면 중앙에 인물을 배치했으며, 필수 인물인 기자들을 모두 누락한 심각한 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "차장검사는 손에 든 보도자료를 내려다보고 있으며, 기자들은 각자의 노트북 화면이나 단상을 향해 시선을 두고 있습니다.",
      "built_space": "지시된 레이아웃에 맞춰 좌측에 단상과 연단이, 우측에 기자석이 올바르게 배치되어 있으며, 노트북과 카메라들이 적절한 방향을 향하고 있습니다.",
      "entities": "차장검사의 외모가 레퍼런스와 정확히 일치하며, 기자들이 명시된 위치에 존재합니다. 단상의 '검찰' 및 문서의 '보도자료' 텍스트가 정확히 구현되었습니다.",
      "hard_violations": [],
      "physics": "검사는 연단 뒤에서 두 손으로 문서를 안정적으로 쥐고 있으며, 기자들은 의자에 앉아 손을 키보드 위에 올린 자연스러운 자세를 유지하고 있습니다."
     },
     {
      "label": "A",
      "direction": "차장검사는 프롬프트 지시와 다르게 정면 카메라 렌즈를 똑바로 응시하고 있습니다.",
      "built_space": "배경은 원본과 같으나 인물이 단상 뒤가 아닌 통로 중앙 바닥에 서 있어 공간 활용이 잘못되었습니다.",
      "entities": "차장검사의 외모는 레퍼런스와 유사하나, 필수적으로 포함되어야 할 기자들이 프레임 내에 전혀 존재하지 않습니다. 지정된 간판 텍스트도 없습니다.",
      "hard_violations": [
       "레이아웃 스케치의 인물 배치 및 구도(3/4 측면 앵글) 완전 무시",
       "필수 인물인 기자들 전원 누락",
       "인물이 단상이 아닌 통로 한가운데에 위치함"
      ],
      "physics": "바닥에 서서 문서를 쥐고 있는 기본 스탠딩 자세이며 물리적 오류는 없으나, 상황에 맞지 않는 자세입니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "레이아웃 스케치의 구도, 인물 배치, 카메라 각도를 완벽하게 구현했으며 지정된 텍스트와 디테일도 훌륭하게 반영된 압도적인 결과물입니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "요구된 3/4 측면 구도와 레이아웃 스케치를 완전히 무시하고 정면 중앙에 인물을 배치했으며, 필수 인물인 기자들을 모두 누락한 심각한 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "차장검사는 손에 든 보도자료를 내려다보고 있으며, 기자들은 각자의 노트북 화면이나 단상을 향해 시선을 두고 있습니다.",
      "built_space": "지시된 레이아웃에 맞춰 좌측에 단상과 연단이, 우측에 기자석이 올바르게 배치되어 있으며, 노트북과 카메라들이 적절한 방향을 향하고 있습니다.",
      "entities": "차장검사의 외모가 레퍼런스와 정확히 일치하며, 기자들이 명시된 위치에 존재합니다. 단상의 '검찰' 및 문서의 '보도자료' 텍스트가 정확히 구현되었습니다.",
      "hard_violations": [],
      "physics": "검사는 연단 뒤에서 두 손으로 문서를 안정적으로 쥐고 있으며, 기자들은 의자에 앉아 손을 키보드 위에 올린 자연스러운 자세를 유지하고 있습니다."
     },
     {
      "label": "B",
      "direction": "차장검사는 프롬프트 지시와 다르게 정면 카메라 렌즈를 똑바로 응시하고 있습니다.",
      "built_space": "배경은 원본과 같으나 인물이 단상 뒤가 아닌 통로 중앙 바닥에 서 있어 공간 활용이 잘못되었습니다.",
      "entities": "차장검사의 외모는 레퍼런스와 유사하나, 필수적으로 포함되어야 할 기자들이 프레임 내에 전혀 존재하지 않습니다. 지정된 간판 텍스트도 없습니다.",
      "hard_violations": [
       "레이아웃 스케치의 인물 배치 및 구도(3/4 측면 앵글) 완전 무시",
       "필수 인물인 기자들 전원 누락",
       "인물이 단상이 아닌 통로 한가운데에 위치함"
      ],
      "physics": "바닥에 서서 문서를 쥐고 있는 기본 스탠딩 자세이며 물리적 오류는 없으나, 상황에 맞지 않는 자세입니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S69sh1__bgfirst_bg.png",
   "bg_asset_id": "70801f0b-27a0-4796-b9bb-8308d956248a",
   "bg_record_key": "S69sh1::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S69sh1::cine": {
  "applied": true,
  "fingerprint": "9111724997ec215de97a8438130d8867826b969f3c411e0c8265fb0ff6074718",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S69sh1_sel.png",
  "source_sha256": "3abd9ce2d97b686aabfedadc51a1b65153151579b52bbcc6cfaf016a53b0bc40",
  "file": "S69sh1_cine.png",
  "latency_ms": 10185
 },
 "S69sh4::signage": {
  "fp": "110ab594c09ee544",
  "inscriptions": [
   {
    "surface_native": "브리핑룸 단상 뒤 현수막 백드롭",
    "text_native": "서울중앙지방검찰청",
    "reason_ko": "장원섭이 바라보는 단상 뒤편에 위치한 공식 기관 명칭 표기로, 이 장소가 검찰청 브리핑룸임을 명확히 보여주기 위해 필요합니다."
   }
  ]
 },
 "S69sh4": {
  "input_fingerprint": "43e79094b12491a9",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 기자들 맨 뒤편에 선 채 팔짱을 끼고 단상 쪽을 응시하는 장원섭의 전경\n\nLOCATION (lock): At the rear of the prosecution briefing room, behind the seated reporters and facing the podium. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated-head height beside the final reporter row, the pan completes only after 장원섭's full figure clears the foreground heads and chairs, holding him on the right in a subtle low-angle three-quarter view toward the podium. He keeps his arms folded and his gaze fixed past the camera's left side at the briefing, while the reporters remain lower and nearer in staggered seated postures.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-right of the frame, midground, looks toward briefing podium; last reporter row in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 마지막 기자석 (Reporters seated at varied heights and typing with nonuniform shoulder, head, and hand positions); used as Forms a low foreground layer from which 장원섭 is cleanly separated; 브리핑 단상 (Briefing in progress) — Seen along 장원섭's viewing axis at the far side of the room; used as Establishes the direction of 장원섭's restrained observation across the room; 열린 노트북들 (Open and being used) — Screens face the reporters, presenting mostly side or rear profiles to the camera; used as Adds repeated planes among the seated reporters without obscuring 장원섭's full figure.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime room light is intermittently disturbed by distant camera flashes, with 장원섭 held in restrained neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same briefing room, podium, press seating, camera flashes, and daylight. Exclude the chief prosecutor as the central subject and frame the other prosecutor standing behind the reporters.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 브리핑룸 단상 뒤 현수막 백드롭: \"서울중앙지방검찰청\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 기자들 맨 뒤편에 선 채 팔짱을 끼고 단상 쪽을 응시하는 장원섭의 전경\n\nLOCATION (lock): At the rear of the prosecution briefing room, behind the seated reporters and facing the podium. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated-head height beside the final reporter row, the pan completes only after 장원섭's full figure clears the foreground heads and chairs, holding him on the right in a subtle low-angle three-quarter view toward the podium. He keeps his arms folded and his gaze fixed past the camera's left side at the briefing, while the reporters remain lower and nearer in staggered seated postures.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-right of the frame, midground, looks toward briefing podium; last reporter row in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 마지막 기자석 (Reporters seated at varied heights and typing with nonuniform shoulder, head, and hand positions); used as Forms a low foreground layer from which 장원섭 is cleanly separated; 브리핑 단상 (Briefing in progress) — Seen along 장원섭's viewing axis at the far side of the room; used as Establishes the direction of 장원섭's restrained observation across the room; 열린 노트북들 (Open and being used) — Screens face the reporters, presenting mostly side or rear profiles to the camera; used as Adds repeated planes among the seated reporters without obscuring 장원섭's full figure.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime room light is intermittently disturbed by distant camera flashes, with 장원섭 held in restrained neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same briefing room, podium, press seating, camera flashes, and daylight. Exclude the chief prosecutor as the central subject and frame the other prosecutor standing behind the reporters.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 브리핑룸 단상 뒤 현수막 백드롭: \"서울중앙지방검찰청\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 기자들 맨 뒤편에 선 채 팔짱을 끼고 단상 쪽을 응시하는 장원섭의 전경\n\nLOCATION (lock): At the rear of the prosecution briefing room, behind the seated reporters and facing the podium. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated-head height beside the final reporter row, the pan completes only after 장원섭's full figure clears the foreground heads and chairs, holding him on the right in a subtle low-angle three-quarter view toward the podium. He keeps his arms folded and his gaze fixed past the camera's left side at the briefing, while the reporters remain lower and nearer in staggered seated postures.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 장원섭 in the middle-right of the frame, midground, looks toward briefing podium; last reporter row in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 마지막 기자석 (Reporters seated at varied heights and typing with nonuniform shoulder, head, and hand positions); used as Forms a low foreground layer from which 장원섭 is cleanly separated; 브리핑 단상 (Briefing in progress) — Seen along 장원섭's viewing axis at the far side of the room; used as Establishes the direction of 장원섭's restrained observation across the room; 열린 노트북들 (Open and being used) — Screens face the reporters, presenting mostly side or rear profiles to the camera; used as Adds repeated planes among the seated reporters without obscuring 장원섭's full figure.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime room light is intermittently disturbed by distant camera flashes, with 장원섭 held in restrained neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same briefing room, podium, press seating, camera flashes, and daylight. Exclude the chief prosecutor as the central subject and frame the other prosecutor standing behind the reporters.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 브리핑룸 단상 뒤 현수막 백드롭: \"서울중앙지방검찰청\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.75
   },
   "adjusted": {
    "A": 0.667,
    "B": 0.5
   },
   "violations": {
    "A": [
     "[gemini-pro] 공간 구조 위반: 먼 배경에 있어야 할 단상과 백드롭이 피사체 바로 뒤에 있음",
     "[gemini-pro] 인물 부재: 브리핑이 진행 중(Briefing in progress)이어야 하나 단상이 비어 있음",
     "[openrouter:x-ai/grok-4.6] PEOPLE 목록 밖 착석 기자 다수 추가(발명된 인물)",
     "[openrouter:x-ai/grok-4.6] 장원섭이 기자 맨 뒤가 아니라 단상 옆에 서서 단상을 응시하지 않음"
    ],
    "B": [
     "[gemini-pro] 공간 구조 위반: 단상과 백드롭이 피사체 바로 뒤에 잘못 배치됨",
     "[gemini-pro] 인물 부재: 브리핑이 진행 중이어야 하나 단상에 아무도 없음",
     "[gemini-pro] 물리적 불가능: 팔짱 낀 손과 시계 부위의 해부학적 왜곡",
     "[openrouter:x-ai/grok-4.6] PEOPLE 목록 밖 착석 기자 다수 추가(발명된 인물)",
     "[openrouter:x-ai/grok-4.6] 장원섭이 기자 맨 뒤가 아니라 단상 옆에 서서 단상을 응시하지 않음"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 667,
   "B": 500
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 667,
    "verdict_ko": "텍스트는 정확하나, 단상과 백드롭이 멀리 있지 않고 피사체 바로 뒤에 배치되어 공간 구조 구현에 완전히 실패함.  ★위반: [gemini-pro] 공간 구조 위반: 먼 배경에 있어야 할 단상과 백드롭이 피사체 바로 뒤에 있음 / [gemini-pro] 인물 부재: 브리핑이 진행 중(Briefing in progress)이어야 하나 단상이 비어 있음 / [openrouter:x-ai/grok-4.6] PEOPLE 목록 밖 착석 기자 다수 추가(발명된 인물) / [openrouter:x-ai/grok-4.6] 장원섭이 기자 맨 뒤가 아니라 단상 옆에 서서 단상을 응시하지 않음"
   },
   {
    "label": "B",
    "score": 500,
    "verdict_ko": "공간 구조가 심각하게 왜곡되었으며, 팔짱 낀 손과 시계 주변의 해부학적 오류까지 겹쳐 부자연스러움.  ★위반: [gemini-pro] 공간 구조 위반: 단상과 백드롭이 피사체 바로 뒤에 잘못 배치됨 / [gemini-pro] 인물 부재: 브리핑이 진행 중이어야 하나 단상에 아무도 없음 / [gemini-pro] 물리적 불가능: 팔짱 낀 손과 시계 부위의 해부학적 왜곡 / [openrouter:x-ai/grok-4.6] PEOPLE 목록 밖 착석 기자 다수 추가(발명된 인물) / [openrouter:x-ai/grok-4.6] 장원섭이 기자 맨 뒤가 아니라 단상 옆에 서서 단상을 응시하지 않음"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S69sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "장원섭이 지시된 '기자들 맨 뒤편'이 아닌 화면 우측의 브리핑 단상 바로 옆에 서서 기자들을 마주보고 있음.",
     "fix_en": "Move Jang Won-seop to the left background behind the seated reporters, filling his current space with the continuation of the blue backdrop and wooden wall. Preserve the existing reporters, podium, lighting, and overall room architecture.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이전 샷(레퍼런스) 스틸과 비교하여 단상과 백드롭, 기자석의 좌우 위치 및 방향이 완전히 반대로 뒤집혀 있음.",
     "fix_en": "Mirror the layout of the podium, background, and reporter seating. Keep the lighting, materials, and character identities unchanged.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경에 세워진 방송용 카메라들의 렌즈가 브리핑 단상이 아닌 화면 왼쪽의 엉뚱한 방향을 향하고 있음.",
     "fix_en": "Rotate the broadcast camera heads so their lenses point right toward the podium. Preserve the tripods, attached lights, and surrounding background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "장원섭이 카메라에 거의 정면으로 양발 균등히 서서 초상처럼 보인다",
     "fix_en": "Adjust Jang Won-seop's lower body to shift his weight onto one leg, breaking the symmetric pose. Preserve his face, suit, folded arms, and the background behind him.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "장원섭이 지시된 '기자들 맨 뒤편'이 아닌 화면 우측의 브리핑 단상 바로 옆에 서서 기자들을 마주보고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 샷(레퍼런스) 스틸과 비교하여 단상과 백드롭, 기자석의 좌우 위치 및 방향이 완전히 반대로 뒤집혀 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "배경에 세워진 방송용 카메라들의 렌즈가 브리핑 단상이 아닌 화면 왼쪽의 엉뚱한 방향을 향하고 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "중경 오른쪽 장원섭이 기자석 맨 뒤가 아니라 단상 바로 옆에 서 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "단상이 시선 축의 먼 쪽이 아니라 장원섭 옆에 붙어 있어 단상 응시가 성립하지 않는다",
     "severity": "critical"
    },
    {
     "issue_ko": "장원섭이 단상이 아니라 프레임 왼쪽 기자·카메라 쪽을 바라본다",
     "severity": "major"
    },
    {
     "issue_ko": "장원섭이 카메라에 거의 정면으로 양발 균등히 서서 초상처럼 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "전경 기자들이 단상이 아닌 반대쪽을 향해 앉아 있다",
     "severity": "major"
    },
    {
     "issue_ko": "이전 스틸의 단상·카메라·기자석 고정 배치가 바뀌어 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 6
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 샷(레퍼런스) 스틸과 비교하여 단상과 백드롭, 기자석의 좌우 위치 및 방향이 완전히 반대로 뒤집혀 있음.",
    "fix_en": "Mirror the layout of the podium, background, and reporter seating. Keep the lighting, materials, and character identities unchanged.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경에 세워진 방송용 카메라들의 렌즈가 브리핑 단상이 아닌 화면 왼쪽의 엉뚱한 방향을 향하고 있음.",
    "fix_en": "Rotate the broadcast camera heads so their lenses point right toward the podium. Preserve the tripods, attached lights, and surrounding background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "장원섭이 카메라에 거의 정면으로 양발 균등히 서서 초상처럼 보인다",
    "fix_en": "Adjust Jang Won-seop's lower body to shift his weight onto one leg, breaking the symmetric pose. Preserve his face, suit, folded arms, and the background behind him.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Move Jang Won-seop to the left background behind the seated reporters, filling his current space with the continuation of the blue backdrop and wooden wall. Preserve the existing reporters, podium, lighting, and overall room architecture.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "장원섭의 외형과 현수막 텍스트는 훌륭하게 렌더링되었으나, 단상이 인물의 시선 방향과 반대편에 위치해 시선이 목표인 단상에 닿지 않는 치명적인 방향(Direction) 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "화면에 동일한 주인공(장원섭)이 두 명으로 복제되어 나타나는 심각한 하드 위반(Hard Violation)이 발생했으며, 시선 방향 또한 목표인 단상을 향하지 않습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭은 팔짱을 낀 채 화면 왼쪽을 응시하고 있으나, 시선의 목표물이 되어야 할 단상은 그의 우측(등 뒤쪽)에 위치하고 있어 시선이 목표에 닿지 않음.",
      "built_space": "이전 샷의 구조가 좌우 반전된 형태로 렌더링됨. 좌측 하단에 노트북을 사용하는 기자석이 있고, 우측에 단상이 배치되었으며 좌측 창가를 따라 여러 대의 방송용 카메라가 서 있음.",
      "entities": "장원섭의 얼굴 특징, 짧은 머리, 짙은 회색 정장과 넥타이가 레퍼런스와 정확히 일치함. 배경 현수막에 '서울중앙지방검찰청', 단상에 '검 찰'이라는 텍스트가 명확하게 적혀 있음.",
      "hard_violations": [],
      "physics": "모든 인물의 발이 바닥에 닿아 있거나 의자에 앉아 무게를 지지하고 있으며, 카메라 삼각대 등 사물들도 바닥에 안정적으로 세워져 있음. 공중에 떠 있거나 지지되지 않은 사물은 없음."
     },
     {
      "label": "B",
      "direction": "화면에 등장하는 두 명의 장원섭 모두 화면 왼쪽을 바라보고 있으며, 시선의 목표물인 단상은 그들의 우측에 있어 어느 쪽도 단상을 향해 시선을 두지 못함.",
      "built_space": "A와 동일하게 이전 샷 대비 좌우가 반전된 구조로 기자석이 좌측, 단상이 우측, 카메라가 좌측 창가 쪽에 배치되어 있음.",
      "entities": "지정된 인물 장원섭과 외형 및 복장이 완전히 동일한 인물이 좌측 후경과 우측 전경에 각각 1명씩 총 2명 등장함. 배경의 '서울중앙지방검찰청' 텍스트는 정확하게 구현됨.",
      "hard_violations": [
       "duplicated or extra bodies (주인공 장원섭이 두 명으로 복제되어 화면에 등장함)"
      ],
      "physics": "인물들은 의자에 앉아 있거나 바닥에 두 발로 서서 체중을 지지하고 있으며, 노트북과 카메라 등의 집기류도 책상과 바닥에 올바르게 지지되어 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "장원섭의 외형과 현수막 텍스트는 훌륭하게 렌더링되었으나, 단상이 인물의 시선 방향과 반대편에 위치해 시선이 목표인 단상에 닿지 않는 치명적인 방향(Direction) 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "화면에 동일한 주인공(장원섭)이 두 명으로 복제되어 나타나는 심각한 하드 위반(Hard Violation)이 발생했으며, 시선 방향 또한 목표인 단상을 향하지 않습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭은 팔짱을 낀 채 화면 왼쪽을 응시하고 있으나, 시선의 목표물이 되어야 할 단상은 그의 우측(등 뒤쪽)에 위치하고 있어 시선이 목표에 닿지 않음.",
      "built_space": "이전 샷의 구조가 좌우 반전된 형태로 렌더링됨. 좌측 하단에 노트북을 사용하는 기자석이 있고, 우측에 단상이 배치되었으며 좌측 창가를 따라 여러 대의 방송용 카메라가 서 있음.",
      "entities": "장원섭의 얼굴 특징, 짧은 머리, 짙은 회색 정장과 넥타이가 레퍼런스와 정확히 일치함. 배경 현수막에 '서울중앙지방검찰청', 단상에 '검 찰'이라는 텍스트가 명확하게 적혀 있음.",
      "hard_violations": [],
      "physics": "모든 인물의 발이 바닥에 닿아 있거나 의자에 앉아 무게를 지지하고 있으며, 카메라 삼각대 등 사물들도 바닥에 안정적으로 세워져 있음. 공중에 떠 있거나 지지되지 않은 사물은 없음."
     },
     {
      "label": "B",
      "direction": "화면에 등장하는 두 명의 장원섭 모두 화면 왼쪽을 바라보고 있으며, 시선의 목표물인 단상은 그들의 우측에 있어 어느 쪽도 단상을 향해 시선을 두지 못함.",
      "built_space": "A와 동일하게 이전 샷 대비 좌우가 반전된 구조로 기자석이 좌측, 단상이 우측, 카메라가 좌측 창가 쪽에 배치되어 있음.",
      "entities": "지정된 인물 장원섭과 외형 및 복장이 완전히 동일한 인물이 좌측 후경과 우측 전경에 각각 1명씩 총 2명 등장함. 배경의 '서울중앙지방검찰청' 텍스트는 정확하게 구현됨.",
      "hard_violations": [
       "duplicated or extra bodies (주인공 장원섭이 두 명으로 복제되어 화면에 등장함)"
      ],
      "physics": "인물들은 의자에 앉아 있거나 바닥에 두 발로 서서 체중을 지지하고 있으며, 노트북과 카메라 등의 집기류도 책상과 바닥에 올바르게 지지되어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "단상과 백드롭이 방의 뒤편인 주인공 바로 옆에 배치되는 공간 왜곡이 있으나, A의 치명적인 인물 복제 오류가 없고 주인공의 얼굴, 자세, 지정된 텍스트를 모두 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "화면 좌측 배경과 우측 전경에 동일한 외모를 가진 장원섭이 두 명으로 복제되어 나타나는 치명적인 규정 위반이 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "우측 전경의 장원섭과 좌측 배경의 복제된 인물 모두 화면 왼쪽 바깥을 응시함. 기자들은 자신들의 노트북 화면을 향해 시선을 고정함.",
      "built_space": "기준 사진의 공간 배치가 완전히 무너져, 기자석과 카메라가 좌측에 위치하고 브리핑 단상 및 백드롭이 주인공과 함께 화면 우측에 잘못 배치됨.",
      "entities": "프롬프트에 유일하게 명시된 주인공 장원섭과 동일한 외형, 복장을 한 인물이 두 명으로 등장함(복제). 백드롭의 '서울중앙지방검찰청' 및 단상의 '검찰' 텍스트는 잘 구현됨.",
      "hard_violations": [
       "duplicated or extra bodies (주인공 장원섭이 화면 좌우에 두 명으로 복제됨)"
      ],
      "physics": "전경 및 배경 인물들 모두 두 발을 바닥에 딛고 있으며, 앉은 기자들의 자세도 의자에 의해 잘 지지됨. 허공에 뜬 객체 없음."
     },
     {
      "label": "B",
      "direction": "장원섭은 팔짱을 낀 채 화면 왼쪽 바깥을 응시하고 있음. 앉아 있는 기자들의 시선은 각자의 노트북 화면을 향함.",
      "built_space": "프롬프트에 따라 장원섭이 기자들 뒤편에 있으나, 방 앞쪽에 있어야 할 단상과 백드롭이 주인공 바로 뒤(화면 우측)로 옮겨와 공간적 구조와 깊이감이 기준 사진과 어긋남.",
      "entities": "장원섭이 단일 인물로 나타나며, 레퍼런스의 얼굴 및 정장과 매우 잘 일치함. 백드롭 텍스트 '서울중앙지방검찰청'과 단상의 '검찰' 글자 모양과 배치가 정확함.",
      "hard_violations": [],
      "physics": "장원섭은 바닥에 안정적으로 서서 체중을 지탱하고 있으며, 기자들의 손은 노트북 위에 자연스럽게 얹혀 있음. 물리적으로 불가능한 지지나 떠 있는 객체는 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "단상과 백드롭이 방의 뒤편인 주인공 바로 옆에 배치되는 공간 왜곡이 있으나, A의 치명적인 인물 복제 오류가 없고 주인공의 얼굴, 자세, 지정된 텍스트를 모두 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "화면 좌측 배경과 우측 전경에 동일한 외모를 가진 장원섭이 두 명으로 복제되어 나타나는 치명적인 규정 위반이 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "우측 전경의 장원섭과 좌측 배경의 복제된 인물 모두 화면 왼쪽 바깥을 응시함. 기자들은 자신들의 노트북 화면을 향해 시선을 고정함.",
      "built_space": "기준 사진의 공간 배치가 완전히 무너져, 기자석과 카메라가 좌측에 위치하고 브리핑 단상 및 백드롭이 주인공과 함께 화면 우측에 잘못 배치됨.",
      "entities": "프롬프트에 유일하게 명시된 주인공 장원섭과 동일한 외형, 복장을 한 인물이 두 명으로 등장함(복제). 백드롭의 '서울중앙지방검찰청' 및 단상의 '검찰' 텍스트는 잘 구현됨.",
      "hard_violations": [
       "duplicated or extra bodies (주인공 장원섭이 화면 좌우에 두 명으로 복제됨)"
      ],
      "physics": "전경 및 배경 인물들 모두 두 발을 바닥에 딛고 있으며, 앉은 기자들의 자세도 의자에 의해 잘 지지됨. 허공에 뜬 객체 없음."
     },
     {
      "label": "A",
      "direction": "장원섭은 팔짱을 낀 채 화면 왼쪽 바깥을 응시하고 있음. 앉아 있는 기자들의 시선은 각자의 노트북 화면을 향함.",
      "built_space": "프롬프트에 따라 장원섭이 기자들 뒤편에 있으나, 방 앞쪽에 있어야 할 단상과 백드롭이 주인공 바로 뒤(화면 우측)로 옮겨와 공간적 구조와 깊이감이 기준 사진과 어긋남.",
      "entities": "장원섭이 단일 인물로 나타나며, 레퍼런스의 얼굴 및 정장과 매우 잘 일치함. 백드롭 텍스트 '서울중앙지방검찰청'과 단상의 '검찰' 글자 모양과 배치가 정확함.",
      "hard_violations": [],
      "physics": "장원섭은 바닥에 안정적으로 서서 체중을 지탱하고 있으며, 기자들의 손은 노트북 위에 자연스럽게 얹혀 있음. 물리적으로 불가능한 지지나 떠 있는 객체는 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 12,
     "B": 3
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S69sh1"
  }
 },
 "S69sh4::cine": {
  "applied": true,
  "fingerprint": "54a9da751d78af212786d3a837a3e44da755e73be455cf7e3047f40149be6acc",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S69sh4_sel.png",
  "source_sha256": "a2767090900ba7a5d81c89caf1fd642e35c9dbf42d45fbd8c8f148d5c2f141bc",
  "file": "S69sh4_cine.png",
  "latency_ms": 10638
 },
 "S70sh1::signage": {
  "fp": "dfcfcddb84ce8741",
  "inscriptions": [
   {
    "surface_native": "낡은 간판",
    "text_native": "자매식당",
    "reason_ko": "지문에 명시된 어두운 골목 안 오래된 식당의 정체성을 나타내기 위해 간판 글자가 필요합니다."
   },
   {
    "surface_native": "네온사인 간판",
    "text_native": "소주·맥주",
    "reason_ko": "한국의 전형적인 먹자골목 밤 풍경을 사실적으로 묘사하기 위해 주변 네온사인 문구가 필요합니다."
   }
  ]
 },
 "S70sh1": {
  "input_fingerprint": "fcf82a42166f77be",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 먹자골목 안, 환한 네온사인들 사이로 낡은 '자매식당' 간판이 빛나는 전경\n\nLOCATION (lock): Outside the old restaurant facade in a lively nighttime dining alley, among brighter neighboring signs. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From above pedestrian head height at the food-alley entrance, the opening portion of the dolly-in looks down the alley in a wide composition, layering signs along both sides while keeping the aged 자매식당 sign visible deeper near the center. The movement follows the same central axis, allowing the surrounding brighter signs to frame rather than eclipse the restaurant sign.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 자매식당 sign in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 자매식당 간판 (Lit and showing signs of age) — Its name-bearing face is visible to the camera down the alley; used as The principal destination of the forward move, held within the brighter field of neighboring signage; 주변 술집의 세련된 간판들 (Brightly lit) — Their illuminated content faces are visible at alternating oblique angles along both sides of the alley; used as Builds layered side framing and visual contrast around the older restaurant sign; 먹자골목 (Nighttime alley with lively sound from surrounding bars) — Viewed lengthwise from its entrance, with both sides receding toward the restaurant sign; used as Defines the long central corridor through which the camera advances.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright neon signage provides the explicitly present night illumination, balanced into a restrained cool-neutral palette without sensational saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 낡은 간판: \"자매식당\"\n- 네온사인 간판: \"소주·맥주\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 먹자골목 안, 환한 네온사인들 사이로 낡은 '자매식당' 간판이 빛나는 전경\n\nLOCATION (lock): Outside the old restaurant facade in a lively nighttime dining alley, among brighter neighboring signs. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From above pedestrian head height at the food-alley entrance, the opening portion of the dolly-in looks down the alley in a wide composition, layering signs along both sides while keeping the aged 자매식당 sign visible deeper near the center. The movement follows the same central axis, allowing the surrounding brighter signs to frame rather than eclipse the restaurant sign.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 자매식당 sign in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 자매식당 간판 (Lit and showing signs of age) — Its name-bearing face is visible to the camera down the alley; used as The principal destination of the forward move, held within the brighter field of neighboring signage; 주변 술집의 세련된 간판들 (Brightly lit) — Their illuminated content faces are visible at alternating oblique angles along both sides of the alley; used as Builds layered side framing and visual contrast around the older restaurant sign; 먹자골목 (Nighttime alley with lively sound from surrounding bars) — Viewed lengthwise from its entrance, with both sides receding toward the restaurant sign; used as Defines the long central corridor through which the camera advances.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright neon signage provides the explicitly present night illumination, balanced into a restrained cool-neutral palette without sensational saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 낡은 간판: \"자매식당\"\n- 네온사인 간판: \"소주·맥주\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 어두운 먹자골목 안, 환한 네온사인들 사이로 낡은 '자매식당' 간판이 빛나는 전경\n\nLOCATION (lock): Outside the old restaurant facade in a lively nighttime dining alley, among brighter neighboring signs. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From above pedestrian head height at the food-alley entrance, the opening portion of the dolly-in looks down the alley in a wide composition, layering signs along both sides while keeping the aged 자매식당 sign visible deeper near the center. The movement follows the same central axis, allowing the surrounding brighter signs to frame rather than eclipse the restaurant sign.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 자매식당 sign in the middle-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 자매식당 간판 (Lit and showing signs of age) — Its name-bearing face is visible to the camera down the alley; used as The principal destination of the forward move, held within the brighter field of neighboring signage; 주변 술집의 세련된 간판들 (Brightly lit) — Their illuminated content faces are visible at alternating oblique angles along both sides of the alley; used as Builds layered side framing and visual contrast around the older restaurant sign; 먹자골목 (Nighttime alley with lively sound from surrounding bars) — Viewed lengthwise from its entrance, with both sides receding toward the restaurant sign; used as Defines the long central corridor through which the camera advances.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Bright neon signage provides the explicitly present night illumination, balanced into a restrained cool-neutral palette without sensational saturation.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 낡은 간판: \"자매식당\"\n- 네온사인 간판: \"소주·맥주\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.571,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.375
   },
   "adjusted": {
    "A": 1.429,
    "B": 1.125
   },
   "violations": {
    "B": [
     "[gemini-pro] 고정된 로케이션 구조 위반 (도로 형태 및 건물 외형 임의 변형)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1429,
   "B": 1125
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1429,
    "verdict_ko": "지정된 로케이션의 구조를 정확히 유지하면서 야간 골목의 분위기와 '자매식당', '소주·맥주' 텍스트를 충실하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1125,
    "verdict_ko": "지정된 텍스트는 잘 반영했으나, 고정 조건인 건물의 외형과 Y자 도로 구조를 대칭형 좁은 골목으로 임의 변형하여 감점되었습니다.  ★위반: [gemini-pro] 고정된 로케이션 구조 위반 (도로 형태 및 건물 외형 임의 변형)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_old_commercial_alley_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우측 '소주·맥주' 간판 옆 초록색 네온사인, 골목 안쪽 붉은색 세로 간판 및 우측 가장자리 초록색 간판 등의 텍스트가 식별할 수 없는 문자로 왜곡되어 있습니다.",
     "fix_en": "Replace the unreadable garbled text on the background green and red signs with indistinct bright neon blooms or simple abstract strokes that stage the text out of legibility. Preserve the legible '자매식당' and '소주·맥주' signs, the building structures, the camera angle, and the lighting.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "자매식당 간판이 화면 정중앙 배경 깊은 곳이 아니라 왼쪽 전경 건물 정면에 크게 붙어 있다",
     "fix_en": "Remove the '자매식당' text from the large left-hand building and place it on a smaller rectangular sign deeper down the alleyway on the right. Maintain the current camera framing, the building shapes, the '소주·맥주' neon sign, and the nighttime lighting.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "골목 입구에서 골목 안을 길이 방향으로 내려다보는 구도가 아니라 위치 참조 사진과 같은 모퉁이 전경이다",
     "fix_en": "Darken the large building on the left and brighten the neon signs deep down the right alley to forcefully redirect visual focus into the corridor. Preserve the exact camera angle, scene geometry, and the existing signage on the left.",
     "severity": "critical",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우측 '소주·맥주' 간판 옆 초록색 네온사인, 골목 안쪽 붉은색 세로 간판 및 우측 가장자리 초록색 간판 등의 텍스트가 식별할 수 없는 문자로 왜곡되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "'자매식당' 간판을 화면 중앙 배경에 배치하라는 지시와 레퍼런스 이미지의 구도를 복사하지 말라는 지시를 어기고, 레퍼런스와 완전히 동일한 구도(간판이 중경 좌측에 위치)로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "자매식당 간판이 화면 정중앙 배경 깊은 곳이 아니라 왼쪽 전경 건물 정면에 크게 붙어 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "골목 입구에서 골목 안을 길이 방향으로 내려다보는 구도가 아니라 위치 참조 사진과 같은 모퉁이 전경이다",
     "severity": "critical"
    },
    {
     "issue_ko": "양옆 네온 간판이 골목을 따라 비스듬히 겹쳐 식당 간판을 감싸는 층위가 없다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "우측 '소주·맥주' 간판 옆 초록색 네온사인, 골목 안쪽 붉은색 세로 간판 및 우측 가장자리 초록색 간판 등의 텍스트가 식별할 수 없는 문자로 왜곡되어 있습니다.",
    "fix_en": "Replace the unreadable garbled text on the background green and red signs with indistinct bright neon blooms or simple abstract strokes that stage the text out of legibility. Preserve the legible '자매식당' and '소주·맥주' signs, the building structures, the camera angle, and the lighting.",
    "severity": "major",
    "observation_index": 0
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the '자매식당' text from the large left-hand building and place it on a smaller rectangular sign deeper down the alleyway on the right. Maintain the current camera framing, the building shapes, the '소주·맥주' neon sign, and the nighttime lighting.\n- Darken the large building on the left and brighten the neon signs deep down the right alley to forcefully redirect visual focus into the corridor. Preserve the exact camera angle, scene geometry, and the existing signage on the left.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "요구된 두 간판의 텍스트를 모두 정확하게 렌더링했으며, '낡은 간판'이라는 지시에 맞게 레퍼런스의 메인 간판을 자연스럽게 활용하여 공간의 현실감을 잘 살렸습니다."
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "'자매식당'을 원경에 배치해 프레이밍 의도는 일부 맞췄으나, 전경의 거대한 메인 간판을 부자연스럽게 백지로 비워두어 장소의 구조적 현실감을 크게 훼손했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.571
     },
     "adjusted": {
      "A": 1.5,
      "B": 1.571
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "요구된 두 간판의 텍스트를 모두 정확하게 렌더링했으며, '낡은 간판'이라는 지시에 맞게 레퍼런스의 메인 간판을 자연스럽게 활용하여 공간의 현실감을 잘 살렸습니다."
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "'자매식당'을 원경에 배치해 프레이밍 의도는 일부 맞췄으나, 전경의 거대한 메인 간판을 부자연스럽게 백지로 비워두어 장소의 구조적 현실감을 크게 훼손했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "프롬프트가 요구한 '자매식당' 텍스트를 가장 잘 보이는 낡은 간판 위치에 정확히 배치하여 지시사항을 훌륭하게 이행했습니다."
     },
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "네온사인 등 전반적인 분위기는 구현했으나, 프레임에서 가장 눈에 띄는 낡은 간판을 빈 공간으로 남겨두어 주요 지시사항을 누락했습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.5
     },
     "adjusted": {
      "A": 1.571,
      "B": 1.5
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "프롬프트가 요구한 '자매식당' 텍스트를 가장 잘 보이는 낡은 간판 위치에 정확히 배치하여 지시사항을 훌륭하게 이행했습니다."
     },
     {
      "label": "B",
      "score": 1571,
      "verdict_ko": "네온사인 등 전반적인 분위기는 구현했으나, 프레임에서 가장 눈에 띄는 낡은 간판을 빈 공간으로 남겨두어 주요 지시사항을 누락했습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 3000,
     "B": 3142
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S70sh1::cine": {
  "applied": true,
  "fingerprint": "6067829e3648f908da9ecff1a78437f3c5a7aa5159b8427a371fa2bba2ea912c",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S70sh1_sel.png",
  "source_sha256": "f947bf5fe3281105762b750af7a62c40e64df76690223e28726dc5a0b34b26f9",
  "file": "S70sh1_cine.png",
  "latency_ms": 10308
 },
 "S71sh4::signage": {
  "fp": "a1084ef51a35873d",
  "inscriptions": [
   {
    "surface_native": "식당 벽면 메뉴판",
    "text_native": "닭볶음탕",
    "reason_ko": "주인공이 먹고 있는 닭볶음탕이 식당 메뉴판에 표기되어 한국 식당 내부의 사실적인 분위기를 살립니다."
   }
  ]
 },
 "S71sh4": {
  "input_fingerprint": "1c2e94b96b04973a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닭볶음탕이 끓는 냄비 앞 테이블에 앉아 소주잔을 입가에 댄 전택수의 상체\n\nLOCATION (lock): At a dining table inside the otherwise empty restaurant, beside the simmering pot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above tabletop height beside 심옥's end of the table, the opening dolly-in holds the boiling pot in the lower foreground at a natural scale and 전택수 beyond it in a medium upper-body frame. 전택수 leans slightly over the table with the soju glass at his lips and his eyes lowered, while the surrounding empty restaurant remains peripheral rather than becoming a second focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 닭볶음탕 냄비 (Boiling on the table); used as Occupies less than the lower third between the lens and 전택수, establishing the meal without obscuring his hands or face; 식탁 (Set for the meal) — Its near edge runs diagonally from the lower frame toward 전택수; used as Provides the diagonal line along which the subsequent camera movement will continue; 소주잔 (Raised to his lips); used as Held at 전택수's mouth as the immediate action detail; 식당 내부 (Empty apart from 전택수 and 심옥); used as Keeps the encounter private and visually isolated.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient restaurant light appropriate to the nighttime interior is rendered naturalistically in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The pot of dak-bokkeum-tang continues simmering on the table, with Taeksu's soju glass in front of him.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면 메뉴판: \"닭볶음탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닭볶음탕이 끓는 냄비 앞 테이블에 앉아 소주잔을 입가에 댄 전택수의 상체\n\nLOCATION (lock): At a dining table inside the otherwise empty restaurant, beside the simmering pot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above tabletop height beside 심옥's end of the table, the opening dolly-in holds the boiling pot in the lower foreground at a natural scale and 전택수 beyond it in a medium upper-body frame. 전택수 leans slightly over the table with the soju glass at his lips and his eyes lowered, while the surrounding empty restaurant remains peripheral rather than becoming a second focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 닭볶음탕 냄비 (Boiling on the table); used as Occupies less than the lower third between the lens and 전택수, establishing the meal without obscuring his hands or face; 식탁 (Set for the meal) — Its near edge runs diagonally from the lower frame toward 전택수; used as Provides the diagonal line along which the subsequent camera movement will continue; 소주잔 (Raised to his lips); used as Held at 전택수's mouth as the immediate action detail; 식당 내부 (Empty apart from 전택수 and 심옥); used as Keeps the encounter private and visually isolated.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient restaurant light appropriate to the nighttime interior is rendered naturalistically in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The pot of dak-bokkeum-tang continues simmering on the table, with Taeksu's soju glass in front of him.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면 메뉴판: \"닭볶음탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 닭볶음탕이 끓는 냄비 앞 테이블에 앉아 소주잔을 입가에 댄 전택수의 상체\n\nLOCATION (lock): At a dining table inside the otherwise empty restaurant, beside the simmering pot. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above tabletop height beside 심옥's end of the table, the opening dolly-in holds the boiling pot in the lower foreground at a natural scale and 전택수 beyond it in a medium upper-body frame. 전택수 leans slightly over the table with the soju glass at his lips and his eyes lowered, while the surrounding empty restaurant remains peripheral rather than becoming a second focal point.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 닭볶음탕 냄비 (Boiling on the table); used as Occupies less than the lower third between the lens and 전택수, establishing the meal without obscuring his hands or face; 식탁 (Set for the meal) — Its near edge runs diagonally from the lower frame toward 전택수; used as Provides the diagonal line along which the subsequent camera movement will continue; 소주잔 (Raised to his lips); used as Held at 전택수's mouth as the immediate action detail; 식당 내부 (Empty apart from 전택수 and 심옥); used as Keeps the encounter private and visually isolated.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Ambient restaurant light appropriate to the nighttime interior is rendered naturalistically in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The pot of dak-bokkeum-tang continues simmering on the table, with Taeksu's soju glass in front of him.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 식당 벽면 메뉴판: \"닭볶음탕\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선은 아래를 향하고 있으며, 소주잔의 끝이 입술에 닿아 있음.",
    "built_space": "식당 내부 테이블 앞에 앉아 있으며, 카메라와 인물 사이에 식탁과 냄비가 위치함.",
    "entities": "전택수(레퍼런스와 무관한 이전 컷의 인물 외모 및 복장), 끓는 닭볶음탕 냄비, 소주잔, 벽면 닭볶음탕 메뉴판.",
    "hard_violations": [
     "이전 컷 인물 금지 및 캐릭터 레퍼런스 준수 지시를 완전히 위반한 인물 등장"
    ],
    "physics": "오른손이 소주잔을 안정적으로 쥐고 허공에서 입가로 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "시선은 아래를 향하고 있으며, 소주잔의 끝이 입술에 닿아 있음.",
    "built_space": "식당 내부 테이블 앞에 앉아 있으며, 카메라와 인물 사이에 식탁과 냄비가 위치함.",
    "entities": "전택수(경찰복과 모자 등 캐릭터 레퍼런스와 일치), 끓는 닭볶음탕 냄비, 소주잔, 유리문에 붙은 닭볶음탕 표지.",
    "hard_violations": [],
    "physics": "오른손이 소주잔을 안정적으로 쥐고 허공에서 입가로 지탱하고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 8,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "제공된 캐릭터 레퍼런스를 정확히 반영하였으며, 지정된 프레이밍과 동작(소주잔을 입에 대고 시선을 내린 상체 샷)을 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "이전 컷에 등장한 인물을 배제하라는 지시를 어기고 해당 인물을 그대로 등장시켜 캐릭터 레퍼런스를 완전히 무시했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S2sh9_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이전 샷(Reference)에서 유리문에 고정되어 있던 '당기시오' 안내문이 사라지고, 그 자리에 프롬프트가 벽면 메뉴판용으로 지시한 '닭볶음탕' 텍스트가 잘못 배치되어 장소 일관성을 훼손합니다.",
     "fix_en": "Replace the '닭볶음탕' sign on the glass door with '당기시오' to match the reference, and add a '닭볶음탕' menu to the wall. Preserve the man, his pose, clothing, the boiling pot, table, and lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "소주잔을 쥐고 있는 인물의 오른손 손가락이 비정상적으로 묘사되었습니다 (잔 앞면을 감싼 손가락이 5개로 보임).",
     "fix_en": "Redraw the right hand to show exactly four fingers wrapping the front of the glass. Preserve the man, his face, clothing, the glass, pot, and background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "인물의 재킷 우측 가슴에 새겨진 영문 텍스트가 캐릭터 레퍼런스 이미지의 'SAUKS'가 아닌 'SALUS'로 잘못 렌더링되었습니다.",
     "fix_en": "Change the right chest text from 'SALUS' to 'SAUKS'. Preserve the man, his clothing, pose, and background.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이전 샷(Reference)에서 유리문에 고정되어 있던 '당기시오' 안내문이 사라지고, 그 자리에 프롬프트가 벽면 메뉴판용으로 지시한 '닭볶음탕' 텍스트가 잘못 배치되어 장소 일관성을 훼손합니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "소주잔을 쥐고 있는 인물의 오른손 손가락이 비정상적으로 묘사되었습니다 (잔 앞면을 감싼 손가락이 5개로 보임).",
     "severity": "critical"
    },
    {
     "issue_ko": "인물의 재킷 우측 가슴에 새겨진 영문 텍스트가 캐릭터 레퍼런스 이미지의 'SAUKS'가 아닌 'SALUS'로 잘못 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 출입문 유리에 이전 샷의 '당기시오' 대신 '닭볶음탕'이 붙어 있고 지정된 벽면 메뉴판은 없다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "인물의 재킷 우측 가슴에 새겨진 영문 텍스트가 캐릭터 레퍼런스 이미지의 'SAUKS'가 아닌 'SALUS'로 잘못 렌더링되었습니다.",
    "fix_en": "Change the right chest text from 'SALUS' to 'SAUKS'. Preserve the man, his clothing, pose, and background.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the '닭볶음탕' sign on the glass door with '당기시오' to match the reference, and add a '닭볶음탕' menu to the wall. Preserve the man, his pose, clothing, the boiling pot, table, and lighting.\n- Redraw the right hand to show exactly four fingers wrapping the front of the glass. Preserve the man, his face, clothing, the glass, pot, and background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "프롬프트가 지시한 대로 벽면 메뉴판에 '닭볶음탕'을 정확히 표기하였으며, 문의 '당기시오' 사인 등 레퍼런스의 디테일도 훌륭하게 유지했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "전체적인 구성과 인물의 모습은 좋으나, 벽면에 있어야 할 '닭볶음탕' 텍스트를 유리문에 잘못 배치하여 프롬프트의 간판 지시사항을 어겼습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 아래쪽을 향하고 있으며, 오른손으로 소주잔을 입가에 가져다 대고 있음.",
      "built_space": "식당 내부, 테이블 앞에 앉아 있으며 배경에 문과 벽면이 보임.",
      "entities": "전택수의 외모, 복장(경찰 잠바, 모자 등)이 레퍼런스와 일치함. 테이블 위에는 닭볶음탕이 끓고 있음. 유리문에 '닭볶음탕', 벽면에 '물은 셀프'라는 텍스트가 있음.",
      "hard_violations": [],
      "physics": "냄비는 테이블 위 버너에 안정적으로 놓여 있고, 소주잔은 전택수의 손에 의해 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 아래쪽을 향하고 있으며, 오른손으로 소주잔을 입가에 가져다 대고 있음.",
      "built_space": "식당 내부, 테이블 앞에 앉아 있으며 배경에 문과 벽면이 보임.",
      "entities": "전택수의 외모, 복장(경찰 잠바, 모자 등)이 레퍼런스와 일치함. 테이블 위에는 닭볶음탕이 끓고 있음. 유리문에 '당기시오', 벽면에 '닭볶음탕'이라는 텍스트가 있어 프롬프트 지시와 완벽히 일치함.",
      "hard_violations": [],
      "physics": "냄비는 테이블 위 버너에 안정적으로 놓여 있고, 소주잔은 전택수의 손에 의해 지탱되고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "프롬프트가 지시한 대로 벽면 메뉴판에 '닭볶음탕'을 정확히 표기하였으며, 문의 '당기시오' 사인 등 레퍼런스의 디테일도 훌륭하게 유지했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "전체적인 구성과 인물의 모습은 좋으나, 벽면에 있어야 할 '닭볶음탕' 텍스트를 유리문에 잘못 배치하여 프롬프트의 간판 지시사항을 어겼습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선은 아래쪽을 향하고 있으며, 오른손으로 소주잔을 입가에 가져다 대고 있음.",
      "built_space": "식당 내부, 테이블 앞에 앉아 있으며 배경에 문과 벽면이 보임.",
      "entities": "전택수의 외모, 복장(경찰 잠바, 모자 등)이 레퍼런스와 일치함. 테이블 위에는 닭볶음탕이 끓고 있음. 유리문에 '닭볶음탕', 벽면에 '물은 셀프'라는 텍스트가 있음.",
      "hard_violations": [],
      "physics": "냄비는 테이블 위 버너에 안정적으로 놓여 있고, 소주잔은 전택수의 손에 의해 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선은 아래쪽을 향하고 있으며, 오른손으로 소주잔을 입가에 가져다 대고 있음.",
      "built_space": "식당 내부, 테이블 앞에 앉아 있으며 배경에 문과 벽면이 보임.",
      "entities": "전택수의 외모, 복장(경찰 잠바, 모자 등)이 레퍼런스와 일치함. 테이블 위에는 닭볶음탕이 끓고 있음. 유리문에 '당기시오', 벽면에 '닭볶음탕'이라는 텍스트가 있어 프롬프트 지시와 완벽히 일치함.",
      "hard_violations": [],
      "physics": "냄비는 테이블 위 버너에 안정적으로 놓여 있고, 소주잔은 전택수의 손에 의해 지탱되고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 대로 벽면에 '닭볶음탕' 텍스트를 정확히 배치하였고, 레퍼런스의 '당기시오' 표지판을 유지하여 장소의 연속성을 잘 살림."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "'닭볶음탕' 텍스트를 지시된 벽면이 아닌 유리문에 배치하여 텍스트 위치 지시를 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 테이블 쪽을 향하며, 오른손이 소주잔을 입가로 정확히 조준하고 있음.",
      "built_space": "참조 이미지와 동일한 식당 내부. 유리문에는 '당기시오', 우측 벽면에는 '닭볶음탕' 표지가 올바르게 위치함.",
      "entities": "전택수(얼굴, 경찰 점퍼, 모자 일치), 끓고 있는 닭볶음탕 냄비, 소주잔 모두 프롬프트에 부합함.",
      "hard_violations": [],
      "physics": "인물은 의자에 안정적으로 앉아 있고, 손은 소주잔을 정상적으로 쥐고 있으며 냄비는 버너 위에서 지지됨."
     },
     {
      "label": "B",
      "direction": "시선은 아래로 향하고, 오른손으로 소주잔을 입가에 댐.",
      "built_space": "식당 내부 구조는 맞으나, 유리문에 '닭볶음탕', 벽면에 '물은 셀프'로 텍스트 위치가 잘못됨.",
      "entities": "전택수(복장 및 외모 일치), 가스버너 위의 닭볶음탕, 소주잔.",
      "hard_violations": [],
      "physics": "손으로 잔을 쥐고 있고 냄비가 버너 위에 놓여 있는 등 물리적 지지가 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 대로 벽면에 '닭볶음탕' 텍스트를 정확히 배치하였고, 레퍼런스의 '당기시오' 표지판을 유지하여 장소의 연속성을 잘 살림."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "'닭볶음탕' 텍스트를 지시된 벽면이 아닌 유리문에 배치하여 텍스트 위치 지시를 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 테이블 쪽을 향하며, 오른손이 소주잔을 입가로 정확히 조준하고 있음.",
      "built_space": "참조 이미지와 동일한 식당 내부. 유리문에는 '당기시오', 우측 벽면에는 '닭볶음탕' 표지가 올바르게 위치함.",
      "entities": "전택수(얼굴, 경찰 점퍼, 모자 일치), 끓고 있는 닭볶음탕 냄비, 소주잔 모두 프롬프트에 부합함.",
      "hard_violations": [],
      "physics": "인물은 의자에 안정적으로 앉아 있고, 손은 소주잔을 정상적으로 쥐고 있으며 냄비는 버너 위에서 지지됨."
     },
     {
      "label": "A",
      "direction": "시선은 아래로 향하고, 오른손으로 소주잔을 입가에 댐.",
      "built_space": "식당 내부 구조는 맞으나, 유리문에 '닭볶음탕', 벽면에 '물은 셀프'로 텍스트 위치가 잘못됨.",
      "entities": "전택수(복장 및 외모 일치), 가스버너 위의 닭볶음탕, 소주잔.",
      "hard_violations": [],
      "physics": "손으로 잔을 쥐고 있고 냄비가 버너 위에 놓여 있는 등 물리적 지지가 자연스러움."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 11,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S2sh9"
  }
 },
 "S71sh4::cine": {
  "applied": true,
  "fingerprint": "4bc7a71c3068cbee4f0bf1a7e5a9c2f3f9ca43d6640a4119a1e0997ee5bd3648",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S71sh4_sel.png",
  "source_sha256": "6c50f43910e5f575433dad79b5f10d9b6f8f45a9c74c6a623586a1192db104c7",
  "file": "S71sh4_cine.png",
  "latency_ms": 11662
 },
 "S71sh10::signage": {
  "fp": "ee020e39c378678f",
  "inscriptions": []
 },
 "S71sh10": {
  "input_fingerprint": "d2c3ce4460a8d72f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 웃고 있는 심옥을 향해 짠한 눈빛을 고정한 전택수의 얼굴\n\nLOCATION (lock): Across the dining table inside the empty restaurant. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close behind and slightly above 심옥's shoulder, the final dolly-in isolates 전택수's face in the center-right with a subtle downward angle, leaving only a narrow, soft shoulder edge at the lower-left as the direction of his gaze. His eyes remain fixed on 심옥's smile with restrained sorrow, while his forward-leaning seated posture keeps the moment intimate rather than posed.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, midground, looks toward 심옥 at the table; 심옥 shoulder edge in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 심옥의 어깨 가장자리 (Softly out of focus); used as A narrow out-of-focus foreground edge establishes whose smile 전택수 is watching without revealing a competing second face; 식탁 가장자리 (Meal in progress) — Only a small diagonal portion remains along the lower frame; used as Supplies minimal spatial continuity beneath the isolated faces.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The nighttime restaurant ambience stays subdued and naturalistic, with gentle tonal separation preserving the sorrow in 전택수's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same empty old restaurant, simmering pot, tableware, and warm night lighting. Exclude the drinking gesture and the seated woman from the close frame, isolating the man's compassionate look.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The simmering dish and the two soju glasses remain on the table after Sim-ok pours Taeksu another drink.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 웃고 있는 심옥을 향해 짠한 눈빛을 고정한 전택수의 얼굴\n\nLOCATION (lock): Across the dining table inside the empty restaurant. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close behind and slightly above 심옥's shoulder, the final dolly-in isolates 전택수's face in the center-right with a subtle downward angle, leaving only a narrow, soft shoulder edge at the lower-left as the direction of his gaze. His eyes remain fixed on 심옥's smile with restrained sorrow, while his forward-leaning seated posture keeps the moment intimate rather than posed.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, midground, looks toward 심옥 at the table; 심옥 shoulder edge in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 심옥의 어깨 가장자리 (Softly out of focus); used as A narrow out-of-focus foreground edge establishes whose smile 전택수 is watching without revealing a competing second face; 식탁 가장자리 (Meal in progress) — Only a small diagonal portion remains along the lower frame; used as Supplies minimal spatial continuity beneath the isolated faces.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The nighttime restaurant ambience stays subdued and naturalistic, with gentle tonal separation preserving the sorrow in 전택수's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same empty old restaurant, simmering pot, tableware, and warm night lighting. Exclude the drinking gesture and the seated woman from the close frame, isolating the man's compassionate look.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The simmering dish and the two soju glasses remain on the table after Sim-ok pours Taeksu another drink.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 웃고 있는 심옥을 향해 짠한 눈빛을 고정한 전택수의 얼굴\n\nLOCATION (lock): Across the dining table inside the empty restaurant. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Close behind and slightly above 심옥's shoulder, the final dolly-in isolates 전택수's face in the center-right with a subtle downward angle, leaving only a narrow, soft shoulder edge at the lower-left as the direction of his gaze. His eyes remain fixed on 심옥's smile with restrained sorrow, while his forward-leaning seated posture keeps the moment intimate rather than posed.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-right of the frame, midground, looks toward 심옥 at the table; 심옥 shoulder edge in the lower-left of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 심옥의 어깨 가장자리 (Softly out of focus); used as A narrow out-of-focus foreground edge establishes whose smile 전택수 is watching without revealing a competing second face; 식탁 가장자리 (Meal in progress) — Only a small diagonal portion remains along the lower frame; used as Supplies minimal spatial continuity beneath the isolated faces.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The nighttime restaurant ambience stays subdued and naturalistic, with gentle tonal separation preserving the sorrow in 전택수's eyes.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same empty old restaurant, simmering pot, tableware, and warm night lighting. Exclude the drinking gesture and the seated woman from the close frame, isolating the man's compassionate look.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The simmering dish and the two soju glasses remain on the table after Sim-ok pours Taeksu another drink.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "전택수의 시선과 얼굴 방향이 좌측 전경에 있는 대상(심옥)을 향해 고정되어 있음.",
    "built_space": "텅 빈 식당 내부 구조, 유리문, 벽면의 안내문('당기시오', '닭볶음탕')의 위치와 조명이 레퍼런스와 정확히 일치함.",
    "entities": "전택수의 인상착의(경찰 점퍼, 모자 등)가 레퍼런스와 일치함. 좌측 전경에 긴 머리를 한 인물(심옥)의 좁은 어깨 가장자리만 배치됨. 테이블 위 전골 냄비와 소주잔 2개가 명확히 존재함.",
    "hard_violations": [],
    "physics": "전택수가 의자에 앉아 팔을 내리고 있는 자세가 안정적이며 냄비와 잔이 테이블 위에 올바르게 놓여 있음."
   },
   {
    "label": "A",
    "direction": "전택수의 시선이 좌측 전경의 대상을 향하고 있음.",
    "built_space": "식당 내부의 배경 요소와 구조, 조명이 레퍼런스와 일치함.",
    "entities": "전택수의 복장과 외모는 레퍼런스와 일치하나, 좌측 전경에 지시된 '좁은 어깨 가장자리' 대신 짧은 머리를 한 남성의 뒷머리와 귀가 넓게 등장함. 테이블 위 전골 냄비와 소주잔 2개가 있음.",
    "hard_violations": [],
    "physics": "인물이 의자에 자연스럽게 앉아 있으며, 사물들이 중력에 맞게 테이블 위에 위치함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "전경 좌측에 심옥의 어깨와 머리카락 가장자리만을 좁고 부드럽게 배치하여 전택수의 시선에 집중하도록 한 연출이 지시어와 잘 부합하나, 프레이밍이 클로즈업보다는 미디엄 샷에 가깝습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "전경 좌측에 짧은 머리의 남성 뒷모습(머리와 귀)이 화면을 크게 차지하여, '경쟁하는 다른 얼굴을 노출하지 않고 좁은 어깨 가장자리만 남기라'는 지시를 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S71sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "전택수의 시선이 심옥의 얼굴이 아닌 아래쪽 식탁과 냄비를 향하고 있습니다.",
     "fix_en": "Adjust Taeksu's eyes to look horizontally toward the upper left foreground, meeting an implied face level, instead of looking down at the table. Preserve his facial features, clothing, the foreground woman, the pot, and the background.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "프레이밍이 지시된 클로즈업이 아니며 상반신 전체가 보이고, 전경의 심옥이 '좌측 하단의 좁은 어깨 가장자리'가 아닌 화면 왼쪽 상당 부분을 차지하고 있습니다.",
     "fix_en": "Reduce the woman's head and back on the left to a small, out-of-focus shoulder edge in the lower-left corner, revealing the background glass door in the freed space. Preserve Taeksu's figure, the table, the pot, and the lighting.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "식탁과 냄비가 하단을 크게 차지해 작은 대각선 가장자리만 남기라는 지시를 어겼다.",
     "fix_en": "Lower the table and pot to form only a narrow diagonal strip along the bottom edge, extending Taeksu's jacket downward into the freed space. Preserve Taeksu's upper body, the foreground woman, the background, and the tableware.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "전택수의 시선이 심옥의 얼굴이 아닌 아래쪽 식탁과 냄비를 향하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "프레이밍이 지시된 클로즈업이 아니며 상반신 전체가 보이고, 전경의 심옥이 '좌측 하단의 좁은 어깨 가장자리'가 아닌 화면 왼쪽 상당 부분을 차지하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴만 고립하는 클로즈업이 아니라 상반신·냄비·식당이 크게 보이는 미디엄 샷이다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 전경에 심옥의 머리와 등이 넓게 보여 흐린 어깨 가장자리만 두라는 구도를 벗어났다.",
     "severity": "major"
    },
    {
     "issue_ko": "식탁과 냄비가 하단을 크게 차지해 작은 대각선 가장자리만 남기라는 지시를 어겼다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "전택수의 시선이 심옥의 얼굴이 아닌 아래쪽 식탁과 냄비를 향하고 있습니다.",
    "fix_en": "Adjust Taeksu's eyes to look horizontally toward the upper left foreground, meeting an implied face level, instead of looking down at the table. Preserve his facial features, clothing, the foreground woman, the pot, and the background.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "프레이밍이 지시된 클로즈업이 아니며 상반신 전체가 보이고, 전경의 심옥이 '좌측 하단의 좁은 어깨 가장자리'가 아닌 화면 왼쪽 상당 부분을 차지하고 있습니다.",
    "fix_en": "Reduce the woman's head and back on the left to a small, out-of-focus shoulder edge in the lower-left corner, revealing the background glass door in the freed space. Preserve Taeksu's figure, the table, the pot, and the lighting.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "식탁과 냄비가 하단을 크게 차지해 작은 대각선 가장자리만 남기라는 지시를 어겼다.",
    "fix_en": "Lower the table and pot to form only a narrow diagonal strip along the bottom edge, extending Taeksu's jacket downward into the freed space. Preserve Taeksu's upper body, the foreground woman, the background, and the tableware.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S71sh4"
  }
 },
 "S71sh10::cine": {
  "applied": true,
  "fingerprint": "977d298068c882b63ef4b80527d8df7b7e9adc09e404f80445b233b0e74c9352",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S71sh10_sel.png",
  "source_sha256": "45940115a2e50472ddfdd249682fbe197fc58a5be8044cea29a50cda6bb0322c",
  "file": "S71sh10_cine.png",
  "latency_ms": 10738
 },
 "S72sh1::signage": {
  "fp": "56843f23369d8a89",
  "inscriptions": [
   {
    "surface_native": "법원 청사 표지석",
    "text_native": "광주지방법원",
    "reason_ko": "웅장한 광주지방법원 법원 청사 전경에 법원의 공식 명칭이 새겨진 표지석이 위치하여 장소의 정체성을 명확히 해줍니다."
   }
  ]
 },
 "S72sh1": {
  "input_fingerprint": "bc1eced27ee3d755",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맑은 대낮, 웅장한 광주지방법원 건물의 전경 위로 '2016년 10월'이라는 하얀색 자막이 띄워진 모습\n\nLOCATION (lock): In the courthouse's open front plaza, facing the full building facade in clear daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a moderately elevated, distant frontal three-quarter position, hold the complete Gwangju District Court building across the lower two-thirds of a wide frame, with additional open space preserved above the roofline. The camera remains static as the white title “2016년 10월” appears clearly separated from the architecture and then disappears.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Gwangju District Court building in the lower-center of the frame, background; white date title in the upper-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Gwangju District Court building (Shown in full exterior view) — Its broad frontal three-quarter exterior faces the camera, with the entire building visible; used as Primary architectural subject spanning the lower portion of the establishing frame; white date title overlay (Appears and then disappears) — The front-facing overlay reads “2016년 10월” above the building; used as Temporarily occupies the open space above the courthouse as the temporal marker.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Clear midday daylight is rendered with restrained neutral-to-cool color and moderate naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 청사 표지석: \"광주지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맑은 대낮, 웅장한 광주지방법원 건물의 전경 위로 '2016년 10월'이라는 하얀색 자막이 띄워진 모습\n\nLOCATION (lock): In the courthouse's open front plaza, facing the full building facade in clear daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a moderately elevated, distant frontal three-quarter position, hold the complete Gwangju District Court building across the lower two-thirds of a wide frame, with additional open space preserved above the roofline. The camera remains static as the white title “2016년 10월” appears clearly separated from the architecture and then disappears.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Gwangju District Court building in the lower-center of the frame, background; white date title in the upper-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Gwangju District Court building (Shown in full exterior view) — Its broad frontal three-quarter exterior faces the camera, with the entire building visible; used as Primary architectural subject spanning the lower portion of the establishing frame; white date title overlay (Appears and then disappears) — The front-facing overlay reads “2016년 10월” above the building; used as Temporarily occupies the open space above the courthouse as the temporal marker.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Clear midday daylight is rendered with restrained neutral-to-cool color and moderate naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 청사 표지석: \"광주지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 맑은 대낮, 웅장한 광주지방법원 건물의 전경 위로 '2016년 10월'이라는 하얀색 자막이 띄워진 모습\n\nLOCATION (lock): In the courthouse's open front plaza, facing the full building facade in clear daylight. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nSTRUCTURE LOOK AUTHORITY: the attached STRUCTURE LOOK photograph is the identity of the fixed structure at this location — wherever that structure appears in the frame, its shape, proportions, openings, materials and colors are LOCKED to it. The LOCATION PHOTOGRAPH remains the authority for this shot's sub-space, surroundings, time of day and lighting. If the two conflict on the structure itself, the STRUCTURE LOOK photo wins; for everything else, the LOCATION PHOTOGRAPH wins.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a moderately elevated, distant frontal three-quarter position, hold the complete Gwangju District Court building across the lower two-thirds of a wide frame, with additional open space preserved above the roofline. The camera remains static as the white title “2016년 10월” appears clearly separated from the architecture and then disappears.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: Gwangju District Court building in the lower-center of the frame, background; white date title in the upper-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: Gwangju District Court building (Shown in full exterior view) — Its broad frontal three-quarter exterior faces the camera, with the entire building visible; used as Primary architectural subject spanning the lower portion of the establishing frame; white date title overlay (Appears and then disappears) — The front-facing overlay reads “2016년 10월” above the building; used as Temporarily occupies the open space above the courthouse as the temporal marker.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Clear midday daylight is rendered with restrained neutral-to-cool color and moderate naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 청사 표지석: \"광주지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "카메라는 광장 중앙에서 법원 건물 정면을 정면 또는 약간 측면에서 넓게 바라보고 있음.",
    "built_space": "위치 사진과 일치하는 넓은 광장과 중앙의 포장된 길, 양옆의 조경이 있음. 건물은 구조물 참조 사진의 특징인 수직 기둥, 중앙 유리창, '자유평등정의'가 적힌 파란색 출입구 캐노피를 반영하여 렌더링됨.",
    "entities": "상단 하늘에 '2016년 10월' 흰색 자막이 명확하게 띄워져 있음. 건물 정면 상단에 '광주지방법원' 텍스트가 부착되어 있음. 프롬프트 지시대로 사람이나 차량은 존재하지 않음.",
    "hard_violations": [],
    "physics": "자막은 하늘에 독립적으로 떠 있으며, 건물과 광장의 구조물들은 바닥에 안정적으로 고정되어 지지되고 있음."
   },
   {
    "label": "B",
    "direction": "카메라는 광장 중앙에서 건물 정면을 넓게 바라보고 있음.",
    "built_space": "광장과 주변 조경은 위치 참조 사진과 일치함. 그러나 건물은 구조물 참조 사진이 아닌 위치 참조 사진의 건물 형태(가로형 창문, 수직 기둥 없음)를 그대로 가져옴.",
    "entities": "상단 하늘에 '2016년 10월' 흰색 자막이 띄워져 있음. 건물 정면 상단에 '광주지방법원' 텍스트가 부착됨. 사람은 존재하지 않음.",
    "hard_violations": [
     "건물의 형태, 재질, 개구부를 STRUCTURE LOOK 사진으로 고정(LOCKED)하라는 지시를 무시하고, LOCATION 사진의 건물을 그대로 렌더링함 (지정된 피사체 외형 위반)."
    ],
    "physics": "자막은 하늘에 떠 있으며, 건물과 모든 시설물은 지면에 정상적으로 지지되어 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 9,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "후보 A는 위치 사진의 공간 및 채광과 구조물 사진의 건축 양식(수직 기둥, 파란색 출입구 등)을 훌륭하게 결합하였으며, 요구된 자막과 간판도 정확하게 구현하여 지침을 가장 잘 따랐습니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "후보 B는 건물의 외관을 STRUCTURE LOOK 사진으로 고정하라는 최우선 지시를 완전히 무시하고 위치 사진의 건물을 그대로 사용하여 핵심 지침을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its spatial layout, surroundings, fixed features, time of day and lighting mood are spatial truth; stage the moment inside this place. If a STRUCTURE LOOK photograph is also attached, that photo wins for the fixed structure itself — this photograph wins for everything around it. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L61B01.png"
   },
   {
    "label": "STRUCTURE LOOK — the confirmed photograph of the fixed structure at this location: wherever the structure appears in the frame, its shape, proportions, materials, colors and openings are LOCKED to this photo. Never copy its camera framing, time of day or lighting — the shot text and the LOCATION PHOTOGRAPH are the authorities for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_district_court_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "건물 중앙 전면부 상단의 둥근 아치 구조와 원형 엠블럼이 누락되고 평평한 형태로 변형됨.",
     "fix_en": "Redraw the top center of the building's facade to include the large arched structure and circular emblem shown in the reference. Preserve the rest of the building's architecture, the entire plaza, the lighting, and the '2016년 10월' text floating in the sky.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 좌측에 레퍼런스(구조 및 장소 사진)에 존재하지 않는 꺾인 형태의 대형 부속 건물이 임의로 추가됨.",
     "fix_en": "Remove the protruding building wing on the left side, replacing it with the background sky and trees. Preserve the central building, the plaza, and the text overlay.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "카메라가 지시된 정면 3/4가 아니라 완전 정면 대칭 구도이다",
     "fix_en": "Change the camera angle to a distant frontal three-quarter position. Preserve the building's structural identity, the open space, and the text overlay.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "건물 상단에 구조 참조에 없는 '광주지방법원' 글자가 있다",
     "fix_en": "Remove the '광주지방법원' text from the building's top center and replace it with '지 방' on the left and '법 원' on the right of the center, matching the reference. Preserve the building facade and the white date title in the sky.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "광장에 요구된 '광주지방법원' 표지석이 보이지 않는다",
     "fix_en": "Add a stone signpost engraved with '광주지방법원' to the foreground plaza. Preserve the building, the plaza ground, and the date text overlay.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "건물 앞 수목의 수와 배치가 로케이션 참조와 다르다",
     "fix_en": "Adjust the pine trees in front of the building to match the exact placement in the location reference. Preserve the building, plaza, and text overlay.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "건물 중앙 전면부 상단의 둥근 아치 구조와 원형 엠블럼이 누락되고 평평한 형태로 변형됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 좌측에 레퍼런스(구조 및 장소 사진)에 존재하지 않는 꺾인 형태의 대형 부속 건물이 임의로 추가됨.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 지시된 정면 3/4가 아니라 완전 정면 대칭 구도이다",
     "severity": "major"
    },
    {
     "issue_ko": "건물 입면이 구조 참조와 다르다. 중앙 아치·문장이 없고 창호와 매스가 다르다",
     "severity": "major"
    },
    {
     "issue_ko": "건물 상단에 구조 참조에 없는 '광주지방법원' 글자가 있다",
     "severity": "major"
    },
    {
     "issue_ko": "광장에 요구된 '광주지방법원' 표지석이 보이지 않는다",
     "severity": "major"
    },
    {
     "issue_ko": "건물 앞 수목의 수와 배치가 로케이션 참조와 다르다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 좌측에 레퍼런스(구조 및 장소 사진)에 존재하지 않는 꺾인 형태의 대형 부속 건물이 임의로 추가됨.",
    "fix_en": "Remove the protruding building wing on the left side, replacing it with the background sky and trees. Preserve the central building, the plaza, and the text overlay.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "카메라가 지시된 정면 3/4가 아니라 완전 정면 대칭 구도이다",
    "fix_en": "Change the camera angle to a distant frontal three-quarter position. Preserve the building's structural identity, the open space, and the text overlay.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "건물 상단에 구조 참조에 없는 '광주지방법원' 글자가 있다",
    "fix_en": "Remove the '광주지방법원' text from the building's top center and replace it with '지 방' on the left and '법 원' on the right of the center, matching the reference. Preserve the building facade and the white date title in the sky.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "광장에 요구된 '광주지방법원' 표지석이 보이지 않는다",
    "fix_en": "Add a stone signpost engraved with '광주지방법원' to the foreground plaza. Preserve the building, the plaza ground, and the date text overlay.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "건물 앞 수목의 수와 배치가 로케이션 참조와 다르다",
    "fix_en": "Adjust the pine trees in front of the building to match the exact placement in the location reference. Preserve the building, plaza, and text overlay.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the top center of the building's facade to include the large arched structure and circular emblem shown in the reference. Preserve the rest of the building's architecture, the entire plaza, the lighting, and the '2016년 10월' text floating in the sky.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "핵심 요소인 '2016년 10월' 자막을 올바르게 배치하여 샷의 목적을 달성했으나, 건물 중앙 상단의 아치 구조를 임의로 변형하고 텍스트를 잘못 배치한 점은 아쉽습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "건물 상단의 아치와 엠블럼 등 건축적 형태는 레퍼런스와 완벽하게 복구되었으나, 프롬프트의 가장 중요한 묘사인 '2016년 10월' 자막이 완전히 누락되어 샷의 의도를 상실했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 광주지방법원 건물의 정면 전경을 향하고 있음.",
      "built_space": "레퍼런스에 있는 건물 중앙 상단의 아치형 구조와 마크가 사라지고 평평하게 변형됨.",
      "entities": "하늘 상단에 '2016년 10월' 흰색 자막이 명확히 존재함. 표지석 대신 건물 외벽 상단에 '광주지방법원' 텍스트가 임의로 기재됨.",
      "hard_violations": [],
      "physics": "자막은 화면 오버레이이므로 물리적 지지가 필요하지 않음."
     },
     {
      "label": "B",
      "direction": "카메라는 광주지방법원 건물의 정면 전경을 향하고 있음.",
      "built_space": "건물 중앙 상단의 아치, 엠블럼, 창문 배열 등 레퍼런스의 구조적 특징을 정확히 반영함.",
      "entities": "프롬프트가 요구한 '2016년 10월' 자막이 화면 어디에도 존재하지 않음.",
      "hard_violations": [],
      "physics": "물리적으로 어색하거나 지지되지 않는 요소 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "핵심 요소인 '2016년 10월' 자막을 올바르게 배치하여 샷의 목적을 달성했으나, 건물 중앙 상단의 아치 구조를 임의로 변형하고 텍스트를 잘못 배치한 점은 아쉽습니다."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "건물 상단의 아치와 엠블럼 등 건축적 형태는 레퍼런스와 완벽하게 복구되었으나, 프롬프트의 가장 중요한 묘사인 '2016년 10월' 자막이 완전히 누락되어 샷의 의도를 상실했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 광주지방법원 건물의 정면 전경을 향하고 있음.",
      "built_space": "레퍼런스에 있는 건물 중앙 상단의 아치형 구조와 마크가 사라지고 평평하게 변형됨.",
      "entities": "하늘 상단에 '2016년 10월' 흰색 자막이 명확히 존재함. 표지석 대신 건물 외벽 상단에 '광주지방법원' 텍스트가 임의로 기재됨.",
      "hard_violations": [],
      "physics": "자막은 화면 오버레이이므로 물리적 지지가 필요하지 않음."
     },
     {
      "label": "B",
      "direction": "카메라는 광주지방법원 건물의 정면 전경을 향하고 있음.",
      "built_space": "건물 중앙 상단의 아치, 엠블럼, 창문 배열 등 레퍼런스의 구조적 특징을 정확히 반영함.",
      "entities": "프롬프트가 요구한 '2016년 10월' 자막이 화면 어디에도 존재하지 않음.",
      "hard_violations": [],
      "physics": "물리적으로 어색하거나 지지되지 않는 요소 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "요구된 자막이 누락되었으나, 지정된 건축물의 핵심 형태(중앙 아치 등)를 변형 없이 보존하여 전체적인 손실이 적음."
     },
     {
      "label": "B",
      "score": 1667,
      "verdict_ko": "자막 추가를 위해 건물 중앙의 아치형 구조를 삭제하고 상단 텍스트를 임의로 변경하여 레퍼런스를 심하게 훼손함."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.667
     },
     "adjusted": {
      "A": 1.5,
      "B": 1.667
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "요구된 자막이 누락되었으나, 지정된 건축물의 핵심 형태(중앙 아치 등)를 변형 없이 보존하여 전체적인 손실이 적음."
     },
     {
      "label": "A",
      "score": 1667,
      "verdict_ko": "자막 추가를 위해 건물 중앙의 아치형 구조를 삭제하고 상단 텍스트를 임의로 변경하여 레퍼런스를 심하게 훼손함."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1673,
     "B": 1504
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "플레이트+seed만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S72sh1::cine": {
  "applied": true,
  "fingerprint": "5ac7ba70ec7027aefc7e9fcab3deda1518ee82ea09b44b832055af48a98ff34b",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S72sh1_sel.png",
  "source_sha256": "221215b06c8b1a8b8186e54f1983e15af92c073454fa792173154df88d5e5df3",
  "file": "S72sh1_cine.png",
  "latency_ms": 13603
 },
 "S73sh5::signage": {
  "fp": "cd48dcdb05289d27",
  "inscriptions": [
   {
    "surface_native": "증인석 명패",
    "text_native": "증인석",
    "reason_ko": "법정 내부의 증인석을 마주보는 구도에서 이곳이 한국의 법정임을 사실적으로 묘사하기 위해 증인석 테이블 앞의 안내 표기가 필요합니다."
   }
  ]
 },
 "era_assess::b28eb33c1d225bea": {
  "subjects": [
   {
    "subject_native": "2000년대 대한민국 법정 내부",
    "search_terms_native": [
     "대한민국 법정 내부",
     "한국 법원 법정 실내",
     "법정 방청석 증인석"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "대한민국의 법정 구조, 증인석, 방청석 배치 및 인테리어는 미국 등 서구식 법정과 형태가 크게 다르므로 실제 한국 법정의 모습을 참조해야 정확한 재현이 가능합니다."
   }
  ]
 },
 "era_ref::78d1ea3f99a920db": {
  "subject": "2000년대 대한민국 법정 내부",
  "terms": [
   "대한민국 법정 내부",
   "한국 법원 법정 실내",
   "법정 방청석 증인석"
  ],
  "queries": [
   [
    "2000년대 대한민국 법정 내부 법정 방청석 증인석",
    "한국 법원 법정 실내 방청석 증인석"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://cdn.mindlenews.com/news/photo/202401/7012_20999_4440.jpg",
  "picked_reason_ko": "1번은 별도의 관람객용 시설 없이 통상적인 현대 대한민국 법정 내부를 정면에서 폭넓게 담아 재료·비례·가구·설비를 가장 명료하게 참고할 수 있다.",
  "sha256": "b091b0c9d58c800425db5b2d760f1a8360af826af0656aca7643054f002a11f0",
  "file": "eraref_78d1ea3f99a920db.png"
 },
 "S73sh5::bgfirst_bg": {
  "input_fingerprint": "742829edb7978f56",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 두 눈을 질끈 감은 채 고통스러운 표정으로 고개를 푹 숙인 심옥의 굳은 상체\n\nLOCATION (lock): In the public gallery of the courtroom, facing the witness stand and projection screen.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the aisle-side dolly-in, frame 심옥 from slightly above her seated head height in a close three-quarter view, her face and rigid upper body occupying the left-center while the end of the seating row recedes behind her. Her shoulders fold inward, her head drops toward her lap, and her tightly closed eyes isolate her distress from the proceedings around her.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: public seating row (Occupied sparsely by spectators) — The row extends away from the camera along the side aisle; used as The edge of the row forms a restrained diagonal behind 심옥 and preserves the courtroom context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is kept subdued and neutral-to-cool, with naturalistic facial detail and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 법정 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 두 눈을 질끈 감은 채 고통스러운 표정으로 고개를 푹 숙인 심옥의 굳은 상체\n\nLOCATION (lock): In the public gallery of the courtroom, facing the witness stand and projection screen.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At the end of the aisle-side dolly-in, frame 심옥 from slightly above her seated head height in a close three-quarter view, her face and rigid upper body occupying the left-center while the end of the seating row recedes behind her. Her shoulders fold inward, her head drops toward her lap, and her tightly closed eyes isolate her distress from the proceedings around her.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: public seating row (Occupied sparsely by spectators) — The row extends away from the camera along the side aisle; used as The edge of the row forms a restrained diagonal behind 심옥 and preserves the courtroom context.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is kept subdued and neutral-to-cool, with naturalistic facial detail and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 대한민국 법정 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S73sh5__bgfirst_bg.png",
  "asset_id": "3591dbbb-33ef-4ae1-a9ad-0764c1522985",
  "input_asset_ids": [
   "bdf2390a-ec74-499f-ba31-186ad85906cf",
   "45909c88-d564-402c-9353-f8b6f753e5d1"
  ],
  "era_research": {
   "subject": "2000년대 대한민국 법정 내부",
   "queries": [
    [
     "2000년대 대한민국 법정 내부 법정 방청석 증인석",
     "한국 법원 법정 실내 방청석 증인석"
    ]
   ],
   "picked_url": "https://cdn.mindlenews.com/news/photo/202401/7012_20999_4440.jpg",
   "sha256": "b091b0c9d58c800425db5b2d760f1a8360af826af0656aca7643054f002a11f0",
   "file": "eraref_78d1ea3f99a920db.png"
  }
 },
 "S73sh5": {
  "input_fingerprint": "ca11be5d069ce478",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 질끈 감은 채 고통스러운 표정으로 고개를 푹 숙인 심옥의 굳은 상체\n\nLOCATION (lock): In the public gallery of the courtroom, facing the witness stand and projection screen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the end of the aisle-side dolly-in, frame 심옥 from slightly above her seated head height in a close three-quarter view, her face and rigid upper body occupying the left-center while the end of the seating row recedes behind her. Her shoulders fold inward, her head drops toward her lap, and her tightly closed eyes isolate her distress from the proceedings around her.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: public seating row (Occupied sparsely by spectators) — The row extends away from the camera along the side aisle; used as The edge of the row forms a restrained diagonal behind 심옥 and preserves the courtroom context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is kept subdued and neutral-to-cool, with naturalistic facial detail and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 질끈 감은 채 고통스러운 표정으로 고개를 푹 숙인 심옥의 굳은 상체\n\nLOCATION (lock): In the public gallery of the courtroom, facing the witness stand and projection screen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the end of the aisle-side dolly-in, frame 심옥 from slightly above her seated head height in a close three-quarter view, her face and rigid upper body occupying the left-center while the end of the seating row recedes behind her. Her shoulders fold inward, her head drops toward her lap, and her tightly closed eyes isolate her distress from the proceedings around her.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: public seating row (Occupied sparsely by spectators) — The row extends away from the camera along the side aisle; used as The edge of the row forms a restrained diagonal behind 심옥 and preserves the courtroom context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is kept subdued and neutral-to-cool, with naturalistic facial detail and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 눈을 질끈 감은 채 고통스러운 표정으로 고개를 푹 숙인 심옥의 굳은 상체\n\nLOCATION (lock): In the public gallery of the courtroom, facing the witness stand and projection screen. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the end of the aisle-side dolly-in, frame 심옥 from slightly above her seated head height in a close three-quarter view, her face and rigid upper body occupying the left-center while the end of the seating row recedes behind her. Her shoulders fold inward, her head drops toward her lap, and her tightly closed eyes isolate her distress from the proceedings around her.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: public seating row (Occupied sparsely by spectators) — The row extends away from the camera along the side aisle; used as The edge of the row forms a restrained diagonal behind 심옥 and preserves the courtroom context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is kept subdued and neutral-to-cool, with naturalistic facial detail and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S73sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S73sh5.png"
    },
    {
     "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:866898>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L62B03.png"
    },
    {
     "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:866898>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "공간 구조와 심옥의 묘사는 우수하나, 샷 텍스트에 명시되지 않은 배경 인물을 추가하여 엄격한 인물 제한 규정을 위반했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "방청석 바로 옆에 증인석 책상이 배치되는 심각한 공간 왜곡이 발생했으며, 역시 지시되지 않은 배경 인물이 등장했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 눈을 질끈 감고 고개를 아래로 푹 숙여 시선과 얼굴이 무릎 쪽을 향하고 있음.",
      "built_space": "방청석 벤치에 앉아 있으나, 근경 우측에 '증인석' 명패가 놓인 책상이 방청석 내부인 것처럼 어색하게 배치되어 레퍼런스의 법정 구조와 맞지 않음.",
      "entities": "심옥의 얼굴, 헤어스타일, 의상(베이지색 스웨터, 남색 바지, 크로스백)이 레퍼런스와 일치함. 샷 텍스트에 없는 관람객 2명이 배경에 존재함. '증인석' 텍스트는 정확함.",
      "hard_violations": [
       "지시되지 않은 추가 인물 등장 (invented people)",
       "물리적으로 불가능한 공간 구조 (방청석 위치에 증인석 책상 배치)"
      ],
      "physics": "벤치에 정상적으로 앉아 있으며 가방 끈과 옷 주름 등은 몸의 지지를 받고 있음."
     },
     {
      "label": "B",
      "direction": "두 눈을 꽉 감고 고통스러운 표정으로 얼굴을 무릎 방향으로 깊게 숙임.",
      "built_space": "레퍼런스 사진과 유사하게 방청석 뒤로 가림막, 증인석, 판사석이 올바른 깊이감을 두고 배치됨. 증인석 책상 전면에 '증인석' 표기가 위치함.",
      "entities": "심옥의 인상착의 및 지정된 의상이 정확히 구현됨. 샷 텍스트에 없는 관람객 3명이 배경에 등장함. '증인석' 텍스트는 정확함.",
      "hard_violations": [
       "지시되지 않은 추가 인물 등장 (invented people)"
      ],
      "physics": "엉덩이를 벤치에 대고 상체를 숙인 자세가 중력에 맞게 자연스럽게 유지되고 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "공간 구조와 심옥의 묘사는 우수하나, 샷 텍스트에 명시되지 않은 배경 인물을 추가하여 엄격한 인물 제한 규정을 위반했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "방청석 바로 옆에 증인석 책상이 배치되는 심각한 공간 왜곡이 발생했으며, 역시 지시되지 않은 배경 인물이 등장했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "두 눈을 질끈 감고 고개를 아래로 푹 숙여 시선과 얼굴이 무릎 쪽을 향하고 있음.",
      "built_space": "방청석 벤치에 앉아 있으나, 근경 우측에 '증인석' 명패가 놓인 책상이 방청석 내부인 것처럼 어색하게 배치되어 레퍼런스의 법정 구조와 맞지 않음.",
      "entities": "심옥의 얼굴, 헤어스타일, 의상(베이지색 스웨터, 남색 바지, 크로스백)이 레퍼런스와 일치함. 샷 텍스트에 없는 관람객 2명이 배경에 존재함. '증인석' 텍스트는 정확함.",
      "hard_violations": [
       "지시되지 않은 추가 인물 등장 (invented people)",
       "물리적으로 불가능한 공간 구조 (방청석 위치에 증인석 책상 배치)"
      ],
      "physics": "벤치에 정상적으로 앉아 있으며 가방 끈과 옷 주름 등은 몸의 지지를 받고 있음."
     },
     {
      "label": "B",
      "direction": "두 눈을 꽉 감고 고통스러운 표정으로 얼굴을 무릎 방향으로 깊게 숙임.",
      "built_space": "레퍼런스 사진과 유사하게 방청석 뒤로 가림막, 증인석, 판사석이 올바른 깊이감을 두고 배치됨. 증인석 책상 전면에 '증인석' 표기가 위치함.",
      "entities": "심옥의 인상착의 및 지정된 의상이 정확히 구현됨. 샷 텍스트에 없는 관람객 3명이 배경에 등장함. '증인석' 텍스트는 정확함.",
      "hard_violations": [
       "지시되지 않은 추가 인물 등장 (invented people)"
      ],
      "physics": "엉덩이를 벤치에 대고 상체를 숙인 자세가 중력에 맞게 자연스럽게 유지되고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 833,
      "verdict_ko": "지시된 인물의 포즈를 잘 구현했으며, 레퍼런스 이미지의 법정 공간 구조(좌측 스크린 등)를 가장 정확하게 재현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남녀 관객 2명을 발명·추가함 / [openrouter:x-ai/grok-4.6] 증인석 명패가 전경 크기로 방청석 옆에 붙어 위치 잠금(방청석에서 증인석을 마주봄)을 깨뜨림"
     },
     {
      "label": "A",
      "score": 1321,
      "verdict_ko": "인물의 감정 표현은 양호하나, 프로젝션 스크린을 정면 벽에 배치하는 등 로케이션 구조를 변형하여 공간 제약을 위반했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남성 관객 3명을 발명·추가함"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.321,
      "B": 0.833
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남성 관객 3명을 발명·추가함"
      ],
      "B": [
       "[openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남녀 관객 2명을 발명·추가함",
       "[openrouter:x-ai/grok-4.6] 증인석 명패가 전경 크기로 방청석 옆에 붙어 위치 잠금(방청석에서 증인석을 마주봄)을 깨뜨림"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 833,
      "verdict_ko": "지시된 인물의 포즈를 잘 구현했으며, 레퍼런스 이미지의 법정 공간 구조(좌측 스크린 등)를 가장 정확하게 재현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남녀 관객 2명을 발명·추가함 / [openrouter:x-ai/grok-4.6] 증인석 명패가 전경 크기로 방청석 옆에 붙어 위치 잠금(방청석에서 증인석을 마주봄)을 깨뜨림"
     },
     {
      "label": "B",
      "score": 1321,
      "verdict_ko": "인물의 감정 표현은 양호하나, 프로젝션 스크린을 정면 벽에 배치하는 등 로케이션 구조를 변형하여 공간 제약을 위반했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남성 관객 3명을 발명·추가함"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 836,
     "B": 1326
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "두 눈을 질끈 감고 고개를 아래로 푹 숙여 시선과 얼굴이 무릎 쪽을 향하고 있음.",
    "built_space": "방청석 벤치에 앉아 있으나, 근경 우측에 '증인석' 명패가 놓인 책상이 방청석 내부인 것처럼 어색하게 배치되어 레퍼런스의 법정 구조와 맞지 않음.",
    "entities": "심옥의 얼굴, 헤어스타일, 의상(베이지색 스웨터, 남색 바지, 크로스백)이 레퍼런스와 일치함. 샷 텍스트에 없는 관람객 2명이 배경에 존재함. '증인석' 텍스트는 정확함.",
    "hard_violations": [
     "지시되지 않은 추가 인물 등장 (invented people)",
     "물리적으로 불가능한 공간 구조 (방청석 위치에 증인석 책상 배치)"
    ],
    "physics": "벤치에 정상적으로 앉아 있으며 가방 끈과 옷 주름 등은 몸의 지지를 받고 있음."
   },
   {
    "label": "B",
    "direction": "두 눈을 꽉 감고 고통스러운 표정으로 얼굴을 무릎 방향으로 깊게 숙임.",
    "built_space": "레퍼런스 사진과 유사하게 방청석 뒤로 가림막, 증인석, 판사석이 올바른 깊이감을 두고 배치됨. 증인석 책상 전면에 '증인석' 표기가 위치함.",
    "entities": "심옥의 인상착의 및 지정된 의상이 정확히 구현됨. 샷 텍스트에 없는 관람객 3명이 배경에 등장함. '증인석' 텍스트는 정확함.",
    "hard_violations": [
     "지시되지 않은 추가 인물 등장 (invented people)"
    ],
    "physics": "엉덩이를 벤치에 대고 상체를 숙인 자세가 중력에 맞게 자연스럽게 유지되고 있음."
   }
  ],
  "totals": {
   "A": 836,
   "B": 1326
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "공간 구조와 심옥의 묘사는 우수하나, 샷 텍스트에 명시되지 않은 배경 인물을 추가하여 엄격한 인물 제한 규정을 위반했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "방청석 바로 옆에 증인석 책상이 배치되는 심각한 공간 왜곡이 발생했으며, 역시 지시되지 않은 배경 인물이 등장했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L62B03.png"
   },
   {
    "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:866898>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트의 인물 제한 규정(SHOT TEXT에 없는 인물 추가 불가)을 위반하여 배경에 3명의 남성 방청객이 묘사됨.",
     "fix_en": "Remove the three men seated in the background on the left, replacing them with empty wooden pews and the plain wall to match the rest of the vacant seating. Preserve the woman, her exact pose, clothing, the wooden bench she is on, the background furniture, flag, lighting, and framing exactly.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "심옥의 상체가 꼿꼿하고 고개를 살짝만 숙이고 있어, '고개를 푹 숙인', '어깨가 안으로 말림', '무릎 쪽으로 고개가 떨어짐'이라는 자세 지시와 일치하지 않음.",
     "fix_en": "Redraw the woman's upper body so her shoulders hunch deeply inward and her head drops heavily toward her lap in distress. Preserve her face, hair, clothing, the wooden benches, the background setting, and lighting exactly.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "배경 법대 옆에 세워진 태극기의 태극 문양이 상하(가로)가 아닌 좌우(세로)로 분할되어 잘못 그려짐.",
     "fix_en": "Redraw the yin-yang symbol on the South Korean flag in the background so it is divided horizontally rather than vertically. Preserve the flag's fabric folds, the woman, her pose, clothing, all courtroom furniture, lighting, and framing exactly.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "무릎 위에 올려놓은 양손의 손가락 형태가 서로 뭉개지고 관절 마디가 불분명하게 렌더링됨.",
     "fix_en": "Refine the woman's hands on her lap to clearly separate the individual fingers and correctly define the knuckles. Preserve the rest of her body, clothing, the wooden bench, and all background details exactly.",
     "severity": "minor",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트의 인물 제한 규정(SHOT TEXT에 없는 인물 추가 불가)을 위반하여 배경에 3명의 남성 방청객이 묘사됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "심옥의 상체가 꼿꼿하고 고개를 살짝만 숙이고 있어, '고개를 푹 숙인', '어깨가 안으로 말림', '무릎 쪽으로 고개가 떨어짐'이라는 자세 지시와 일치하지 않음.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 법대 옆에 세워진 태극기의 태극 문양이 상하(가로)가 아닌 좌우(세로)로 분할되어 잘못 그려짐.",
     "severity": "major"
    },
    {
     "issue_ko": "무릎 위에 올려놓은 양손의 손가락 형태가 서로 뭉개지고 관절 마디가 불분명하게 렌더링됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "왼쪽 뒤 방청석에 샷 텍스트에 없는 남성 세 명이 앉아 있다",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "심옥의 상체가 꼿꼿하고 고개를 살짝만 숙이고 있어, '고개를 푹 숙인', '어깨가 안으로 말림', '무릎 쪽으로 고개가 떨어짐'이라는 자세 지시와 일치하지 않음.",
    "fix_en": "Redraw the woman's upper body so her shoulders hunch deeply inward and her head drops heavily toward her lap in distress. Preserve her face, hair, clothing, the wooden benches, the background setting, and lighting exactly.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "배경 법대 옆에 세워진 태극기의 태극 문양이 상하(가로)가 아닌 좌우(세로)로 분할되어 잘못 그려짐.",
    "fix_en": "Redraw the yin-yang symbol on the South Korean flag in the background so it is divided horizontally rather than vertically. Preserve the flag's fabric folds, the woman, her pose, clothing, all courtroom furniture, lighting, and framing exactly.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "무릎 위에 올려놓은 양손의 손가락 형태가 서로 뭉개지고 관절 마디가 불분명하게 렌더링됨.",
    "fix_en": "Refine the woman's hands on her lap to clearly separate the individual fingers and correctly define the knuckles. Preserve the rest of her body, clothing, the wooden bench, and all background details exactly.",
    "severity": "minor",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the three men seated in the background on the left, replacing them with empty wooden pews and the plain wall to match the rest of the vacant seating. Preserve the woman, her exact pose, clothing, the wooden bench she is on, the background furniture, flag, lighting, and framing exactly.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "SHOT TEXT에 없는 인물을 배제하라는 최우선 규칙을 정확히 준수하여 배경의 방청객을 주변 구조물 훼손 없이 깔끔하게 제거한 성공적인 수정본입니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경 설정에 방청객이 언급되었으나, SHOT TEXT에 명시되지 않은 인물의 등장을 엄격히 금지하는 'PEOPLE' 규칙을 위반하여 감점되었습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "심옥은 고개를 푹 숙인 채 무릎 쪽을 향하고 있으며, 두 눈을 질끈 감고 있음.",
      "built_space": "법정 내부의 방청석, 증인석, 판사석이 카메라 시점에 맞게 자연스럽게 배치되어 있으며, 심옥 뒤로 이어지는 벤치는 비어 있음.",
      "entities": "심옥의 얼굴, 헤어스타일, 의상(베이지색 상의, 짙은 바지, 가방)이 레퍼런스와 완벽히 일치하며, 명패에 '증인석' 텍스트가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "심옥은 방청석 벤치에 안정적으로 앉아 체중을 지지하고 있으며, 가방은 어깨에 걸쳐진 상태임."
     },
     {
      "label": "A",
      "direction": "심옥은 고개를 푹 숙인 채 무릎 쪽을 향하고 있으며, 두 눈을 질끈 감고 있음.",
      "built_space": "법정 내부의 방청석, 증인석, 판사석이 배치되어 있으며, 배경 좌측 벤치에 여러 사람들이 앉아 있음.",
      "entities": "심옥의 외모와 의상이 레퍼런스와 일치하고 '증인석' 명패도 정확하나, 배경에 SHOT TEXT가 요구하지 않은 추가 인물 3명이 등장함.",
      "hard_violations": [
       "extra bodies (SHOT TEXT에 없는 배경 인물 3명 등장)"
      ],
      "physics": "심옥과 배경 인물들 모두 벤치에 몸을 지탱하고 앉아 있으며, 가방 끈이 어깨에 자연스럽게 놓여 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "SHOT TEXT에 없는 인물을 배제하라는 최우선 규칙을 정확히 준수하여 배경의 방청객을 주변 구조물 훼손 없이 깔끔하게 제거한 성공적인 수정본입니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경 설정에 방청객이 언급되었으나, SHOT TEXT에 명시되지 않은 인물의 등장을 엄격히 금지하는 'PEOPLE' 규칙을 위반하여 감점되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "심옥은 고개를 푹 숙인 채 무릎 쪽을 향하고 있으며, 두 눈을 질끈 감고 있음.",
      "built_space": "법정 내부의 방청석, 증인석, 판사석이 카메라 시점에 맞게 자연스럽게 배치되어 있으며, 심옥 뒤로 이어지는 벤치는 비어 있음.",
      "entities": "심옥의 얼굴, 헤어스타일, 의상(베이지색 상의, 짙은 바지, 가방)이 레퍼런스와 완벽히 일치하며, 명패에 '증인석' 텍스트가 정확히 적혀 있음.",
      "hard_violations": [],
      "physics": "심옥은 방청석 벤치에 안정적으로 앉아 체중을 지지하고 있으며, 가방은 어깨에 걸쳐진 상태임."
     },
     {
      "label": "A",
      "direction": "심옥은 고개를 푹 숙인 채 무릎 쪽을 향하고 있으며, 두 눈을 질끈 감고 있음.",
      "built_space": "법정 내부의 방청석, 증인석, 판사석이 배치되어 있으며, 배경 좌측 벤치에 여러 사람들이 앉아 있음.",
      "entities": "심옥의 외모와 의상이 레퍼런스와 일치하고 '증인석' 명패도 정확하나, 배경에 SHOT TEXT가 요구하지 않은 추가 인물 3명이 등장함.",
      "hard_violations": [
       "extra bodies (SHOT TEXT에 없는 배경 인물 3명 등장)"
      ],
      "physics": "심옥과 배경 인물들 모두 벤치에 몸을 지탱하고 앉아 있으며, 가방 끈이 어깨에 자연스럽게 놓여 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "샷 텍스트에 명시된 심옥만을 정확히 묘사하며, 엄격한 인물 추가 배제 지침과 카메라 구도를 충실히 준수함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "샷 텍스트에 언급되지 않은 인물들(배경의 남성 방청객 3명)을 임의로 추가하여 인물 생성 제한 지침을 심각하게 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 눈을 질끈 감고 고개를 아래로 푹 숙임.",
      "built_space": "법정 방청석에 위치하며, 뒤쪽으로 증인석과 법대 구조물이 올바르게 배치됨.",
      "entities": "심옥(레퍼런스와 일치하는 인상, 의상, 가방), '증인석' 명패 일치. 다른 인물 없음.",
      "hard_violations": [],
      "physics": "방청석 나무 의자에 앉아 체중을 자연스럽게 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "두 눈을 질끈 감고 고개를 아래로 푹 숙임.",
      "built_space": "법정 방청석에 위치하며, 증인석과 법대 구조물이 배경에 올바르게 배치됨.",
      "entities": "심옥 일치, '증인석' 명패 일치. 샷 텍스트에 없는 남성 방청객 3명 추가됨.",
      "hard_violations": [
       "invented people (샷 텍스트에 명시되지 않은 배경 인물 3명 추가)"
      ],
      "physics": "방청석 의자에 안정적으로 앉아 체중을 지탱함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "샷 텍스트에 명시된 심옥만을 정확히 묘사하며, 엄격한 인물 추가 배제 지침과 카메라 구도를 충실히 준수함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "샷 텍스트에 언급되지 않은 인물들(배경의 남성 방청객 3명)을 임의로 추가하여 인물 생성 제한 지침을 심각하게 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "두 눈을 질끈 감고 고개를 아래로 푹 숙임.",
      "built_space": "법정 방청석에 위치하며, 뒤쪽으로 증인석과 법대 구조물이 올바르게 배치됨.",
      "entities": "심옥(레퍼런스와 일치하는 인상, 의상, 가방), '증인석' 명패 일치. 다른 인물 없음.",
      "hard_violations": [],
      "physics": "방청석 나무 의자에 앉아 체중을 자연스럽게 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "두 눈을 질끈 감고 고개를 아래로 푹 숙임.",
      "built_space": "법정 방청석에 위치하며, 증인석과 법대 구조물이 배경에 올바르게 배치됨.",
      "entities": "심옥 일치, '증인석' 명패 일치. 샷 텍스트에 없는 남성 방청객 3명 추가됨.",
      "hard_violations": [
       "invented people (샷 텍스트에 명시되지 않은 배경 인물 3명 추가)"
      ],
      "physics": "방청석 의자에 안정적으로 앉아 체중을 지탱함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S73sh5__bgfirst_bg.png",
   "bg_asset_id": "3591dbbb-33ef-4ae1-a9ad-0764c1522985",
   "bg_record_key": "S73sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S73sh5::cine": {
  "applied": true,
  "fingerprint": "9c26acdbb57f35137f569254740934caa363e52d743f67d46e824bdba9a017cf",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S73sh5_sel.png",
  "source_sha256": "c72d6e8992c32dfe04e48a8df8c64e36a0435f012cdbdc520e1d8e5954fc2ec2",
  "file": "S73sh5_cine.png",
  "latency_ms": 14786
 },
 "S73sh6::signage": {
  "fp": "b441fd81cf6c87bb",
  "inscriptions": [
   {
    "surface_native": "빔프로젝터 스크린 화면",
    "text_native": "증 제12호증: 부검 사진",
    "reason_ko": "법정 내 스크린에 투사된 부검 사진에 실제 재판에서 쓰이는 증거 번호와 제목을 표시하여 장면의 사실성을 높이기 위함입니다."
   }
  ]
 },
 "S73sh6": {
  "input_fingerprint": "6ed9e46472db0a1a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 화면 속 꼬리뼈에 피가 고인 부검 사진을 향해 손가락을 뻗은 장원섭의 상체\n\nLOCATION (lock): Before the witness stand and projection screen inside the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track beside 장원섭 at standing chest height in a tightened medium frame, placing his upper body on the left and the courtroom screen on the right along an oblique diagonal. His extended finger leads directly to the referenced area of the displayed autopsy photograph while his eyes remain directed toward the medical examiner outside the frame, keeping the gesture evidentiary rather than sensational.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground, points to courtroom screen; courtroom screen in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying photograph number 3) — The display face is visible obliquely to camera and shows the referenced autopsy photograph, including the tailbone area with pooled blood; used as Receives the line of 장원섭’s pointing arm and keeps the referenced evidence legible at a restrained scale; witness stand (Occupied by the medical examiner) — Its open witness-facing side is angled away from the camera; used as Provides the procedural location just beyond 장원섭’s gesture.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience remains restrained and neutral-to-cool, avoiding heightened emphasis on the displayed evidence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photograph showing pooled blood near the tailbone remains displayed on the courtroom screen as Wonseop points to it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 빔프로젝터 스크린 화면: \"증 제12호증: 부검 사진\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 화면 속 꼬리뼈에 피가 고인 부검 사진을 향해 손가락을 뻗은 장원섭의 상체\n\nLOCATION (lock): Before the witness stand and projection screen inside the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track beside 장원섭 at standing chest height in a tightened medium frame, placing his upper body on the left and the courtroom screen on the right along an oblique diagonal. His extended finger leads directly to the referenced area of the displayed autopsy photograph while his eyes remain directed toward the medical examiner outside the frame, keeping the gesture evidentiary rather than sensational.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground, points to courtroom screen; courtroom screen in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying photograph number 3) — The display face is visible obliquely to camera and shows the referenced autopsy photograph, including the tailbone area with pooled blood; used as Receives the line of 장원섭’s pointing arm and keeps the referenced evidence legible at a restrained scale; witness stand (Occupied by the medical examiner) — Its open witness-facing side is angled away from the camera; used as Provides the procedural location just beyond 장원섭’s gesture.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience remains restrained and neutral-to-cool, avoiding heightened emphasis on the displayed evidence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photograph showing pooled blood near the tailbone remains displayed on the courtroom screen as Wonseop points to it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 빔프로젝터 스크린 화면: \"증 제12호증: 부검 사진\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 화면 속 꼬리뼈에 피가 고인 부검 사진을 향해 손가락을 뻗은 장원섭의 상체\n\nLOCATION (lock): Before the witness stand and projection screen inside the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track beside 장원섭 at standing chest height in a tightened medium frame, placing his upper body on the left and the courtroom screen on the right along an oblique diagonal. His extended finger leads directly to the referenced area of the displayed autopsy photograph while his eyes remain directed toward the medical examiner outside the frame, keeping the gesture evidentiary rather than sensational.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, midground, points to courtroom screen; courtroom screen in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying photograph number 3) — The display face is visible obliquely to camera and shows the referenced autopsy photograph, including the tailbone area with pooled blood; used as Receives the line of 장원섭’s pointing arm and keeps the referenced evidence legible at a restrained scale; witness stand (Occupied by the medical examiner) — Its open witness-facing side is angled away from the camera; used as Provides the procedural location just beyond 장원섭’s gesture.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience remains restrained and neutral-to-cool, avoiding heightened emphasis on the displayed evidence.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photograph showing pooled blood near the tailbone remains displayed on the courtroom screen as Wonseop points to it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 빔프로젝터 스크린 화면: \"증 제12호증: 부검 사진\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "장원섭의 시선은 화면 밖 우측을 향하고, 오른손 검지는 스크린의 부검 사진을 정확히 가리키고 있음.",
    "built_space": "법정 내부. 화면 우측 하단에 이전 샷과 동일한 '증인석' 명패가 붙은 구조물이 있으며, 배경에 태극기가 2개 나란히 배치됨.",
    "entities": "장원섭의 얼굴은 레퍼런스와 일치하나 법복이 아닌 정장을 입고 있음. 스크린 텍스트 '증 제12호증: 부검 사진'이 완벽히 표기되었고, 사진 내 꼬리뼈에 피가 고인 상처도 사실적으로 묘사됨.",
    "hard_violations": [
     "배경에 태극기가 2개 중복되어 생성됨"
    ],
    "physics": "바닥에 서서 스스로 체중을 지탱하며 자연스러운 삿대질 자세를 취하고 있음."
   },
   {
    "label": "A",
    "direction": "장원섭의 시선은 화면 밖 우측을 향하고, 오른손 검지는 스크린을 향해 뻗어 있음.",
    "built_space": "법정 내부. 화면 우측 하단에 증인석이 있으나 명패가 누락되었고, 배경에 태극기가 1개 있음.",
    "entities": "장원섭의 얼굴은 일치하나 법복 대신 정장을 착용함. 스크린의 텍스트는 정확하나, 부검 사진이 인체의 꼬리뼈가 아닌 분리된 짐승의 뼈나 관절 부위처럼 잘못 묘사됨.",
    "hard_violations": [
     "프롬프트가 명확히 요구한 인체의 '꼬리뼈' 대신 정체불명의 뼈가 스크린에 발명되어 나타남"
    ],
    "physics": "바닥에 서서 스스로 체중을 지탱하며 팔을 뻗은 자세를 취함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 6,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "의상이 레퍼런스의 법복 대신 정장으로 대체되고 태극기가 중복된 오류가 있으나, 꼬리뼈 부검 사진과 지정된 텍스트, 증인석 명패 등 핵심 요소를 훨씬 정확하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "스크린에 요구된 인체의 꼬리뼈 대신 정체불명의 뼈가 나타나고 증인석 명패가 누락되어 핵심 묘사에 크게 실패했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S73sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:796050>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "장원섭이 캐릭터 레퍼런스에 지정된 법복이 아닌 일반 정장을 입고 있음.",
     "fix_en": "Replace Wonseop's grey suit and tie with the black courtroom robe and wide white pleated jabot from the character reference. Preserve Wonseop's face, pose, pointing hand, the projection screen contents, the courtroom set, lighting, and camera framing exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이전 장면 레퍼런스에 없던 두 개의 태극기가 화면 중앙의 증인석 뒤편 배경에 임의로 추가됨.",
     "fix_en": "Would remove the two South Korean flags from behind the witness stand, leaving the plain wood-paneled wall seen in the previous shot.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "이전 장면 레퍼런스에서 나무 단상 위에 올려져 있던 '증인석' 명패가 생성된 이미지에서는 단상 앞면 패널에 매립된 형태로 구조가 변경됨.",
     "fix_en": "Would convert the '증인석' sign from an embedded panel into a freestanding wooden nameplate resting on top of the witness stand desk.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "장원섭의 시선이 프레임 밖 감정인이 아니라 스크린·가리키는 쪽을 향한다",
     "fix_en": "Would redirect Wonseop's eyes to look off-camera to the left, toward the medical examiner, rather than at the screen.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "스크린 글자의 12가 판독 불가여야 하는데 선명하게 읽힌다",
     "fix_en": "Would apply shallow focus to the screen text so the decisive numeral '12' becomes illegible.",
     "severity": "minor",
     "observation_index": 7
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "장원섭이 캐릭터 레퍼런스에 지정된 법복이 아닌 일반 정장을 입고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 장면 레퍼런스에 없던 두 개의 태극기가 화면 중앙의 증인석 뒤편 배경에 임의로 추가됨.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 장면 레퍼런스에서 나무 단상 위에 올려져 있던 '증인석' 명패가 생성된 이미지에서는 단상 앞면 패널에 매립된 형태로 구조가 변경됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "장원섭이 레퍼런스의 검은 법복·흰 jabot이 아니라 회색 양복과 넥타이를 입고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "장원섭의 시선이 프레임 밖 감정인이 아니라 스크린·가리키는 쪽을 향한다",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷에 없던 태극기가 하나 더 있어 깃발이 두 개다",
     "severity": "major"
    },
    {
     "issue_ko": "스크린의 부검 사진이 절제된 크기가 아니라 프레임 오른쪽을 크게 차지한다",
     "severity": "major"
    },
    {
     "issue_ko": "스크린 글자의 12가 판독 불가여야 하는데 선명하게 읽힌다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "이전 장면 레퍼런스에 없던 두 개의 태극기가 화면 중앙의 증인석 뒤편 배경에 임의로 추가됨.",
    "fix_en": "Would remove the two South Korean flags from behind the witness stand, leaving the plain wood-paneled wall seen in the previous shot.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "이전 장면 레퍼런스에서 나무 단상 위에 올려져 있던 '증인석' 명패가 생성된 이미지에서는 단상 앞면 패널에 매립된 형태로 구조가 변경됨.",
    "fix_en": "Would convert the '증인석' sign from an embedded panel into a freestanding wooden nameplate resting on top of the witness stand desk.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "장원섭의 시선이 프레임 밖 감정인이 아니라 스크린·가리키는 쪽을 향한다",
    "fix_en": "Would redirect Wonseop's eyes to look off-camera to the left, toward the medical examiner, rather than at the screen.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "스크린 글자의 12가 판독 불가여야 하는데 선명하게 읽힌다",
    "fix_en": "Would apply shallow focus to the screen text so the decisive numeral '12' becomes illegible.",
    "severity": "minor",
    "observation_index": 7
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace Wonseop's grey suit and tie with the black courtroom robe and wide white pleated jabot from the character reference. Preserve Wonseop's face, pose, pointing hand, the projection screen contents, the courtroom set, lighting, and camera framing exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "장원섭의 얼굴뿐만 아니라 레퍼런스에 지정된 법복 의상까지 완벽하게 재현했으며, 지정된 스크린 텍스트와 부검 사진, 법정 배경까지 지시문을 모두 충실히 이행했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경과 스크린의 묘사는 훌륭하지만, 캐릭터 레퍼런스의 법복 의상을 일반 정장으로 임의 변경하여 아이덴티티 유지에 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭이 오른쪽 스크린의 부검 사진을 손가락으로 가리키고 있으며 시선은 화면 밖 정면을 향함.",
      "built_space": "이전 샷 레퍼런스와 동일한 법정 구조. 스크린과 증인석의 위치, 각도가 정확함.",
      "entities": "스크린의 텍스트와 부검 사진 일치. 장원섭의 얼굴은 맞으나 의상이 캐릭터 레퍼런스(법복)와 다른 정장임.",
      "hard_violations": [],
      "physics": "바닥에 서서 지탱하는 체중과 뻗은 팔의 자세가 자연스러움."
     },
     {
      "label": "B",
      "direction": "장원섭이 오른쪽 스크린의 부검 사진을 가리키고 시선은 화면 밖을 향함.",
      "built_space": "이전 샷 레퍼런스와 동일한 법정. 증인석과 스크린의 배치 및 디테일이 정확함.",
      "entities": "장원섭의 얼굴 및 의상(법복)이 캐릭터 레퍼런스와 완벽히 일치함. 스크린의 텍스트와 사진 지시문 부합.",
      "hard_violations": [],
      "physics": "서 있는 자세와 팔을 뻗은 동작의 지지 상태가 정상적이고 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "장원섭의 얼굴뿐만 아니라 레퍼런스에 지정된 법복 의상까지 완벽하게 재현했으며, 지정된 스크린 텍스트와 부검 사진, 법정 배경까지 지시문을 모두 충실히 이행했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "배경과 스크린의 묘사는 훌륭하지만, 캐릭터 레퍼런스의 법복 의상을 일반 정장으로 임의 변경하여 아이덴티티 유지에 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭이 오른쪽 스크린의 부검 사진을 손가락으로 가리키고 있으며 시선은 화면 밖 정면을 향함.",
      "built_space": "이전 샷 레퍼런스와 동일한 법정 구조. 스크린과 증인석의 위치, 각도가 정확함.",
      "entities": "스크린의 텍스트와 부검 사진 일치. 장원섭의 얼굴은 맞으나 의상이 캐릭터 레퍼런스(법복)와 다른 정장임.",
      "hard_violations": [],
      "physics": "바닥에 서서 지탱하는 체중과 뻗은 팔의 자세가 자연스러움."
     },
     {
      "label": "B",
      "direction": "장원섭이 오른쪽 스크린의 부검 사진을 가리키고 시선은 화면 밖을 향함.",
      "built_space": "이전 샷 레퍼런스와 동일한 법정. 증인석과 스크린의 배치 및 디테일이 정확함.",
      "entities": "장원섭의 얼굴 및 의상(법복)이 캐릭터 레퍼런스와 완벽히 일치함. 스크린의 텍스트와 사진 지시문 부합.",
      "hard_violations": [],
      "physics": "서 있는 자세와 팔을 뻗은 동작의 지지 상태가 정상적이고 자연스러움."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "캐릭터 레퍼런스의 법복 의상을 정확하게 유지하며 프롬프트가 요구한 구도와 스크린 텍스트를 충실히 구현함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "구도와 텍스트는 정확하나, 캐릭터 레퍼런스에 명시된 법복을 임의로 정장으로 변경하여 큰 감점 요인이 됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 화면 밖 우측을 향하고, 뻗은 손가락은 스크린의 부검 사진을 정확히 가리킴.",
      "built_space": "법정 내부 공간의 증인석과 스크린이 명시된 위치에 자연스럽게 배치됨.",
      "entities": "인물의 외모와 법복 의상이 레퍼런스와 일치하며, 스크린의 텍스트와 부검 사진이 지시대로 표현됨.",
      "hard_violations": [],
      "physics": "서 있는 상태에서 팔을 뻗어 가리키는 자세가 지지면 위에서 안정적임."
     },
     {
      "label": "B",
      "direction": "시선은 화면 밖 우측을 향하고, 뻗은 손가락은 스크린의 부검 사진을 정확히 가리킴.",
      "built_space": "법정 내부 공간의 증인석과 스크린이 명시된 위치에 자연스럽게 배치됨.",
      "entities": "스크린 텍스트와 인물 외모는 일치하나, 레퍼런스의 법복 대신 정장을 입고 있음.",
      "hard_violations": [],
      "physics": "서 있는 상태에서 팔을 뻗어 가리키는 자세가 지지면 위에서 안정적임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "캐릭터 레퍼런스의 법복 의상을 정확하게 유지하며 프롬프트가 요구한 구도와 스크린 텍스트를 충실히 구현함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "구도와 텍스트는 정확하나, 캐릭터 레퍼런스에 명시된 법복을 임의로 정장으로 변경하여 큰 감점 요인이 됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "시선은 화면 밖 우측을 향하고, 뻗은 손가락은 스크린의 부검 사진을 정확히 가리킴.",
      "built_space": "법정 내부 공간의 증인석과 스크린이 명시된 위치에 자연스럽게 배치됨.",
      "entities": "인물의 외모와 법복 의상이 레퍼런스와 일치하며, 스크린의 텍스트와 부검 사진이 지시대로 표현됨.",
      "hard_violations": [],
      "physics": "서 있는 상태에서 팔을 뻗어 가리키는 자세가 지지면 위에서 안정적임."
     },
     {
      "label": "A",
      "direction": "시선은 화면 밖 우측을 향하고, 뻗은 손가락은 스크린의 부검 사진을 정확히 가리킴.",
      "built_space": "법정 내부 공간의 증인석과 스크린이 명시된 위치에 자연스럽게 배치됨.",
      "entities": "스크린 텍스트와 인물 외모는 일치하나, 레퍼런스의 법복 대신 정장을 입고 있음.",
      "hard_violations": [],
      "physics": "서 있는 상태에서 팔을 뻗어 가리키는 자세가 지지면 위에서 안정적임."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S73sh5"
  }
 },
 "S73sh6::cine": {
  "applied": false,
  "fingerprint": "f27a82902fe711af64df0412a7045454d01c709381e578ac837b7c4002bd48f2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S73sh6_sel.png",
  "source_sha256": "bad5cc44948092c0da385cd9b44034f67a4741e840cb149352b55616290821ec",
  "error": "RuntimeError: Grok image API error 400: {\"error\":{\"message\":\"Provider returned error\",\"code\":400,\"metadata\":{\"raw\":\"{\\\"code\\\":\\\"imagine:content-moderated\\\",\\\"error\\\":\\\"Generated image rejected by content moderation.\\\",\\\"usage\\\":{\\\"cost_in_usd_ticks\\\":700000000}}\",\"provider_name\":\"xAI\",\"is_byok\":false}},\"user_id\":\"org_3HnYgDbzLu9jaNLuyBrNsWTAFk2\"}"
 },
 "S73sh9::signage": {
  "fp": "e90e63931f85d1c5",
  "inscriptions": [
   {
    "surface_native": "증인석 명패",
    "text_native": "증인석",
    "reason_ko": "법정 내부의 증인석에 앉아 신문을 받고 있는 부검의의 상황과 위치를 시각적으로 명확히 전달하기 위해 증인석 표지판이 필요합니다."
   }
  ]
 },
 "S73sh9": {
  "input_fingerprint": "bbfaaf72d5d6c1cf",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현의 변호인의 시선을 피한 채 입술을 꾹 다문 김민호 부검의의 곤혹스러운 얼굴 클로즈업\n\nLOCATION (lock): At the witness stand inside the courtroom, facing the questioning attorney. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from just behind 지국현의 변호인’s shoulder, leaving the attorney as a narrow, soft foreground edge while 김민호 부검의’s off-center face occupies most of the frame across the witness stand. 김민호 부검의 presses his lips together and turns his eyes down and away from the attorney, with a small amount of look room preserving the direction of the avoided confrontation.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현의 변호인 in the middle-left of the frame, foreground, looks toward medical examiner; 김민호 부검의 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: witness stand (Occupied by 김민호 부검의) — The camera looks across its attorney-facing side toward the seated witness; used as A partial boundary beneath 김민호 부검의 keeps him located at the witness position without distracting from his expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime courtroom ambience holds natural facial texture and restrained contrast through the hesitant reaction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, projection screen, witness stand, daylight, and formal court lighting. Exclude the prosecutor's pointing gesture and isolate the medical examiner's uneasy face.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photographs remain displayed on the courtroom screen during the medical examiner's cross-examination.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김민호 부검의 (Korean 남성, 50대 중반, 차분한 얼굴, 짧게 정돈된 검은색 머리) — wearing: 법의학자 및 부검의가 착용하는 흰색 실험실 가운. 안에는 일상적인 셔츠와 슬랙스를 받쳐 입는다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현의 변호인의 시선을 피한 채 입술을 꾹 다문 김민호 부검의의 곤혹스러운 얼굴 클로즈업\n\nLOCATION (lock): At the witness stand inside the courtroom, facing the questioning attorney. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from just behind 지국현의 변호인’s shoulder, leaving the attorney as a narrow, soft foreground edge while 김민호 부검의’s off-center face occupies most of the frame across the witness stand. 김민호 부검의 presses his lips together and turns his eyes down and away from the attorney, with a small amount of look room preserving the direction of the avoided confrontation.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현의 변호인 in the middle-left of the frame, foreground, looks toward medical examiner; 김민호 부검의 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: witness stand (Occupied by 김민호 부검의) — The camera looks across its attorney-facing side toward the seated witness; used as A partial boundary beneath 김민호 부검의 keeps him located at the witness position without distracting from his expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime courtroom ambience holds natural facial texture and restrained contrast through the hesitant reaction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, projection screen, witness stand, daylight, and formal court lighting. Exclude the prosecutor's pointing gesture and isolate the medical examiner's uneasy face.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photographs remain displayed on the courtroom screen during the medical examiner's cross-examination.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김민호 부검의 (Korean 남성, 50대 중반, 차분한 얼굴, 짧게 정돈된 검은색 머리) — wearing: 법의학자 및 부검의가 착용하는 흰색 실험실 가운. 안에는 일상적인 셔츠와 슬랙스를 받쳐 입는다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현의 변호인의 시선을 피한 채 입술을 꾹 다문 김민호 부검의의 곤혹스러운 얼굴 클로즈업\n\nLOCATION (lock): At the witness stand inside the courtroom, facing the questioning attorney. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the dolly-in from just behind 지국현의 변호인’s shoulder, leaving the attorney as a narrow, soft foreground edge while 김민호 부검의’s off-center face occupies most of the frame across the witness stand. 김민호 부검의 presses his lips together and turns his eyes down and away from the attorney, with a small amount of look room preserving the direction of the avoided confrontation.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 지국현의 변호인 in the middle-left of the frame, foreground, looks toward medical examiner; 김민호 부검의 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: witness stand (Occupied by 김민호 부검의) — The camera looks across its attorney-facing side toward the seated witness; used as A partial boundary beneath 김민호 부검의 keeps him located at the witness position without distracting from his expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime courtroom ambience holds natural facial texture and restrained contrast through the hesitant reaction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, projection screen, witness stand, daylight, and formal court lighting. Exclude the prosecutor's pointing gesture and isolate the medical examiner's uneasy face.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The autopsy photographs remain displayed on the courtroom screen during the medical examiner's cross-examination.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 김민호 부검의 (Korean 남성, 50대 중반, 차분한 얼굴, 짧게 정돈된 검은색 머리) — wearing: 법의학자 및 부검의가 착용하는 흰색 실험실 가운. 안에는 일상적인 셔츠와 슬랙스를 받쳐 입는다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.429,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.429,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.179,
    "B": 1.571
   },
   "violations": {
    "A": [
     "[gemini-pro] 인물이 지시된 공간(증인석)에 위치하지 않고 다른 곳에 서 있으며, 배경에 빈 증인석이 노출됨"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1571,
   "A": 1179
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "부검의가 증인석에 올바르게 배치되어 구조물이 자연스러운 경계를 이루며 시선 처리도 훌륭하나, 요구된 클로즈업 샷보다 프레임이 다소 넓게(미디엄 샷) 렌더링되었습니다."
   },
   {
    "label": "A",
    "score": 1179,
    "verdict_ko": "클로즈업 척도에는 더 가깝지만, 인물이 증인석을 벗어난 엉뚱한 곳에 위치하고 배경에 빈 증인석이 그대로 노출되는 치명적인 공간 배치 오류가 발생했습니다.  ★위반: [gemini-pro] 인물이 지시된 공간(증인석)에 위치하지 않고 다른 곳에 서 있으며, 배경에 빈 증인석이 노출됨"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S73sh6_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "부검의가 변호인의 시선을 피하지 않고 똑바로 응시하고 있음 (지시문은 시선을 아래로 피할 것을 명시함).",
     "fix_en": "Redraw the medical examiner's eyes so his pupils point downward and away from the attorney, keeping his lips pressed together, while preserving his facial features, hair, clothing, the foreground attorney, and the background screen.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "증인석에 앉아 있어야(seated) 할 부검의가 일어서서 단상에 손을 짚고 있음.",
     "fix_en": "Lower the medical examiner's body so he appears seated behind the witness stand, removing his hands from the ledge, while preserving his face, clothing, the foreground attorney, the background screen, and the camera position.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신·증인석·스크린이 다 보이는 미디엄 샷이라 김민호 얼굴이 화면 대부분을 차지하지 않는다.",
     "fix_en": "Apply a stronger depth-of-field blur to the foreground attorney and the background elements to better isolate the medical examiner's face, while preserving the current camera framing, character poses, and room geometry.",
     "severity": "critical",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "부검의가 변호인의 시선을 피하지 않고 똑바로 응시하고 있음 (지시문은 시선을 아래로 피할 것을 명시함).",
     "severity": "critical"
    },
    {
     "issue_ko": "증인석에 앉아 있어야(seated) 할 부검의가 일어서서 단상에 손을 짚고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "부검의의 얼굴이 화면 대부분을 차지하는 클로즈업이어야 하나, 허리 위까지 보이는 미디엄 샷으로 연출됨.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신·증인석·스크린이 다 보이는 미디엄 샷이라 김민호 얼굴이 화면 대부분을 차지하지 않는다.",
     "severity": "critical"
    },
    {
     "issue_ko": "김민호가 변호인 시선을 피하거나 눈을 내리깔지 않고 왼쪽 변호인 쪽을 정면으로 바라본다.",
     "severity": "major"
    },
    {
     "issue_ko": "김민호가 증인석에 앉지 않고 포디엄 왼쪽 바깥 난간에 양손을 얹고 서 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "변호인 뒷모습이 좁고 흐린 전경 가장자리가 아니라 왼쪽을 크게 차지하는 선명한 인물이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the medical examiner's eyes so his pupils point downward and away from the attorney, keeping his lips pressed together, while preserving his facial features, hair, clothing, the foreground attorney, and the background screen.\n- Lower the medical examiner's body so he appears seated behind the witness stand, removing his hands from the ledge, while preserving his face, clothing, the foreground attorney, the background screen, and the camera position.\n- Apply a stronger depth-of-field blur to the foreground attorney and the background elements to better isolate the medical examiner's face, while preserving the current camera framing, character poses, and room geometry.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "카메라 프레임 지시(클로즈업, 흐릿한 전경)를 완전히 무시하고 레퍼런스의 넓은 구도를 답습했으며, 피사체를 증인석이 아닌 검사석에 세우고 명패마저 이동시키는 치명적인 공간 왜곡이 발생했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "좌측 벽면에 부검 사진이 중복 생성되는 환각 오류가 있으나, 지시된 카메라 워크(흐릿하고 좁은 전경, 클로즈업)와 부검의의 곤혹스러운 표정 연기, 정확한 위치 선정을 충실히 구현하여 더 나은 결과물입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "변호인은 부검의를 향하고 있으며, 부검의는 입술을 다물고 시선을 아래로 피하고 있음.",
      "built_space": "부검의가 레퍼런스의 좌측 검사석 위치에 서 있으며, 우측에 있어야 할 증인석 명패가 좌측 단상으로 잘못 이동됨.",
      "entities": "김민호 부검의(50대 남성, 백가운, 셔츠 차림 일치), 변호인(초점이 맞는 넓고 선명한 전경), 부검 사진 스크린, 증인석 명패.",
      "hard_violations": [
       "지정된 증인석(착석)이 아닌 좌측 검사석에 서서 위치함 (Staging violation)",
       "증인석 명패가 원래 위치(우측 책상)에서 좌측 단상으로 임의 이동됨 (Built space violation)"
      ],
      "physics": "부검의는 두 발로 서서 단상에 양손을 얹고 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "변호인은 부검의를 향하고, 부검의는 입술을 꾹 다문 채 시선을 아래로 피하고 있음.",
      "built_space": "부검의가 스크린 앞 증인석에 올바르게 위치하나, 좌측 빈 벽면에 부검 사진이 담긴 액자가 잘못 생성됨.",
      "entities": "김민호 부검의(50대 남성, 백가운, 셔츠 차림 일치), 변호인(아웃포커싱 된 좁고 부드러운 전경), 부검 사진 스크린, 증인석 명패(일부 가려짐), 벽면의 중복된 부검 사진 액자.",
      "hard_violations": [
       "스크린의 부검 사진이 좌측 벽면 액자에도 중복되어 나타남 (Leaked marker / Invented object)"
      ],
      "physics": "부검의는 증인석 뒤에 자리 잡고 있으며 책상에 의해 하단이 자연스럽게 가려져 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "카메라 프레임 지시(클로즈업, 흐릿한 전경)를 완전히 무시하고 레퍼런스의 넓은 구도를 답습했으며, 피사체를 증인석이 아닌 검사석에 세우고 명패마저 이동시키는 치명적인 공간 왜곡이 발생했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "좌측 벽면에 부검 사진이 중복 생성되는 환각 오류가 있으나, 지시된 카메라 워크(흐릿하고 좁은 전경, 클로즈업)와 부검의의 곤혹스러운 표정 연기, 정확한 위치 선정을 충실히 구현하여 더 나은 결과물입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "변호인은 부검의를 향하고 있으며, 부검의는 입술을 다물고 시선을 아래로 피하고 있음.",
      "built_space": "부검의가 레퍼런스의 좌측 검사석 위치에 서 있으며, 우측에 있어야 할 증인석 명패가 좌측 단상으로 잘못 이동됨.",
      "entities": "김민호 부검의(50대 남성, 백가운, 셔츠 차림 일치), 변호인(초점이 맞는 넓고 선명한 전경), 부검 사진 스크린, 증인석 명패.",
      "hard_violations": [
       "지정된 증인석(착석)이 아닌 좌측 검사석에 서서 위치함 (Staging violation)",
       "증인석 명패가 원래 위치(우측 책상)에서 좌측 단상으로 임의 이동됨 (Built space violation)"
      ],
      "physics": "부검의는 두 발로 서서 단상에 양손을 얹고 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "변호인은 부검의를 향하고, 부검의는 입술을 꾹 다문 채 시선을 아래로 피하고 있음.",
      "built_space": "부검의가 스크린 앞 증인석에 올바르게 위치하나, 좌측 빈 벽면에 부검 사진이 담긴 액자가 잘못 생성됨.",
      "entities": "김민호 부검의(50대 남성, 백가운, 셔츠 차림 일치), 변호인(아웃포커싱 된 좁고 부드러운 전경), 부검 사진 스크린, 증인석 명패(일부 가려짐), 벽면의 중복된 부검 사진 액자.",
      "hard_violations": [
       "스크린의 부검 사진이 좌측 벽면 액자에도 중복되어 나타남 (Leaked marker / Invented object)"
      ],
      "physics": "부검의는 증인석 뒤에 자리 잡고 있으며 책상에 의해 하단이 자연스럽게 가려져 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "요구된 오버더숄더 구도와 증인의 곤혹스러운 표정을 잘 구현했으며, 법정 배경의 일관성과 한글 표기('증인석')도 정확히 묘사하여 가장 우수합니다."
     },
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "요구된 구도를 시도했으나, 왼쪽 벽에 프롬프트의 부검 사진이 그대로 복제된 정체불명의 액자가 생성되는 치명적인 오류(Hard Violation)가 발생해 탈락입니다.  ★위반: [gemini-pro] 왼쪽 벽에 부검 사진이 복제된 액자가 잘못 생성됨 (leaked markers/invented objects)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.25,
      "B": 1.5
     },
     "adjusted": {
      "A": 1.0,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[gemini-pro] 왼쪽 벽에 부검 사진이 복제된 액자가 잘못 생성됨 (leaked markers/invented objects)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "요구된 오버더숄더 구도와 증인의 곤혹스러운 표정을 잘 구현했으며, 법정 배경의 일관성과 한글 표기('증인석')도 정확히 묘사하여 가장 우수합니다."
     },
     {
      "label": "B",
      "score": 1000,
      "verdict_ko": "요구된 구도를 시도했으나, 왼쪽 벽에 프롬프트의 부검 사진이 그대로 복제된 정체불명의 액자가 생성되는 치명적인 오류(Hard Violation)가 발생해 탈락입니다.  ★위반: [gemini-pro] 왼쪽 벽에 부검 사진이 복제된 액자가 잘못 생성됨 (leaked markers/invented objects)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1503,
     "B": 1006
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S73sh6"
  }
 },
 "S73sh9::cine": {
  "applied": true,
  "fingerprint": "ac50d2b2c7b00518d9efac515df50a49f3a6e3952455dfc4864713debb866194",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S73sh9_sel.png",
  "source_sha256": "32a76ebce667f8e653078290cf51eda0f0a1d46cba34d243bc2f95e0cb6cdfad",
  "file": "S73sh9_cine.png",
  "latency_ms": 15724
 },
 "S74sh3::signage": {
  "fp": "f7aa93fbb6d86f26",
  "inscriptions": []
 },
 "S74sh3": {
  "input_fingerprint": "b60cbe982b83f0bd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고원효 쪽을 향해 서늘하고 살벌한 눈빛을 쏘아보는 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height and slightly below 지국현’s gaze, tighten on his face from the witness-side three-quarter angle, keeping him in sharp focus near center-right. 고원효 remains only as a soft witness-side foreground presence at the left edge, while 지국현’s rigid jaw and fixed eyes drive across the shared space toward him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 고원효 in the middle-left of the frame, foreground, looks toward 지국현; 지국현 in the middle-center of the frame, midground, looks toward 고원효.\n- KEY BACKGROUND ELEMENTS: witness stand (Occupied by 고원효) — The side facing the defendant area is visible at the edge of frame; used as Its partial edge establishes the witness-side axis behind 고원효’s foreground shoulder.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is subdued and neutral-to-cool, holding detail in 지국현’s fixed expression without melodramatic shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, screen, and neutral daylight. Exclude the medical examiner and prosecutor from the close frame, isolating the defendant's hostile stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고원효 쪽을 향해 서늘하고 살벌한 눈빛을 쏘아보는 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height and slightly below 지국현’s gaze, tighten on his face from the witness-side three-quarter angle, keeping him in sharp focus near center-right. 고원효 remains only as a soft witness-side foreground presence at the left edge, while 지국현’s rigid jaw and fixed eyes drive across the shared space toward him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 고원효 in the middle-left of the frame, foreground, looks toward 지국현; 지국현 in the middle-center of the frame, midground, looks toward 고원효.\n- KEY BACKGROUND ELEMENTS: witness stand (Occupied by 고원효) — The side facing the defendant area is visible at the edge of frame; used as Its partial edge establishes the witness-side axis behind 고원효’s foreground shoulder.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is subdued and neutral-to-cool, holding detail in 지국현’s fixed expression without melodramatic shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, screen, and neutral daylight. Exclude the medical examiner and prosecutor from the close frame, isolating the defendant's hostile stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고원효 쪽을 향해 서늘하고 살벌한 눈빛을 쏘아보는 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height and slightly below 지국현’s gaze, tighten on his face from the witness-side three-quarter angle, keeping him in sharp focus near center-right. 고원효 remains only as a soft witness-side foreground presence at the left edge, while 지국현’s rigid jaw and fixed eyes drive across the shared space toward him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 고원효 in the middle-left of the frame, foreground, looks toward 지국현; 지국현 in the middle-center of the frame, midground, looks toward 고원효.\n- KEY BACKGROUND ELEMENTS: witness stand (Occupied by 고원효) — The side facing the defendant area is visible at the edge of frame; used as Its partial edge establishes the witness-side axis behind 고원효’s foreground shoulder.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is subdued and neutral-to-cool, holding detail in 지국현’s fixed expression without melodramatic shadow.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, screen, and neutral daylight. Exclude the medical examiner and prosecutor from the close frame, isolating the defendant's hostile stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "화면 오른쪽의 양복을 입은 남성은 왼쪽을 바라보고 있으며, 화면 왼쪽의 죄수복을 입은 남성(지국현)은 오른쪽을 바라봄.",
    "built_space": "법정 내부. 뒤편에 태극기와 부검 사진이 띄워진 스크린이 보임.",
    "entities": "왼쪽 전경에 흐리게 처리된 지국현(죄수복 1234)이 있고, 오른쪽에 이전 컷 부검의의 얼굴을 한 남성이 선명하게 렌더링됨.",
    "hard_violations": [
     "프레이밍 및 구도 지시 위반 (지국현이 중앙 우측에서 선명하게 보여야 하나 왼쪽에서 흐리게 렌더링됨)",
     "이전 컷 사진의 인물 얼굴을 재사용함 (금지된 사항)"
    ],
    "physics": "두 인물 모두 안정적인 자세로 위치해 있으며 물리적 지지에 이상 없음."
   },
   {
    "label": "B",
    "direction": "화면 중앙 우측의 지국현이 왼쪽 가장자리에 위치한 전경의 인물을 향해 날카로운 시선을 쏘아봄.",
    "built_space": "법정 내부. 뒤편에 태극기와 스크린, 오른쪽 아래에 증인석 명패가 올바르게 위치함.",
    "entities": "중앙 우측에 지국현이 캐릭터 레퍼런스와 정확히 일치하는 외모와 죄수복('교도소 1234') 차림으로 선명하게 나타남. 왼쪽 전경에 검은 법복을 입은 인물의 뒷모습이 흐리게 배치됨.",
    "hard_violations": [],
    "physics": "지국현의 자세와 어깨선이 자연스러우며 물리적인 오류 없이 공간에 잘 자리잡고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 9,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 9,
    "verdict_ko": "지국현을 화면 우측 중앙에 선명하게 배치하고 왼쪽 전경의 인물을 향해 시선을 던지도록 하여, 요구된 구도와 캐릭터 레퍼런스를 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "초점과 위치가 반대로 설정되어 지국현이 흐리게 처리되었으며, 이전 컷의 인물 얼굴을 그대로 가져오는 치명적인 오류를 범했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S73sh9_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라가 증인석 쪽에서 피고인석에 있는 지국현을 바라보는 구도여야 하나, 지국현의 배경에 스크린, 태극기, 증인석('인석' 표지)이 배치되어 그가 피고인석이 아닌 증인석 위치에 있는 것처럼 공간이 잘못 구성됨.",
     "fix_en": "Defocus background flags and screen to deemphasize incorrect arrangement. Preserve character poses, faces, and lighting.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이전 컷 인물의 의상을 절대 가져오지 말라는 지시(never carry their clothing)에도 불구하고, 왼쪽 전경 인물이 이전 컷의 인물과 동일한 주름진 검은색 법복을 입고 있음.",
     "fix_en": "Replace foreground black robe with a standard dark suit jacket. Preserve character positions, Ji Guk-hyun's face, background, and lighting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "지국현의 죄수복 명찰('교도소 1234')이 캐릭터 레퍼런스(가슴 상단)와 달리 복부에 가까운 맨 아래쪽으로 지나치게 낮게 부착되어 있음.",
     "fix_en": "Move name tag from bottom of shirt to upper left chest. Preserve face, pose, foreground, and background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "배경에 있는 태극기의 검은색 괘가 정상적인 선 형태가 아닌 형태를 알아볼 수 없는 덩어리로 뭉개져 있음.",
     "fix_en": "Redraw flag trigrams as distinct parallel bars. Preserve character, screen, and lighting.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "지국현 얼굴 클로즈업이 아니라 상반신과 법정이 넓게 보이는 미디엄 투샷이다.",
     "fix_en": "Darken lower torso and background edges to simulate tighter framing. Preserve face, foreground, and lighting.",
     "severity": "critical",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라가 증인석 쪽에서 피고인석에 있는 지국현을 바라보는 구도여야 하나, 지국현의 배경에 스크린, 태극기, 증인석('인석' 표지)이 배치되어 그가 피고인석이 아닌 증인석 위치에 있는 것처럼 공간이 잘못 구성됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 컷 인물의 의상을 절대 가져오지 말라는 지시(never carry their clothing)에도 불구하고, 왼쪽 전경 인물이 이전 컷의 인물과 동일한 주름진 검은색 법복을 입고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "지국현의 죄수복 명찰('교도소 1234')이 캐릭터 레퍼런스(가슴 상단)와 달리 복부에 가까운 맨 아래쪽으로 지나치게 낮게 부착되어 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "배경에 있는 태극기의 검은색 괘가 정상적인 선 형태가 아닌 형태를 알아볼 수 없는 덩어리로 뭉개져 있음.",
     "severity": "minor"
    },
    {
     "issue_ko": "지국현 얼굴 클로즈업이 아니라 상반신과 법정이 넓게 보이는 미디엄 투샷이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "지국현이 피고인석이 아니라 증인석·스크린 앞 중앙에 서 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 전경이 증인석의 고원효가 아니라 이전 샷과 같은 검은 법복 인물의 등이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "증인석이 고원효 어깨 뒤 왼쪽 가장자리가 아니라 화면 오른쪽 배경에 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "지국현의 죄수복 명찰('교도소 1234')이 캐릭터 레퍼런스(가슴 상단)와 달리 복부에 가까운 맨 아래쪽으로 지나치게 낮게 부착되어 있음.",
    "fix_en": "Move name tag from bottom of shirt to upper left chest. Preserve face, pose, foreground, and background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "배경에 있는 태극기의 검은색 괘가 정상적인 선 형태가 아닌 형태를 알아볼 수 없는 덩어리로 뭉개져 있음.",
    "fix_en": "Redraw flag trigrams as distinct parallel bars. Preserve character, screen, and lighting.",
    "severity": "minor",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Defocus background flags and screen to deemphasize incorrect arrangement. Preserve character poses, faces, and lighting.\n- Replace foreground black robe with a standard dark suit jacket. Preserve character positions, Ji Guk-hyun's face, background, and lighting.\n- Darken lower torso and background edges to simulate tighter framing. Preserve face, foreground, and lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 600,
      "verdict_ko": "전경 인물에 이전 샷의 법복이 남아있는 것은 부정 지시 위반 소지가 있으나, 지시된 중립적이고 자연스러운 조명과 시네마틱한 분위기를 완벽히 구현하여 우수합니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 피고인석이 아닌 증인석 앞에 배치됨 / [openrouter:x-ai/grok-4.6] 고원효가 증인석을 점유하지 않고 검사 쪽 전경에 있음 / [openrouter:x-ai/grok-4.6] 제외하라던 이전 스틸 검사의 법복을 왼쪽 인물에게 입힘 / [openrouter:x-ai/grok-4.6] 증인석 측 3/4·얼굴 클로즈업이 아니라 이전 샷 프레이밍 복제"
     },
     {
      "label": "B",
      "score": 964,
      "verdict_ko": "전경 인물의 의상을 변경하여 연속성 오류는 피했으나, 하단에 인위적인 검은색 그라데이션을 덮어씌워 '과장된 그림자 없는 조명' 지시를 심각하게 훼손했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 피고인석이 아닌 증인석 앞에 배치됨 / [openrouter:x-ai/grok-4.6] 고원효가 증인석을 점유하지 않음 / [openrouter:x-ai/grok-4.6] 지정된 증인석 측 착석 눈높이 얼굴 클로즈업이 아닌 이전 샷 구도 복제"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.6,
      "B": 1.714
     },
     "adjusted": {
      "A": 0.6,
      "B": 0.964
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 지국현이 피고인석이 아닌 증인석 앞에 배치됨",
       "[openrouter:x-ai/grok-4.6] 고원효가 증인석을 점유하지 않고 검사 쪽 전경에 있음",
       "[openrouter:x-ai/grok-4.6] 제외하라던 이전 스틸 검사의 법복을 왼쪽 인물에게 입힘",
       "[openrouter:x-ai/grok-4.6] 증인석 측 3/4·얼굴 클로즈업이 아니라 이전 샷 프레이밍 복제"
      ],
      "B": [
       "[openrouter:x-ai/grok-4.6] 지국현이 피고인석이 아닌 증인석 앞에 배치됨",
       "[openrouter:x-ai/grok-4.6] 고원효가 증인석을 점유하지 않음",
       "[openrouter:x-ai/grok-4.6] 지정된 증인석 측 착석 눈높이 얼굴 클로즈업이 아닌 이전 샷 구도 복제"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.4,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 600,
      "verdict_ko": "전경 인물에 이전 샷의 법복이 남아있는 것은 부정 지시 위반 소지가 있으나, 지시된 중립적이고 자연스러운 조명과 시네마틱한 분위기를 완벽히 구현하여 우수합니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 피고인석이 아닌 증인석 앞에 배치됨 / [openrouter:x-ai/grok-4.6] 고원효가 증인석을 점유하지 않고 검사 쪽 전경에 있음 / [openrouter:x-ai/grok-4.6] 제외하라던 이전 스틸 검사의 법복을 왼쪽 인물에게 입힘 / [openrouter:x-ai/grok-4.6] 증인석 측 3/4·얼굴 클로즈업이 아니라 이전 샷 프레이밍 복제"
     },
     {
      "label": "B",
      "score": 964,
      "verdict_ko": "전경 인물의 의상을 변경하여 연속성 오류는 피했으나, 하단에 인위적인 검은색 그라데이션을 덮어씌워 '과장된 그림자 없는 조명' 지시를 심각하게 훼손했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 지국현이 피고인석이 아닌 증인석 앞에 배치됨 / [openrouter:x-ai/grok-4.6] 고원효가 증인석을 점유하지 않음 / [openrouter:x-ai/grok-4.6] 지정된 증인석 측 착석 눈높이 얼굴 클로즈업이 아닌 이전 샷 구도 복제"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "이전 컷의 의상을 배제하라는 지시를 준수해 전경 인물에 정장을 입혔으며, 캐릭터와 프레이밍 요구사항을 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "캐릭터 외모는 우수하나, 이전 컷 인물의 법복을 전경 인물에게 그대로 적용하여 부정 지시어를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "화면 우측의 지국현이 좌측 전경의 고원효를 정확히 노려봄.",
      "built_space": "배경에 부검 사진 스크린과 태극기, 우측 하단에 증인석이 올바르게 배치됨.",
      "entities": "지국현의 얼굴과 수의(1234)가 일치함. 전경 인물이 정장을 입어 이전 컷 의상 배제 지시를 충족함.",
      "hard_violations": [],
      "physics": "인물들의 자세가 자연스럽고 안정적으로 지지됨."
     },
     {
      "label": "B",
      "direction": "화면 우측의 지국현이 좌측 전경의 고원효를 정확히 노려봄.",
      "built_space": "배경에 부검 사진 스크린과 태극기, 우측 하단에 증인석이 올바르게 배치됨.",
      "entities": "지국현의 얼굴과 수의(1234)가 일치함. 단, 전경 인물이 이전 컷과 동일한 주름진 법복을 입어 지시를 위반함.",
      "hard_violations": [],
      "physics": "인물들의 자세가 자연스럽고 안정적으로 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 컷의 의상을 배제하라는 지시를 준수해 전경 인물에 정장을 입혔으며, 캐릭터와 프레이밍 요구사항을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "캐릭터 외모는 우수하나, 이전 컷 인물의 법복을 전경 인물에게 그대로 적용하여 부정 지시어를 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "화면 우측의 지국현이 좌측 전경의 고원효를 정확히 노려봄.",
      "built_space": "배경에 부검 사진 스크린과 태극기, 우측 하단에 증인석이 올바르게 배치됨.",
      "entities": "지국현의 얼굴과 수의(1234)가 일치함. 전경 인물이 정장을 입어 이전 컷 의상 배제 지시를 충족함.",
      "hard_violations": [],
      "physics": "인물들의 자세가 자연스럽고 안정적으로 지지됨."
     },
     {
      "label": "A",
      "direction": "화면 우측의 지국현이 좌측 전경의 고원효를 정확히 노려봄.",
      "built_space": "배경에 부검 사진 스크린과 태극기, 우측 하단에 증인석이 올바르게 배치됨.",
      "entities": "지국현의 얼굴과 수의(1234)가 일치함. 단, 전경 인물이 이전 컷과 동일한 주름진 법복을 입어 지시를 위반함.",
      "hard_violations": [],
      "physics": "인물들의 자세가 자연스럽고 안정적으로 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 606,
     "B": 972
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S73sh9"
  }
 },
 "S74sh3::cine": {
  "applied": true,
  "fingerprint": "7d11863a687f8c438de258bf7c469f8719b4910deca8aabb9b0329fc85d705fe",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S74sh3_sel.png",
  "source_sha256": "586fd8ed47c69c666be5a863b5ba2c19e0e54d8aeb6f878c65f5bd6ba2d4a654",
  "file": "S74sh3_cine.png",
  "latency_ms": 10109
 },
 "S74sh9::signage": {
  "fp": "fb2814e66727a68c",
  "inscriptions": [
   {
    "surface_native": "검사석 명패",
    "text_native": "검사",
    "reason_ko": "대한민국의 법정 내부 검사석 테이블 위에 놓인 직책 표시 명패로, 주인공 장원섭의 신분과 위치를 시각적으로 명확히 나타내기 위해 필요합니다."
   }
  ]
 },
 "S74sh9": {
  "input_fingerprint": "5734e26f582a0fe5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자리에서 벌떡 일어선 채 지국현의 변호인 쪽을 향해 단호하게 입을 벌린 장원섭의 상체\n\nLOCATION (lock): At the prosecution table inside the courtroom, turned toward the defense attorney and witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From beside 장원섭’s position at seated shoulder height, finish the upward tilt at the peak of his sudden rise and isolate his upper body on a low three-quarter line. 장원섭’s torso pitches toward 지국현의 변호인 at the opposite edge, his mouth open mid-objection and his eyes fixed on the attorney rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: prosecution table (장원섭 has just risen from his place beside it) — Its courtroom-facing edge runs beneath 장원섭’s rising torso; used as The partially visible prosecution position anchors the abrupt upward movement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime courtroom ambience preserves the procedural realism of the objection with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, screen, and formal lighting. Exclude the earlier autopsy-image emphasis and show the prosecutor standing to object.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 검사석 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자리에서 벌떡 일어선 채 지국현의 변호인 쪽을 향해 단호하게 입을 벌린 장원섭의 상체\n\nLOCATION (lock): At the prosecution table inside the courtroom, turned toward the defense attorney and witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From beside 장원섭’s position at seated shoulder height, finish the upward tilt at the peak of his sudden rise and isolate his upper body on a low three-quarter line. 장원섭’s torso pitches toward 지국현의 변호인 at the opposite edge, his mouth open mid-objection and his eyes fixed on the attorney rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: prosecution table (장원섭 has just risen from his place beside it) — Its courtroom-facing edge runs beneath 장원섭’s rising torso; used as The partially visible prosecution position anchors the abrupt upward movement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime courtroom ambience preserves the procedural realism of the objection with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, screen, and formal lighting. Exclude the earlier autopsy-image emphasis and show the prosecutor standing to object.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 검사석 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자리에서 벌떡 일어선 채 지국현의 변호인 쪽을 향해 단호하게 입을 벌린 장원섭의 상체\n\nLOCATION (lock): At the prosecution table inside the courtroom, turned toward the defense attorney and witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From beside 장원섭’s position at seated shoulder height, finish the upward tilt at the peak of his sudden rise and isolate his upper body on a low three-quarter line. 장원섭’s torso pitches toward 지국현의 변호인 at the opposite edge, his mouth open mid-objection and his eyes fixed on the attorney rather than the lens.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: prosecution table (장원섭 has just risen from his place beside it) — Its courtroom-facing edge runs beneath 장원섭’s rising torso; used as The partially visible prosecution position anchors the abrupt upward movement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral-to-cool daytime courtroom ambience preserves the procedural realism of the objection with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, screen, and formal lighting. Exclude the earlier autopsy-image emphasis and show the prosecutor standing to object.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 검사석 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선은 왼쪽 전경에 위치한 뒷모습의 인물을 향하며, 입을 크게 벌리고 소리치는 방향과 일치함.",
    "built_space": "검사석 뒤로 국기 2개와 텍스트가 띄워진 스크린이 보이며, '검사' 명패가 책상 위에 놓여 있음.",
    "entities": "장원섭과 더불어 텍스트가 지시하지 않은 남성의 뒷모습이 왼쪽 전경에 등장함. 스크린에 의미 불명의 문자가 있음.",
    "hard_violations": [
     "지시되지 않은 인물(프레임 왼쪽 전경의 남성) 임의 추가",
     "스크린에 생성된 의미 불명의 문자들(invented text)"
    ],
    "physics": "양손을 책상에 짚고 상체를 앞으로 굽히며 일어나는 동작이 책상을 통해 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "화면 왼쪽 바깥쪽을 강하게 응시하며 입을 벌리고 있어, 프레임 밖 상대방에게 항의하는 방향성이 뚜렷함.",
    "built_space": "검사석과 국기, 이전 컷의 부검 사진이 떠 있는 스크린이 적절한 심도로 배치됨. 벽면 재질이 레퍼런스와 다소 다름.",
    "entities": "지시된 대로 장원섭 한 명만 프레임에 존재하며, 지정된 복장과 명패의 '검사' 텍스트 모두 완벽히 일치함.",
    "hard_violations": [],
    "physics": "책상을 짚은 양손에 체중이 실려 있고, 상체를 앞으로 쑥 내민 기립 자세가 안정적으로 지탱됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 인물 단독 샷을 유지하고 자세와 시선을 정확히 연출하여 훌륭하게 완성함."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "지시되지 않은 인물을 전경에 추가하고 스크린에 임의의 문자를 생성한 점이 치명적인 감점 요인임."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S74sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:796050>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트에서 부검 사진이 강조되는 것을 배제하라고 명시했으나(Exclude the earlier autopsy-image emphasis), 배경 우측에 부검 사진이 띄워진 대형 스크린이 여전히 큰 비중으로 노출되어 있습니다.",
     "fix_en": "Change the content on the background screen from the graphic autopsy photo to a neutral text document or a blank turned-off monitor, while preserving the prosecutor, his expression, pose, clothing, the flags, the wooden paneling, the nameplate, and the camera framing.",
     "severity": "major",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트에서 부검 사진이 강조되는 것을 배제하라고 명시했으나(Exclude the earlier autopsy-image emphasis), 배경 우측에 부검 사진이 띄워진 대형 스크린이 여전히 큰 비중으로 노출되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "우측 배경 화면에 제외하라고 한 이전 샷의 부검 사진이 크게 강조되어 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트에서 부검 사진이 강조되는 것을 배제하라고 명시했으나(Exclude the earlier autopsy-image emphasis), 배경 우측에 부검 사진이 띄워진 대형 스크린이 여전히 큰 비중으로 노출되어 있습니다.",
    "fix_en": "Change the content on the background screen from the graphic autopsy photo to a neutral text document or a blank turned-off monitor, while preserving the prosecutor, his expression, pose, clothing, the flags, the wooden paneling, the nameplate, and the camera framing.",
    "severity": "major",
    "observation_index": 0
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S74sh3"
  }
 },
 "S74sh9::cine": {
  "applied": true,
  "fingerprint": "7651e39cc4767c3df2dc28721ab581404a4a0ec751127f95eb26aac14dcdfb2f",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S74sh9_sel.png",
  "source_sha256": "c0853a15d977b7461d9ff6937dfe1559f49292b85e70031bd66d7d7a6a2ad4d2",
  "file": "S74sh9_cine.png",
  "latency_ms": 18555
 },
 "S74sh12::signage": {
  "fp": "a56054b8d8e7c813",
  "inscriptions": [
   {
    "surface_native": "법정 벽면 안내판",
    "text_native": "정숙",
    "reason_ko": "법정 내 질서와 침묵을 요구하는 공식 표지판을 배경에 배치하여, 방청석에서 고함을 지르며 소란을 피우는 인물의 분노와 극적인 대비를 이루기 위함입니다."
   }
  ]
 },
 "S74sh12": {
  "input_fingerprint": "43ef75658a98342e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 방청석 자리에 벌떡 일어선 채 앞을 향해 삿대질하며 고함치는 신경희의 분노한 얼굴\n\nLOCATION (lock): In the courtroom public gallery amid the suddenly protesting supporters. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin the backward move close to 신경희 at the aisle-side edge of the public seating area, her angry face and thrusting shoulder held in tight three-quarter framing before the frame starts to breathe wider. She surges up from her seat and points toward the front of the courtroom while shouting; behind her, surrounding congregants are caught at varied phases of rising rather than in a uniform row.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 in the middle-center of the frame, foreground, points to front of courtroom; front of courtroom in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: public seating rows (Surrounding congregants are rising from their seats at different phases of the same outburst) — The rows extend diagonally behind 신경희 toward the rear of the courtroom; used as Frames 신경희’s sudden rise and provides staggered depth for the widening move; side aisle (Open for the backward camera move) — The aisle recedes backward from 신경희’s row along the camera’s withdrawal path; used as Creates the diagonal aisle-side camera path for the dolly-out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience stays neutral-to-cool and naturalistic, allowing the anger to register without heightened spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom gallery, rows of seats, subdued daylight, and the woman's clothing. Exclude the grieving mother from emphasis and show the supporter standing and shouting.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 벽면 안내판: \"정숙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 방청석 자리에 벌떡 일어선 채 앞을 향해 삿대질하며 고함치는 신경희의 분노한 얼굴\n\nLOCATION (lock): In the courtroom public gallery amid the suddenly protesting supporters. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin the backward move close to 신경희 at the aisle-side edge of the public seating area, her angry face and thrusting shoulder held in tight three-quarter framing before the frame starts to breathe wider. She surges up from her seat and points toward the front of the courtroom while shouting; behind her, surrounding congregants are caught at varied phases of rising rather than in a uniform row.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 in the middle-center of the frame, foreground, points to front of courtroom; front of courtroom in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: public seating rows (Surrounding congregants are rising from their seats at different phases of the same outburst) — The rows extend diagonally behind 신경희 toward the rear of the courtroom; used as Frames 신경희’s sudden rise and provides staggered depth for the widening move; side aisle (Open for the backward camera move) — The aisle recedes backward from 신경희’s row along the camera’s withdrawal path; used as Creates the diagonal aisle-side camera path for the dolly-out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience stays neutral-to-cool and naturalistic, allowing the anger to register without heightened spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom gallery, rows of seats, subdued daylight, and the woman's clothing. Exclude the grieving mother from emphasis and show the supporter standing and shouting.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 벽면 안내판: \"정숙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 방청석 자리에 벌떡 일어선 채 앞을 향해 삿대질하며 고함치는 신경희의 분노한 얼굴\n\nLOCATION (lock): In the courtroom public gallery amid the suddenly protesting supporters. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin the backward move close to 신경희 at the aisle-side edge of the public seating area, her angry face and thrusting shoulder held in tight three-quarter framing before the frame starts to breathe wider. She surges up from her seat and points toward the front of the courtroom while shouting; behind her, surrounding congregants are caught at varied phases of rising rather than in a uniform row.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 신경희 in the middle-center of the frame, foreground, points to front of courtroom; front of courtroom in the upper-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: public seating rows (Surrounding congregants are rising from their seats at different phases of the same outburst) — The rows extend diagonally behind 신경희 toward the rear of the courtroom; used as Frames 신경희’s sudden rise and provides staggered depth for the widening move; side aisle (Open for the backward camera move) — The aisle recedes backward from 신경희’s row along the camera’s withdrawal path; used as Creates the diagonal aisle-side camera path for the dolly-out.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience stays neutral-to-cool and naturalistic, allowing the anger to register without heightened spectacle.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom gallery, rows of seats, subdued daylight, and the woman's clothing. Exclude the grieving mother from emphasis and show the supporter standing and shouting.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 벽면 안내판: \"정숙\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "신경희가 법정 앞쪽(화면 좌측 상단)을 향해 오른손으로 삿대질하고 있음. 목표 방향이 정확함.",
    "built_space": "법정 방청석 구조가 정확하며, 우측에 뒤로 이어지는 통로가 열려 있음. 좌측 상단에 판사석과 모니터가 올바르게 위치함.",
    "entities": "신경희는 참조 이미지의 얼굴 및 의상과 일치하나 목걸이가 누락됨. 주변 방청객들은 일어서서 항의하는 지시와 달리 일제히 조용히 앉아 있음. 벽면 안내판 텍스트('정숙')는 잘 구현됨.",
    "hard_violations": [
     "프롬프트에서 명시적으로 제외하라고 지시한 이전 컷의 인물(법복을 입은 검사/판사)이 좌측 상단 배경에 그대로 등장함 (금지된 인물 및 의상 포함)."
    ],
    "physics": "신경희와 방청객 모두 물리적으로 지지되어 안정적인 자세를 취하고 있으며, 공중에 떠 있는 물체나 인물은 없음."
   },
   {
    "label": "B",
    "direction": "신경희가 법정 앞쪽(화면 좌측 상단)을 향해 오른손으로 삿대질하며 고함치고 있음. 방향 설정이 정확함.",
    "built_space": "법정 방청석의 사선 배열과 우측 통로가 잘 묘사됨. 좌측 상단에 판사석과 법정 구조가 올바르게 위치함.",
    "entities": "신경희의 얼굴, 의상, 목걸이가 참조 이미지와 완벽히 일치함. 주변 방청객들이 일어서서 주먹을 쥐고 항의하는 모습이 지시대로 잘 표현됨. 벽면 안내판('정숙')과 모니터 내용이 올바르게 나타남.",
    "hard_violations": [
     "프롬프트에서 명시적으로 제외하라고 지시한 이전 컷의 인물(법복과 흰색 넥타이를 착용한 검사/판사)이 좌측 상단 배경에 그대로 등장함 (금지된 인물 및 의상 포함)."
    ],
    "physics": "신경희의 서 있는 자세와 삿대질, 그리고 일어서는 방청객들의 동선 모두 물리적 지지가 명확하고 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 4,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "방청객의 항의 행동과 캐릭터의 디테일(목걸이)을 훌륭하게 구현했으나, 프레이밍이 클로즈업보다 넓고 프롬프트에서 명시적으로 금지한 이전 컷의 인물이 배경에 등장하는 치명적 오류가 있습니다."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "방청객들이 항의하며 일어서는 지시를 완전히 무시하고 얌전히 앉아있으며, 명시적으로 배제된 이전 컷의 인물이 배경에 그대로 등장하는 치명적 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S74sh9_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:843594>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "좌측 상단 배경에 이전 레퍼런스 이미지에 있던 인물(검사)이 동일한 복장으로 등장하여, 이전 사진의 인물을 이 샷에 포함하지 말라는 지시를 위반함.",
     "fix_en": "Remove the man in black robes from the desk in the top-left background, leaving only the empty wooden desk and the wall with flags behind it. Preserve Shin Kyung-hee's exact pose, face, expression, clothing, the background crowd, the screen, the lighting, the framing, and the '정숙' sign.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "지정된 얼굴 클로즈업이 아니라 상반신과 법정 전체가 보이는 중원경이다",
     "fix_en": "Push the camera in to a tight close-up on the screaming woman's face and raised shoulder, cropping out her lower torso and the wider room. Preserve her face, expression, clothing, and the lighting.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "좌측 상단 배경에 이전 레퍼런스 이미지에 있던 인물(검사)이 동일한 복장으로 등장하여, 이전 사진의 인물을 이 샷에 포함하지 말라는 지시를 위반함.",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 스틸의 남성 검사가 좌상단 재판석에 같은 얼굴과 법복으로 등장한다",
     "severity": "critical"
    },
    {
     "issue_ko": "지정된 얼굴 클로즈업이 아니라 상반신과 법정 전체가 보이는 중원경이다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "지정된 얼굴 클로즈업이 아니라 상반신과 법정 전체가 보이는 중원경이다",
    "fix_en": "Push the camera in to a tight close-up on the screaming woman's face and raised shoulder, cropping out her lower torso and the wider room. Preserve her face, expression, clothing, and the lighting.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the man in black robes from the desk in the top-left background, leaving only the empty wooden desk and the wall with flags behind it. Preserve Shin Kyung-hee's exact pose, face, expression, clothing, the background crowd, the screen, the lighting, the framing, and the '정숙' sign.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "이전 샷의 인물(검사)을 화면에서 성공적으로 제거하여 명시적 제외 지시를 완벽히 준수했으며, 요구된 구도와 캐릭터의 분노한 감정선이 매우 잘 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "캐릭터와 구도는 훌륭하나, 이전 샷 스틸에 등장했던 인물(검사)을 포함하지 말라는 명시적 제약 조건을 위반하고 좌측 상단에 그대로 노출시켰습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "신경희가 화면 좌측 상단의 법정 전면을 향해 팔을 뻗어 삿대질하고 있으며, 시선과 고함치는 방향이 이에 일치함.",
      "built_space": "대각선으로 배치된 방청석 의자들과 뒤로 이어지는 측면 통로. 좌측 상단에 법대, 스크린, 국기가 위치함.",
      "entities": "신경희의 얼굴, 헤어스타일, 의상은 참조와 일치함. 벽면에 '정숙' 텍스트가 정확히 구현됨. 그러나 이전 샷의 검사가 법대에 서 있음.",
      "hard_violations": [
       "이전 샷 스틸에 등장한 인물(검사)을 배제하라는 지시 위반 및 엑스트라 추가 금지 위반"
      ],
      "physics": "신경희는 두 다리로 바닥을 지지하며 앞으로 기울인 역동적인 자세를 취했고, 배경의 방청객들도 각자의 자리에서 자연스럽게 일어나는 물리적 상태를 유지함."
     },
     {
      "label": "B",
      "direction": "신경희가 좌측 상단의 법정 앞쪽을 향해 삿대질을 하고 시선을 고정한 채 고함침.",
      "built_space": "방청석 좌석 배열과 통로가 지시된 대각선 깊이를 형성함. 좌측 상단의 법대는 인물 없이 비어 있음.",
      "entities": "신경희의 외모와 의상, 장신구가 참조 이미지와 일치함. 배경 방청객들이 지시에 맞게 등장했으며, 벽면의 '정숙' 텍스트가 명확함.",
      "hard_violations": [],
      "physics": "서서 항의하는 신경희의 자세가 안정적으로 지탱되고 있으며, 주변 방청객들의 일어나는 동작도 의자와 바닥에 물리적으로 올바르게 접촉되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "이전 샷의 인물(검사)을 화면에서 성공적으로 제거하여 명시적 제외 지시를 완벽히 준수했으며, 요구된 구도와 캐릭터의 분노한 감정선이 매우 잘 표현되었습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "캐릭터와 구도는 훌륭하나, 이전 샷 스틸에 등장했던 인물(검사)을 포함하지 말라는 명시적 제약 조건을 위반하고 좌측 상단에 그대로 노출시켰습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "신경희가 화면 좌측 상단의 법정 전면을 향해 팔을 뻗어 삿대질하고 있으며, 시선과 고함치는 방향이 이에 일치함.",
      "built_space": "대각선으로 배치된 방청석 의자들과 뒤로 이어지는 측면 통로. 좌측 상단에 법대, 스크린, 국기가 위치함.",
      "entities": "신경희의 얼굴, 헤어스타일, 의상은 참조와 일치함. 벽면에 '정숙' 텍스트가 정확히 구현됨. 그러나 이전 샷의 검사가 법대에 서 있음.",
      "hard_violations": [
       "이전 샷 스틸에 등장한 인물(검사)을 배제하라는 지시 위반 및 엑스트라 추가 금지 위반"
      ],
      "physics": "신경희는 두 다리로 바닥을 지지하며 앞으로 기울인 역동적인 자세를 취했고, 배경의 방청객들도 각자의 자리에서 자연스럽게 일어나는 물리적 상태를 유지함."
     },
     {
      "label": "B",
      "direction": "신경희가 좌측 상단의 법정 앞쪽을 향해 삿대질을 하고 시선을 고정한 채 고함침.",
      "built_space": "방청석 좌석 배열과 통로가 지시된 대각선 깊이를 형성함. 좌측 상단의 법대는 인물 없이 비어 있음.",
      "entities": "신경희의 외모와 의상, 장신구가 참조 이미지와 일치함. 배경 방청객들이 지시에 맞게 등장했으며, 벽면의 '정숙' 텍스트가 명확함.",
      "hard_violations": [],
      "physics": "서서 항의하는 신경희의 자세가 안정적으로 지탱되고 있으며, 주변 방청객들의 일어나는 동작도 의자와 바닥에 물리적으로 올바르게 접촉되어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "이전 샷의 인물을 정확히 배제하고 레퍼런스와 일치하는 신경희의 분노한 모습과 삿대질을 명시된 구도에 맞게 잘 구현함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "이전 샷 스틸에 등장한 인물을 화면에 포함하지 말라는 명시적인 제한 지침을 어기고 배경에 검사를 등장시켜 감점됨."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "신경희가 화면 좌측 상단의 법정 앞부분을 향해 손가락을 가리키며 뚜렷한 시선을 던지고 있음.",
      "built_space": "방청석 의자들이 대각선으로 배열되어 깊이감을 주며, 좌측 상단 배경에 스크린과 국기가 있는 법정 전면이 보임.",
      "entities": "신경희의 얼굴, 헤어스타일, 의상이 캐릭터 레퍼런스와 정확히 일치함. 벽면에 '정숙' 안내판이 제대로 표기되었으며, 이전 샷의 인물은 등장하지 않음.",
      "hard_violations": [],
      "physics": "신경희와 배경의 군중들이 바닥이나 의자에 안정적으로 지탱된 채 일어서거나 앉아 있는 자세를 자연스럽게 유지함."
     },
     {
      "label": "B",
      "direction": "신경희가 좌측 상단 법정 앞쪽을 향해 삿대질하며 강한 시선으로 고함치고 있음.",
      "built_space": "대각선으로 배치된 방청석과 뒤로 이어지는 통로, 상단 좌측의 법정 전면부가 지침대로 구성됨.",
      "entities": "신경희의 외모와 복장이 레퍼런스와 부합하고 '정숙' 안내판이 존재함. 그러나 배경 좌측 상단에 이전 샷의 검사가 그대로 나타남.",
      "hard_violations": [
       "이전 샷 스틸에 등장했던 인물(검사)을 배경에 포함하여 '절대 이전 샷의 인물을 포함하지 말 것'이라는 지침을 위반함."
      ],
      "physics": "모든 인물이 중력에 맞게 바닥과 의자에 물리적으로 올바르게 접촉하여 동작을 취하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "이전 샷의 인물을 정확히 배제하고 레퍼런스와 일치하는 신경희의 분노한 모습과 삿대질을 명시된 구도에 맞게 잘 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "이전 샷 스틸에 등장한 인물을 화면에 포함하지 말라는 명시적인 제한 지침을 어기고 배경에 검사를 등장시켜 감점됨."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "신경희가 화면 좌측 상단의 법정 앞부분을 향해 손가락을 가리키며 뚜렷한 시선을 던지고 있음.",
      "built_space": "방청석 의자들이 대각선으로 배열되어 깊이감을 주며, 좌측 상단 배경에 스크린과 국기가 있는 법정 전면이 보임.",
      "entities": "신경희의 얼굴, 헤어스타일, 의상이 캐릭터 레퍼런스와 정확히 일치함. 벽면에 '정숙' 안내판이 제대로 표기되었으며, 이전 샷의 인물은 등장하지 않음.",
      "hard_violations": [],
      "physics": "신경희와 배경의 군중들이 바닥이나 의자에 안정적으로 지탱된 채 일어서거나 앉아 있는 자세를 자연스럽게 유지함."
     },
     {
      "label": "A",
      "direction": "신경희가 좌측 상단 법정 앞쪽을 향해 삿대질하며 강한 시선으로 고함치고 있음.",
      "built_space": "대각선으로 배치된 방청석과 뒤로 이어지는 통로, 상단 좌측의 법정 전면부가 지침대로 구성됨.",
      "entities": "신경희의 외모와 복장이 레퍼런스와 부합하고 '정숙' 안내판이 존재함. 그러나 배경 좌측 상단에 이전 샷의 검사가 그대로 나타남.",
      "hard_violations": [
       "이전 샷 스틸에 등장했던 인물(검사)을 배경에 포함하여 '절대 이전 샷의 인물을 포함하지 말 것'이라는 지침을 위반함."
      ],
      "physics": "모든 인물이 중력에 맞게 바닥과 의자에 물리적으로 올바르게 접촉하여 동작을 취하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S74sh9"
  }
 },
 "S74sh12::cine": {
  "applied": true,
  "fingerprint": "e1a9f3e80444f337548ae436efbc4207cf4bbdd4b8adf07890e2de6f2a0014ff",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S74sh12_sel.png",
  "source_sha256": "d6629e097397d6365d0bad9e67be8253fbbc3921b4410857e2e5e3d123c01f31",
  "file": "S74sh12_cine.png",
  "latency_ms": 16520
 },
 "S75sh4::signage": {
  "fp": "3fce283a65d16523",
  "inscriptions": []
 },
 "S75sh4": {
  "input_fingerprint": "2accbb95671d31e7",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린을 올려다보며 눈을 크게 뜬 채 당황한 기색이 역력한 장원섭의 얼굴 클로즈업.\n\nLOCATION (lock): At the prosecution side of the courtroom, looking up toward the large evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the screen side and slightly above seated eye height, hold a close downward three-quarter view of 장원섭’s face, positioned off-center with look room above his lifted gaze. He remains seated with his chin sharply raised toward the screen, eyes widened and mouth held in a visibly unsettled reaction.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying the dated group photograph) — The display face angles away above him and carries the alibi photograph; used as A partial, softly rendered edge above 장원섭 motivates his upward eyeline without competing with his face; public seating (Audience members are reacting to the testimony) — The seat backs recede behind 장원섭 from the screen-side angle; used as Keeps the reaction located within the courtroom audience area.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is restrained and neutral-to-cool, preserving natural facial detail in 장원섭’s startled reaction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness area, and formal lighting. Exclude the prior witness confrontation and isolate the prosecutor reacting to the displayed photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dated alibi photograph remains projected on the courtroom's large screen while Wonseop reacts to it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린을 올려다보며 눈을 크게 뜬 채 당황한 기색이 역력한 장원섭의 얼굴 클로즈업.\n\nLOCATION (lock): At the prosecution side of the courtroom, looking up toward the large evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the screen side and slightly above seated eye height, hold a close downward three-quarter view of 장원섭’s face, positioned off-center with look room above his lifted gaze. He remains seated with his chin sharply raised toward the screen, eyes widened and mouth held in a visibly unsettled reaction.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying the dated group photograph) — The display face angles away above him and carries the alibi photograph; used as A partial, softly rendered edge above 장원섭 motivates his upward eyeline without competing with his face; public seating (Audience members are reacting to the testimony) — The seat backs recede behind 장원섭 from the screen-side angle; used as Keeps the reaction located within the courtroom audience area.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is restrained and neutral-to-cool, preserving natural facial detail in 장원섭’s startled reaction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness area, and formal lighting. Exclude the prior witness confrontation and isolate the prosecutor reacting to the displayed photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dated alibi photograph remains projected on the courtroom's large screen while Wonseop reacts to it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린을 올려다보며 눈을 크게 뜬 채 당황한 기색이 역력한 장원섭의 얼굴 클로즈업.\n\nLOCATION (lock): At the prosecution side of the courtroom, looking up toward the large evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the screen side and slightly above seated eye height, hold a close downward three-quarter view of 장원섭’s face, positioned off-center with look room above his lifted gaze. He remains seated with his chin sharply raised toward the screen, eyes widened and mouth held in a visibly unsettled reaction.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying the dated group photograph) — The display face angles away above him and carries the alibi photograph; used as A partial, softly rendered edge above 장원섭 motivates his upward eyeline without competing with his face; public seating (Audience members are reacting to the testimony) — The seat backs recede behind 장원섭 from the screen-side angle; used as Keeps the reaction located within the courtroom audience area.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience is restrained and neutral-to-cool, preserving natural facial detail in 장원섭’s startled reaction.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness area, and formal lighting. Exclude the prior witness confrontation and isolate the prosecutor reacting to the displayed photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The dated alibi photograph remains projected on the courtroom's large screen while Wonseop reacts to it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "시선이 위쪽 스크린을 향함.",
    "built_space": "법정 내부 방청석 및 뒷배경의 스크린이 보이며 앵글이 약간 위에서 내려다보는 구도임.",
    "entities": "장원섭의 외모와 법복이 레퍼런스와 일치하며, 스크린에는 단체 사진이 띄워져 있음. 배경의 방청객들은 이전 샷과 다름.",
    "hard_violations": [],
    "physics": "자연스럽게 좌석에 앉아있는 자세임."
   },
   {
    "label": "A",
    "direction": "시선이 위쪽 스크린을 향함.",
    "built_space": "법정 내부 구조와 스크린, 국기가 보이며 레퍼런스와 동일한 공간 배치임.",
    "entities": "장원섭의 외모와 복장은 일치하나, 배경에 이전 샷(레퍼런스)에 등장했던 방청객들이 동일한 의상과 자세(주먹을 쥔 모습)로 그대로 등장함.",
    "hard_violations": [
     "이전 샷의 인물들을 포함하지 말라는 지시를 어기고 레퍼런스 이미지의 배경 인물과 포즈를 그대로 복사함."
    ],
    "physics": "의자에 안정적으로 앉아있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "클로즈업 프레이밍과 카메라 앵글을 정확히 구현했으며, 이전 샷의 인물들을 배제하라는 지시를 잘 따랐습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "주인공의 외모는 일치하나, 프레이밍이 너무 넓고 이전 샷의 배경 인물들을 그대로 복사하여 지침을 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S74sh12_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:796050>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "지문(Shot Text)에 명시되지 않은 여러 명의 인물들(방청객)이 배경에 임의로 추가되었습니다.",
     "fix_en": "Remove all background people, replacing them with empty wooden benches. Preserve Wonseop's face, pose, clothing, the courtroom interior, and lighting.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "카메라가 피사체의 눈높이보다 약간 위에서 내려다보는(downward) 앵글이어야 하지만, 실제로는 피사체의 턱 아래쪽이 보일 정도로 아래에서 올려다보는 앵글로 렌더링되었습니다.",
     "fix_en": "Tilt Wonseop's head slightly down. Preserve character, clothing, and background.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 가슴 위 상반신과 법정 배경이 넓게 담긴 미디엄 클로즈업이다.",
     "fix_en": "Crop tightly around Wonseop's face. Preserve expression and background.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "증거 스크린이 얼굴 위쪽의 흐릿한 가장자리가 아니라 왼쪽 배경에 선명하고 크게 보인다.",
     "fix_en": "Heavily blur the background screen. Preserve Wonseop and lighting.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "지문(Shot Text)에 명시되지 않은 여러 명의 인물들(방청객)이 배경에 임의로 추가되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "카메라가 피사체의 눈높이보다 약간 위에서 내려다보는(downward) 앵글이어야 하지만, 실제로는 피사체의 턱 아래쪽이 보일 정도로 아래에서 올려다보는 앵글로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 가슴 위 상반신과 법정 배경이 넓게 담긴 미디엄 클로즈업이다.",
     "severity": "major"
    },
    {
     "issue_ko": "증거 스크린이 얼굴 위쪽의 흐릿한 가장자리가 아니라 왼쪽 배경에 선명하고 크게 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "카메라가 피사체의 눈높이보다 약간 위에서 내려다보는(downward) 앵글이어야 하지만, 실제로는 피사체의 턱 아래쪽이 보일 정도로 아래에서 올려다보는 앵글로 렌더링되었습니다.",
    "fix_en": "Tilt Wonseop's head slightly down. Preserve character, clothing, and background.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "얼굴 클로즈업이 아니라 가슴 위 상반신과 법정 배경이 넓게 담긴 미디엄 클로즈업이다.",
    "fix_en": "Crop tightly around Wonseop's face. Preserve expression and background.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "증거 스크린이 얼굴 위쪽의 흐릿한 가장자리가 아니라 왼쪽 배경에 선명하고 크게 보인다.",
    "fix_en": "Heavily blur the background screen. Preserve Wonseop and lighting.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove all background people, replacing them with empty wooden benches. Preserve Wonseop's face, pose, clothing, the courtroom interior, and lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "명시되지 않은 인물을 배제하라는 엄격한 지침을 완벽히 따랐으며, 주인공의 당황한 표정과 시선, 지정된 의상을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "주인공의 묘사와 시선 처리는 우수하나, 샷 텍스트에 없는 인물을 추가하지 말라는 지침을 위반하고 배경에 방청객을 생성하여 치명적인 감점 요인이 되었습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "장원섭의 시선이 왼쪽 상단에 위치한 스크린을 명확히 향하고 있습니다.",
      "built_space": "법정 내부의 빈 방청석과 왼쪽 상단의 대형 스크린이 카메라 앵글에 맞게 배치되어 있습니다.",
      "entities": "장원섭의 얼굴, 갸름한 형태, 짧은 머리, 입고 있는 법복이 레퍼런스와 정확히 일치합니다.",
      "hard_violations": [],
      "physics": "의자에 앉아 고개를 위로 든 자세와 무게 중심이 물리적으로 타당하게 묘사되었습니다."
     },
     {
      "label": "A",
      "direction": "장원섭의 시선이 왼쪽 상단에 위치한 스크린을 명확히 향하고 있습니다.",
      "built_space": "법정 내부의 방청석과 왼쪽 상단의 대형 스크린이 카메라 앵글에 맞게 배치되어 있습니다.",
      "entities": "장원섭의 얼굴, 갸름한 형태, 짧은 머리, 입고 있는 법복이 레퍼런스와 정확히 일치합니다.",
      "hard_violations": [
       "발명된 인물 (샷 텍스트에 없는 배경 방청객들이 다수 추가됨)"
      ],
      "physics": "의자에 앉아 고개를 위로 든 자세와 무게 중심이 물리적으로 타당하게 묘사되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "명시되지 않은 인물을 배제하라는 엄격한 지침을 완벽히 따랐으며, 주인공의 당황한 표정과 시선, 지정된 의상을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "주인공의 묘사와 시선 처리는 우수하나, 샷 텍스트에 없는 인물을 추가하지 말라는 지침을 위반하고 배경에 방청객을 생성하여 치명적인 감점 요인이 되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "장원섭의 시선이 왼쪽 상단에 위치한 스크린을 명확히 향하고 있습니다.",
      "built_space": "법정 내부의 빈 방청석과 왼쪽 상단의 대형 스크린이 카메라 앵글에 맞게 배치되어 있습니다.",
      "entities": "장원섭의 얼굴, 갸름한 형태, 짧은 머리, 입고 있는 법복이 레퍼런스와 정확히 일치합니다.",
      "hard_violations": [],
      "physics": "의자에 앉아 고개를 위로 든 자세와 무게 중심이 물리적으로 타당하게 묘사되었습니다."
     },
     {
      "label": "A",
      "direction": "장원섭의 시선이 왼쪽 상단에 위치한 스크린을 명확히 향하고 있습니다.",
      "built_space": "법정 내부의 방청석과 왼쪽 상단의 대형 스크린이 카메라 앵글에 맞게 배치되어 있습니다.",
      "entities": "장원섭의 얼굴, 갸름한 형태, 짧은 머리, 입고 있는 법복이 레퍼런스와 정확히 일치합니다.",
      "hard_violations": [
       "발명된 인물 (샷 텍스트에 없는 배경 방청객들이 다수 추가됨)"
      ],
      "physics": "의자에 앉아 고개를 위로 든 자세와 무게 중심이 물리적으로 타당하게 묘사되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "시선 방향과 카메라 위치가 지시문과 어긋나지만, 허용되지 않은 인물을 추가하지 않아 상대적으로 낫습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "시선과 카메라 위치 오류에 더해, 명시되지 않은 방청객을 추가하여 인물 제한 규칙을 심각하게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선이 화면 밖 왼쪽 위를 향하나, 목표물인 스크린은 우측 배경에 있어 시선이 닿지 않음.",
      "built_space": "법정 내부. 카메라가 스크린 쪽에서 인물을 향해야 하나, 반대로 방청석에서 스크린을 바라보는 구도로 잘못 배치됨.",
      "entities": "장원섭의 외모와 복장은 레퍼런스와 일치. 스크린에 단체 사진이 있으며, 추가 인물은 없음.",
      "hard_violations": [],
      "physics": "의자에 앉아 체중이 정상적으로 지탱된 자세."
     },
     {
      "label": "B",
      "direction": "시선이 화면 밖 왼쪽 위를 향하며, 우측 배경의 스크린과 일치하지 않음.",
      "built_space": "법정 내부. 카메라가 지시된 스크린 쪽이 아닌 방청석 쪽에 잘못 위치함.",
      "entities": "장원섭은 레퍼런스와 일치하나, 텍스트에 없는 방청객 여러 명이 배경에 임의로 등장함.",
      "hard_violations": [
       "지시문에 명시되지 않은 임의의 인물(배경 방청객) 추가"
      ],
      "physics": "의자에 앉아 체중이 정상적으로 지탱된 자세."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "시선 방향과 카메라 위치가 지시문과 어긋나지만, 허용되지 않은 인물을 추가하지 않아 상대적으로 낫습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "시선과 카메라 위치 오류에 더해, 명시되지 않은 방청객을 추가하여 인물 제한 규칙을 심각하게 위반했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "장원섭의 시선이 화면 밖 왼쪽 위를 향하나, 목표물인 스크린은 우측 배경에 있어 시선이 닿지 않음.",
      "built_space": "법정 내부. 카메라가 스크린 쪽에서 인물을 향해야 하나, 반대로 방청석에서 스크린을 바라보는 구도로 잘못 배치됨.",
      "entities": "장원섭의 외모와 복장은 레퍼런스와 일치. 스크린에 단체 사진이 있으며, 추가 인물은 없음.",
      "hard_violations": [],
      "physics": "의자에 앉아 체중이 정상적으로 지탱된 자세."
     },
     {
      "label": "A",
      "direction": "시선이 화면 밖 왼쪽 위를 향하며, 우측 배경의 스크린과 일치하지 않음.",
      "built_space": "법정 내부. 카메라가 지시된 스크린 쪽이 아닌 방청석 쪽에 잘못 위치함.",
      "entities": "장원섭은 레퍼런스와 일치하나, 텍스트에 없는 방청객 여러 명이 배경에 임의로 등장함.",
      "hard_violations": [
       "지시문에 명시되지 않은 임의의 인물(배경 방청객) 추가"
      ],
      "physics": "의자에 앉아 체중이 정상적으로 지탱된 자세."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 11
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S74sh12"
  }
 },
 "S75sh4::cine": {
  "applied": true,
  "fingerprint": "fa39fc1da3f897f9fe7da251b8fc8d01fe424bd9cb2643bb9ef197c8b07652d6",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S75sh4_sel.png",
  "source_sha256": "d59a3ea46c7b53260d48922bfd784d4fa6027e8c7f7220b92f467d008e55b12f",
  "file": "S75sh4_cine.png",
  "latency_ms": 19197
 },
 "S75sh5::signage": {
  "fp": "8adf73baed9b8f26",
  "inscriptions": [
   {
    "surface_native": "법정 증거 화면의 디지털 레이블",
    "text_native": "증거 제3호",
    "reason_ko": "법정 전면의 대형 스크린에 제시된 증거 자료임을 나타내기 위해 화면에 증거 번호 표기가 필요합니다."
   }
  ]
 },
 "S75sh5": {
  "input_fingerprint": "95c217d3a74ebd7d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 화면 속 10대 후반 한국인 여학생의 눈 밑에 선명하게 찍힌 검은 점 클로즈업.\n\nLOCATION (lock): On the large evidence screen at the front of the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the initial closest point of the pullback, photograph the courtroom screen nearly square but slightly off its central axis, isolating the Korean schoolgirl’s eye area and the distinct black mark beneath it. The projected facial detail carries the frame’s visual emphasis while a narrow surrounding portion of the photograph remains visible enough to identify it as screen content rather than a live face.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: black mark beneath the schoolgirl's eye on the courtroom screen in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: courtroom screen image (Displaying the alibi photograph dated 2001.02.04) — The camera-facing display shows a close portion of the alibi photograph, specifically the schoolgirl’s eye and the clear black mark beneath it; used as Carries the isolated evidentiary detail that matches the upward eyeline from the preceding reaction shot.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The projected photograph is presented in restrained neutral tones with moderate contrast so the black facial mark remains clear without stylization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The projected alibi photograph remains unchanged, including the schoolgirl's large dark mole beneath her eye and the printed date “2001.02.04.”\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 지국현의 변호인 (Korean 남성, 40대 초반, 반듯한 얼굴, 단정한 짧은 검은색 머리) — wearing: 말끔한 블랙 정장, 화이트 셔츠, 신뢰감을 주는 푸른 계열의 넥타이; 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 증거 화면의 디지털 레이블: \"증거 제3호\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 화면 속 10대 후반 한국인 여학생의 눈 밑에 선명하게 찍힌 검은 점 클로즈업.\n\nLOCATION (lock): On the large evidence screen at the front of the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the initial closest point of the pullback, photograph the courtroom screen nearly square but slightly off its central axis, isolating the Korean schoolgirl’s eye area and the distinct black mark beneath it. The projected facial detail carries the frame’s visual emphasis while a narrow surrounding portion of the photograph remains visible enough to identify it as screen content rather than a live face.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: black mark beneath the schoolgirl's eye on the courtroom screen in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: courtroom screen image (Displaying the alibi photograph dated 2001.02.04) — The camera-facing display shows a close portion of the alibi photograph, specifically the schoolgirl’s eye and the clear black mark beneath it; used as Carries the isolated evidentiary detail that matches the upward eyeline from the preceding reaction shot.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The projected photograph is presented in restrained neutral tones with moderate contrast so the black facial mark remains clear without stylization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The projected alibi photograph remains unchanged, including the schoolgirl's large dark mole beneath her eye and the printed date “2001.02.04.”\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 지국현의 변호인 (Korean 남성, 40대 초반, 반듯한 얼굴, 단정한 짧은 검은색 머리) — wearing: 말끔한 블랙 정장, 화이트 셔츠, 신뢰감을 주는 푸른 계열의 넥타이; 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 증거 화면의 디지털 레이블: \"증거 제3호\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 화면 속 10대 후반 한국인 여학생의 눈 밑에 선명하게 찍힌 검은 점 클로즈업.\n\nLOCATION (lock): On the large evidence screen at the front of the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At the initial closest point of the pullback, photograph the courtroom screen nearly square but slightly off its central axis, isolating the Korean schoolgirl’s eye area and the distinct black mark beneath it. The projected facial detail carries the frame’s visual emphasis while a narrow surrounding portion of the photograph remains visible enough to identify it as screen content rather than a live face.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: black mark beneath the schoolgirl's eye on the courtroom screen in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: courtroom screen image (Displaying the alibi photograph dated 2001.02.04) — The camera-facing display shows a close portion of the alibi photograph, specifically the schoolgirl’s eye and the clear black mark beneath it; used as Carries the isolated evidentiary detail that matches the upward eyeline from the preceding reaction shot.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The projected photograph is presented in restrained neutral tones with moderate contrast so the black facial mark remains clear without stylization.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The projected alibi photograph remains unchanged, including the schoolgirl's large dark mole beneath her eye and the printed date “2001.02.04.”\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 지국현의 변호인 (Korean 남성, 40대 초반, 반듯한 얼굴, 단정한 짧은 검은색 머리) — wearing: 말끔한 블랙 정장, 화이트 셔츠, 신뢰감을 주는 푸른 계열의 넥타이; 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 증거 화면의 디지털 레이블: \"증거 제3호\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.6,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.4,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.15,
    "B": 1.571
   },
   "violations": {
    "A": [
     "[openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남성 인물을 프레임에 추가함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1150,
   "B": 1571
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1150,
    "verdict_ko": "눈과 점을 클로즈업하라는 프레이밍 지시와 지정된 텍스트를 정확히 구현하여 제시된 요구사항에 가장 부합함.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 남성 인물을 프레임에 추가함"
   },
   {
    "label": "B",
    "score": 1571,
    "verdict_ko": "눈 주변의 디테일한 클로즈업 샷을 요구한 프레이밍 지시를 무시하고 전체 얼굴을 노출하여 핵심 조건을 충족하지 못함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S75sh4_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "재질 묘사 위반: 눈 밑의 검은 점이 사진 속 인물의 실제 점(mole)이 아니라, 화면 위에 평면적인 원형 그래픽 스티커를 덧붙인 것처럼 이질적으로 렌더링되었습니다.",
     "fix_en": "Replace the flat, geometric black dot with a natural, organic dark mole that integrates with the skin's texture. Preserve the girl's face, the screen lighting, and the digital text.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "눈과 눈밑 검은 점만 고립한 인서트 클로즈업이 아니라 모니터 전체와 여학생 얼굴 전체가 프레임을 채운다.",
     "fix_en": "Redraw the monitor's display content to show only an extreme close-up of the girl's eye and the black mole, maximizing their size within the screen area. Preserve the camera's current framing of the room, the monitor's physical bezel, the wooden background wall, and the digital text overlays.",
     "severity": "critical",
     "observation_index": 2,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라 프레이밍 위반: 소녀의 눈 부위와 검은 점만을 고립시키는 인서트 클로즈업(insert close-up)을 요구했으나, 얼굴 전체와 상반신 일부까지 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "재질 묘사 위반: 눈 밑의 검은 점이 사진 속 인물의 실제 점(mole)이 아니라, 화면 위에 평면적인 원형 그래픽 스티커를 덧붙인 것처럼 이질적으로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "눈과 눈밑 검은 점만 고립한 인서트 클로즈업이 아니라 모니터 전체와 여학생 얼굴 전체가 프레임을 채운다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "재질 묘사 위반: 눈 밑의 검은 점이 사진 속 인물의 실제 점(mole)이 아니라, 화면 위에 평면적인 원형 그래픽 스티커를 덧붙인 것처럼 이질적으로 렌더링되었습니다.",
    "fix_en": "Replace the flat, geometric black dot with a natural, organic dark mole that integrates with the skin's texture. Preserve the girl's face, the screen lighting, and the digital text.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the monitor's display content to show only an extreme close-up of the girl's eye and the black mole, maximizing their size within the screen area. Preserve the camera's current framing of the room, the monitor's physical bezel, the wooden background wall, and the digital text overlays.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "요청된 '인서트 클로즈업' 샷 크기를 완벽하게 준수하여 눈과 점 부위만 스크린에 화면 가득 채워 연출함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "눈 부위만 고립시키는 클로즈업 샷을 요구했으나, 소녀의 얼굴 전체와 상반신까지 나오는 와이드한 프레임을 사용하여 프레이밍 지시를 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 법정 내 증거 화면을 향해 있으며, 화면 속 소녀의 시선은 약간 아래쪽을 향함.",
      "built_space": "나무 패널 벽면 앞에 검은색 스탠드로 지지된 증거용 디스플레이가 위치함.",
      "entities": "증거 화면 텍스트('증거 제3호', '2001.02.04')가 올바르게 렌더링됨. 화면에는 10대 여학생의 얼굴 전체가 보이며, 오른쪽 눈 밑에 점이 있음.",
      "hard_violations": [],
      "physics": "증거 화면 디스플레이는 스탠드에 안정적으로 고정되어 있음."
     },
     {
      "label": "B",
      "direction": "카메라는 법정 내 증거 화면을 정면에서 약간 치우친 각도로 바라보고 있음.",
      "built_space": "나무 패널 벽면 앞에 검은색 스탠드로 지지된 증거용 디스플레이가 위치함.",
      "entities": "증거 화면 텍스트('증거 제3호', '2001.02.04')가 명확히 렌더링됨. 화면에는 지시된 대로 여학생의 한쪽 눈과 그 밑의 뚜렷한 검은 점이 클로즈업되어 나타남.",
      "hard_violations": [],
      "physics": "증거 화면 디스플레이는 스탠드에 안정적으로 고정되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "요청된 '인서트 클로즈업' 샷 크기를 완벽하게 준수하여 눈과 점 부위만 스크린에 화면 가득 채워 연출함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "눈 부위만 고립시키는 클로즈업 샷을 요구했으나, 소녀의 얼굴 전체와 상반신까지 나오는 와이드한 프레임을 사용하여 프레이밍 지시를 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 법정 내 증거 화면을 향해 있으며, 화면 속 소녀의 시선은 약간 아래쪽을 향함.",
      "built_space": "나무 패널 벽면 앞에 검은색 스탠드로 지지된 증거용 디스플레이가 위치함.",
      "entities": "증거 화면 텍스트('증거 제3호', '2001.02.04')가 올바르게 렌더링됨. 화면에는 10대 여학생의 얼굴 전체가 보이며, 오른쪽 눈 밑에 점이 있음.",
      "hard_violations": [],
      "physics": "증거 화면 디스플레이는 스탠드에 안정적으로 고정되어 있음."
     },
     {
      "label": "B",
      "direction": "카메라는 법정 내 증거 화면을 정면에서 약간 치우친 각도로 바라보고 있음.",
      "built_space": "나무 패널 벽면 앞에 검은색 스탠드로 지지된 증거용 디스플레이가 위치함.",
      "entities": "증거 화면 텍스트('증거 제3호', '2001.02.04')가 명확히 렌더링됨. 화면에는 지시된 대로 여학생의 한쪽 눈과 그 밑의 뚜렷한 검은 점이 클로즈업되어 나타남.",
      "hard_violations": [],
      "physics": "증거 화면 디스플레이는 스탠드에 안정적으로 고정되어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "프롬프트가 요구한 '눈 밑에 선명하게 찍힌 검은 점 클로즈업'이라는 디테일 인서트 샷의 프레이밍을 정확하게 구현했으며, 지정된 텍스트와 배경 구조물도 훌륭하게 반영했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "눈 부위를 고립시키는 클로즈업을 요구한 프롬프트의 지시를 무시하고 인물의 얼굴 전체를 보여주어 샷의 크기와 구도에서 감점되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 법정 내 스크린을 정면으로 바라보고 있음.",
      "built_space": "나무 패널 벽면을 배경으로 두 개의 기둥이 달린 스탠드 위에 평면 스크린이 설치되어 있음.",
      "entities": "스크린 화면에는 사람의 눈과 그 아래 선명한 검은 점이 클로즈업되어 있으며, 좌측 상단에 '증거 제3호', 우측 하단에 '2001.02.04' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "스크린은 하단의 검은색 스탠드 기둥에 의해 정상적으로 지탱되고 있음."
     },
     {
      "label": "B",
      "direction": "카메라는 법정 내 스크린을 정면으로 바라보고 있음.",
      "built_space": "나무 패널 벽면을 배경으로 두 개의 기둥이 달린 스탠드 위에 평면 스크린이 설치되어 있음.",
      "entities": "스크린 화면에는 교복을 입은 여학생의 얼굴 전체와 어깨가 보이며, 오른쪽 눈 밑에 작은 검은 점이 있음. 좌측 상단에 '증거 제3호', 우측 하단에 '2001.02.04' 텍스트가 표기됨.",
      "hard_violations": [],
      "physics": "스크린은 하단의 검은색 스탠드 기둥에 의해 정상적으로 지탱되고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "프롬프트가 요구한 '눈 밑에 선명하게 찍힌 검은 점 클로즈업'이라는 디테일 인서트 샷의 프레이밍을 정확하게 구현했으며, 지정된 텍스트와 배경 구조물도 훌륭하게 반영했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "눈 부위를 고립시키는 클로즈업을 요구한 프롬프트의 지시를 무시하고 인물의 얼굴 전체를 보여주어 샷의 크기와 구도에서 감점되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "카메라는 법정 내 스크린을 정면으로 바라보고 있음.",
      "built_space": "나무 패널 벽면을 배경으로 두 개의 기둥이 달린 스탠드 위에 평면 스크린이 설치되어 있음.",
      "entities": "스크린 화면에는 사람의 눈과 그 아래 선명한 검은 점이 클로즈업되어 있으며, 좌측 상단에 '증거 제3호', 우측 하단에 '2001.02.04' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "스크린은 하단의 검은색 스탠드 기둥에 의해 정상적으로 지탱되고 있음."
     },
     {
      "label": "A",
      "direction": "카메라는 법정 내 스크린을 정면으로 바라보고 있음.",
      "built_space": "나무 패널 벽면을 배경으로 두 개의 기둥이 달린 스탠드 위에 평면 스크린이 설치되어 있음.",
      "entities": "스크린 화면에는 교복을 입은 여학생의 얼굴 전체와 어깨가 보이며, 오른쪽 눈 밑에 작은 검은 점이 있음. 좌측 상단에 '증거 제3호', 우측 하단에 '2001.02.04' 텍스트가 표기됨.",
      "hard_violations": [],
      "physics": "스크린은 하단의 검은색 스탠드 기둥에 의해 정상적으로 지탱되고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 19
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S75sh4"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S75sh5::cine": {
  "applied": true,
  "fingerprint": "5ee9aa33f52743596c9ac0dfc40f4de7198aac1139af4fa771701adfdc486584",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S75sh5_sel.png",
  "source_sha256": "ee2d32f99e7f8e5968d9d6b0ad3f9e21a62c9c4e813c7b6cf2261a75676527cc",
  "file": "S75sh5_cine.png",
  "latency_ms": 13377
 },
 "S75sh6::signage": {
  "fp": "34bd494262a30208",
  "inscriptions": [
   {
    "surface_native": "증인석 명패",
    "text_native": "증인석",
    "reason_ko": "법정 내부의 증인석을 사실적으로 식별하기 위해 해당 명패에 한국어 표기가 필요합니다."
   },
   {
    "surface_native": "변호인석 명패",
    "text_native": "변호인석",
    "reason_ko": "변호사가 활동하는 구역을 명확히 구분하고 법정의 현실감을 높이기 위해 테이블 명패에 한국어 표기가 필요합니다."
   }
  ]
 },
 "S75sh6": {
  "input_fingerprint": "9aea01e0b62c5353",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 쪽을 손가락으로 가리킨 채 증인석을 향해 매서운 눈빛으로 입을 벌린 지국현의 변호인의 전경.\n\nLOCATION (lock): Between the defense table and witness stand inside the courtroom, beneath the evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the lateral track at roughly seated head height on the defense attorney’s screen-side flank, using a slight low angle with 지국현의 변호인 large in the left foreground and the witness stand receding diagonally behind him. His torso twists between the two targets: one arm points back toward the screen at upper left while his face and forceful gaze remain directed toward the witness outside the right side of frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 지국현의 변호인 in the middle-left of the frame, foreground, points to courtroom screen; courtroom screen in the upper-left of the frame, background; witness stand in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying the alibi photograph) — Its display face is visible at the upper-left angle and shows the alibi group photograph with the date 2001.02.04; used as Receives the attorney’s pointing line while remaining secondary to his foreground assertion; witness stand (Occupied by the young female witness) — Its attorney-facing side is visible in the right background; used as Establishes the destination of the attorney’s spoken question along the deep diagonal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience remains neutral-to-cool and moderately contrasted, keeping the forceful questioning grounded in procedural realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness stand, and daylight. Exclude the prosecutor's reaction and show the defense attorney pointing toward the screen.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same dated alibi photograph remains on the screen as the defense attorney points toward it during the niece's testimony.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현의 변호인 (Korean 남성, 40대 초반, 반듯한 얼굴, 단정한 짧은 검은색 머리) — wearing: 말끔한 블랙 정장, 화이트 셔츠, 신뢰감을 주는 푸른 계열의 넥타이 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n- 변호인석 명패: \"변호인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 쪽을 손가락으로 가리킨 채 증인석을 향해 매서운 눈빛으로 입을 벌린 지국현의 변호인의 전경.\n\nLOCATION (lock): Between the defense table and witness stand inside the courtroom, beneath the evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the lateral track at roughly seated head height on the defense attorney’s screen-side flank, using a slight low angle with 지국현의 변호인 large in the left foreground and the witness stand receding diagonally behind him. His torso twists between the two targets: one arm points back toward the screen at upper left while his face and forceful gaze remain directed toward the witness outside the right side of frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 지국현의 변호인 in the middle-left of the frame, foreground, points to courtroom screen; courtroom screen in the upper-left of the frame, background; witness stand in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying the alibi photograph) — Its display face is visible at the upper-left angle and shows the alibi group photograph with the date 2001.02.04; used as Receives the attorney’s pointing line while remaining secondary to his foreground assertion; witness stand (Occupied by the young female witness) — Its attorney-facing side is visible in the right background; used as Establishes the destination of the attorney’s spoken question along the deep diagonal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience remains neutral-to-cool and moderately contrasted, keeping the forceful questioning grounded in procedural realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness stand, and daylight. Exclude the prosecutor's reaction and show the defense attorney pointing toward the screen.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same dated alibi photograph remains on the screen as the defense attorney points toward it during the niece's testimony.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현의 변호인 (Korean 남성, 40대 초반, 반듯한 얼굴, 단정한 짧은 검은색 머리) — wearing: 말끔한 블랙 정장, 화이트 셔츠, 신뢰감을 주는 푸른 계열의 넥타이 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n- 변호인석 명패: \"변호인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 쪽을 손가락으로 가리킨 채 증인석을 향해 매서운 눈빛으로 입을 벌린 지국현의 변호인의 전경.\n\nLOCATION (lock): Between the defense table and witness stand inside the courtroom, beneath the evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Settle the lateral track at roughly seated head height on the defense attorney’s screen-side flank, using a slight low angle with 지국현의 변호인 large in the left foreground and the witness stand receding diagonally behind him. His torso twists between the two targets: one arm points back toward the screen at upper left while his face and forceful gaze remain directed toward the witness outside the right side of frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 지국현의 변호인 in the middle-left of the frame, foreground, points to courtroom screen; courtroom screen in the upper-left of the frame, background; witness stand in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: courtroom screen (Displaying the alibi photograph) — Its display face is visible at the upper-left angle and shows the alibi group photograph with the date 2001.02.04; used as Receives the attorney’s pointing line while remaining secondary to his foreground assertion; witness stand (Occupied by the young female witness) — Its attorney-facing side is visible in the right background; used as Establishes the destination of the attorney’s spoken question along the deep diagonal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime courtroom ambience remains neutral-to-cool and moderately contrasted, keeping the forceful questioning grounded in procedural realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness stand, and daylight. Exclude the prosecutor's reaction and show the defense attorney pointing toward the screen.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The same dated alibi photograph remains on the screen as the defense attorney points toward it during the niece's testimony.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현의 변호인 (Korean 남성, 40대 초반, 반듯한 얼굴, 단정한 짧은 검은색 머리) — wearing: 말끔한 블랙 정장, 화이트 셔츠, 신뢰감을 주는 푸른 계열의 넥타이 — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n- 변호인석 명패: \"변호인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "변호인의 시선은 오른쪽 증인석을 향하고, 손가락은 화면 왼쪽을 가리킵니다(스크린이 변호인의 팔 뒤쪽에 위치).",
    "built_space": "좌측 전경에 변호인석, 우측 배경에 증인석, 좌측 상단 배경에 스크린이 지시된 레이아웃대로 배치되어 있습니다.",
    "entities": "변호인(40대, 흑색 정장, 푸른 넥타이), 증인(젊은 여성), 증거 스크린 이미지 및 텍스트, 명패 텍스트가 모두 프롬프트와 레퍼런스에 일치합니다.",
    "hard_violations": [],
    "physics": "모든 인물과 구조물이 바닥과 가구에 올바르게 지지되어 있습니다."
   },
   {
    "label": "B",
    "direction": "변호인의 시선은 오른쪽을 향하고 손가락은 좌측 상단을 가리키나, 스크린이 머리 뒤(중앙)에 있어 가리키는 대상이 일치하지 않습니다.",
    "built_space": "스크린이 좌측 상단이 아닌 중앙에 배치되었고, 우측 하단에 증인석이 위치합니다.",
    "entities": "변호인과 증인의 외양, 명패, 스크린 속 이미지와 텍스트 모두 프롬프트 요구사항을 충족합니다.",
    "hard_violations": [],
    "physics": "인물의 자세와 사물이 물리적으로 안정적인 상태를 유지하고 있습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "스크린을 화면 좌측 상단에 배치하는 레이아웃 지시를 정확히 따랐으며, 인물들의 외양과 명패 및 스크린 텍스트를 성공적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "스크린이 화면 중앙에 배치되어 좌측 상단이라는 레이아웃 지시를 어겼으며, 변호인이 가리키는 방향과 스크린 위치가 어긋납니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S75sh5_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "레퍼런스 이미지와 달리 화면 좌측 상단 '증거 제3호' 텍스트의 검은색 배경 상자가 누락되었고, 우측 하단 날짜 텍스트의 색상이 검은색에서 흰색으로 변경됨.",
     "fix_en": "Add a black rectangular background behind the top-left text on the screen, and change the bottom-right date text to black. Preserve the eye image, the attorney, the witness stand, and the framing.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 좌측 변호인이 가리키는 손의 마디와 구조가 뭉개져 해부학적으로 부자연스러움.",
     "fix_en": "Redraw the attorney's pointing hand to have an anatomically correct structure with distinct, naturally folded fingers. Preserve the attorney's face, suit, the screen, the courtroom background, and lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "샷 텍스트·PEOPLE·카메라가 증인을 프레임 밖으로 두라고 했으나 오른쪽 증인석에 여성 증인이 앉아 있다.",
     "fix_en": "Remove the woman seated in the witness stand on the right, replacing her with the empty wooden witness stand and the paneled wall behind it. Preserve the defense attorney, his exact pose and clothing, the screen displaying the eye, the courtroom lighting, and the framing.",
     "severity": "critical",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "레퍼런스 이미지와 달리 화면 좌측 상단 '증거 제3호' 텍스트의 검은색 배경 상자가 누락되었고, 우측 하단 날짜 텍스트의 색상이 검은색에서 흰색으로 변경됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "화면 좌측 변호인이 가리키는 손의 마디와 구조가 뭉개져 해부학적으로 부자연스러움.",
     "severity": "major"
    },
    {
     "issue_ko": "샷 텍스트·PEOPLE·카메라가 증인을 프레임 밖으로 두라고 했으나 오른쪽 증인석에 여성 증인이 앉아 있다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "레퍼런스 이미지와 달리 화면 좌측 상단 '증거 제3호' 텍스트의 검은색 배경 상자가 누락되었고, 우측 하단 날짜 텍스트의 색상이 검은색에서 흰색으로 변경됨.",
    "fix_en": "Add a black rectangular background behind the top-left text on the screen, and change the bottom-right date text to black. Preserve the eye image, the attorney, the witness stand, and the framing.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "화면 좌측 변호인이 가리키는 손의 마디와 구조가 뭉개져 해부학적으로 부자연스러움.",
    "fix_en": "Redraw the attorney's pointing hand to have an anatomically correct structure with distinct, naturally folded fingers. Preserve the attorney's face, suit, the screen, the courtroom background, and lighting.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the woman seated in the witness stand on the right, replacing her with the empty wooden witness stand and the paneled wall behind it. Preserve the defense attorney, his exact pose and clothing, the screen displaying the eye, the courtroom lighting, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "샷 텍스트에 명시된 인물만 묘사하는 규칙을 완벽히 준수하여 빈 증인석을 묘사했으며, 시선과 방향성이 정확합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "구도와 포즈는 훌륭하나, 샷 텍스트에 명시되지 않은 여성 증인을 배경에 추가하여 엄격한 인물 제한 규칙을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 우측 증인석을, 뻗은 손가락은 좌측 상단 화면을 정확히 향함.",
      "built_space": "좌측 전경 변호인석, 상단 화면, 우측 배경 증인석이 요구된 위치에 배치됨.",
      "entities": "변호인의 외모와 복장, 화면 내용이 일치하나, 샷 텍스트에 없는 여성 증인이 우측에 등장함.",
      "hard_violations": [
       "invented people (샷 텍스트에 명시되지 않은 여성 증인 추가)"
      ],
      "physics": "의자에 앉아 하중을 지지하며, 가리키는 왼팔은 어깨를 통해 정상적으로 지탱됨."
     },
     {
      "label": "B",
      "direction": "시선은 우측 증인석을, 뻗은 손가락은 좌측 상단 화면을 정확히 향함.",
      "built_space": "좌측 전경 변호인석, 상단 화면, 우측 배경 증인석이 요구된 위치에 배치됨.",
      "entities": "변호인의 설정과 화면 내용이 완벽히 일치하며, 규칙에 따라 증인석은 비어 있음.",
      "hard_violations": [],
      "physics": "의자에 앉아 하중을 지지하며, 가리키는 왼팔은 어깨를 통해 정상적으로 지탱됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "샷 텍스트에 명시된 인물만 묘사하는 규칙을 완벽히 준수하여 빈 증인석을 묘사했으며, 시선과 방향성이 정확합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "구도와 포즈는 훌륭하나, 샷 텍스트에 명시되지 않은 여성 증인을 배경에 추가하여 엄격한 인물 제한 규칙을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 우측 증인석을, 뻗은 손가락은 좌측 상단 화면을 정확히 향함.",
      "built_space": "좌측 전경 변호인석, 상단 화면, 우측 배경 증인석이 요구된 위치에 배치됨.",
      "entities": "변호인의 외모와 복장, 화면 내용이 일치하나, 샷 텍스트에 없는 여성 증인이 우측에 등장함.",
      "hard_violations": [
       "invented people (샷 텍스트에 명시되지 않은 여성 증인 추가)"
      ],
      "physics": "의자에 앉아 하중을 지지하며, 가리키는 왼팔은 어깨를 통해 정상적으로 지탱됨."
     },
     {
      "label": "B",
      "direction": "시선은 우측 증인석을, 뻗은 손가락은 좌측 상단 화면을 정확히 향함.",
      "built_space": "좌측 전경 변호인석, 상단 화면, 우측 배경 증인석이 요구된 위치에 배치됨.",
      "entities": "변호인의 설정과 화면 내용이 완벽히 일치하며, 규칙에 따라 증인석은 비어 있음.",
      "hard_violations": [],
      "physics": "의자에 앉아 하중을 지지하며, 가리키는 왼팔은 어깨를 통해 정상적으로 지탱됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프레임 우측 밖으로 향하는 시선과 프레임 내 단일 인물 제한 규칙을 완벽하게 준수하며 명패와 스크린 디테일을 잘 살렸습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "프레임 밖에 있어야 할 증인을 화면 내에 등장시켜 지시되지 않은 인물 추가 금지 규정을 심각하게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "변호사가 왼팔을 들어 상단의 스크린을 가리키고 있으며, 시선은 우측 프레임 밖을 강하게 응시하고 있음.",
      "built_space": "법정 내부. 좌측 전경에 '변호인석', 우측 배경에 빈 '증인석' 명패가 있고 좌측 상단에 스크린이 배치됨.",
      "entities": "지시된 복장과 외모의 변호인 1명만 존재함. 스크린의 증거 사진, 날짜(2001.02.04), 두 개의 명패 텍스트가 모두 정확함.",
      "hard_violations": [],
      "physics": "손을 뻗은 변호인의 자세와 몸통의 비틀림이 안정적으로 지지되어 있음."
     },
     {
      "label": "B",
      "direction": "변호사가 스크린을 향해 손을 뻗고 있으며, 시선은 우측을 향함.",
      "built_space": "좌측에 변호인석, 우측에 증인석, 좌측 상단에 스크린이 정상적으로 배치되어 있음.",
      "entities": "변호인, 스크린, 명패는 모두 정확하게 묘사되었으나, 우측 증인석에 여성 인물이 묘사되어 있음.",
      "hard_violations": [
       "지시문에 명시되지 않은 인물(여성 증인)이 프레임 안에 임의로 추가됨",
       "증인이 프레임 우측 밖에 있어야 한다는 카메라 구도 지시 위반"
      ],
      "physics": "두 인물의 자세와 착석 상태 모두 물리적으로 자연스러움."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "프레임 우측 밖으로 향하는 시선과 프레임 내 단일 인물 제한 규칙을 완벽하게 준수하며 명패와 스크린 디테일을 잘 살렸습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프레임 밖에 있어야 할 증인을 화면 내에 등장시켜 지시되지 않은 인물 추가 금지 규정을 심각하게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "변호사가 왼팔을 들어 상단의 스크린을 가리키고 있으며, 시선은 우측 프레임 밖을 강하게 응시하고 있음.",
      "built_space": "법정 내부. 좌측 전경에 '변호인석', 우측 배경에 빈 '증인석' 명패가 있고 좌측 상단에 스크린이 배치됨.",
      "entities": "지시된 복장과 외모의 변호인 1명만 존재함. 스크린의 증거 사진, 날짜(2001.02.04), 두 개의 명패 텍스트가 모두 정확함.",
      "hard_violations": [],
      "physics": "손을 뻗은 변호인의 자세와 몸통의 비틀림이 안정적으로 지지되어 있음."
     },
     {
      "label": "A",
      "direction": "변호사가 스크린을 향해 손을 뻗고 있으며, 시선은 우측을 향함.",
      "built_space": "좌측에 변호인석, 우측에 증인석, 좌측 상단에 스크린이 정상적으로 배치되어 있음.",
      "entities": "변호인, 스크린, 명패는 모두 정확하게 묘사되었으나, 우측 증인석에 여성 인물이 묘사되어 있음.",
      "hard_violations": [
       "지시문에 명시되지 않은 인물(여성 증인)이 프레임 안에 임의로 추가됨",
       "증인이 프레임 우측 밖에 있어야 한다는 카메라 구도 지시 위반"
      ],
      "physics": "두 인물의 자세와 착석 상태 모두 물리적으로 자연스러움."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S75sh5"
  }
 },
 "S75sh6::cine": {
  "applied": true,
  "fingerprint": "84c22e283a2819eca8a7af18b7438080cbdd01112d69b039eb072a9429172ec7",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S75sh6_sel.png",
  "source_sha256": "45f1c9aee7df26764e4e821e6595da4031387665af7c40f07f7a38d11b869564",
  "file": "S75sh6_cine.png",
  "latency_ms": 11426
 },
 "S76sh2::signage": {
  "fp": "3b82cfbb65e40142",
  "inscriptions": [
   {
    "surface_native": "단상 위 현수막",
    "text_native": "안전한 나주, 신뢰받는 나주경찰",
    "reason_ko": "나주경찰서 강당에서 진행되는 공식 집회나 회의 분위기를 사실적으로 묘사하기 위해 무대 뒤편 현수막에 들어갈 경찰 표어가 필요합니다."
   }
  ]
 },
 "era_assess::614b3c8ade89a0ec": {
  "subjects": [
   {
    "subject_native": "대한민국 지방경찰청 대강당 (2001-2017)",
    "search_terms_native": [
     "경찰서 대강당",
     "지방경찰청 강당",
     "경찰청 행사 현수막"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 검색해야 하며 영어나 다른 언어로 번역하지 마십시오.",
    "reason_ko": "한국 경찰 고유의 참수리 문양, 특유의 행사 현수막 디자인, 강당 내부 인테리어는 일반적인 해외 강당 이미지와 크게 다릅니다."
   }
  ]
 },
 "era_ref::ce0d416ddde31fcd": {
  "subject": "대한민국 지방경찰청 대강당 (2001-2017)",
  "terms": [
   "경찰서 대강당",
   "지방경찰청 강당",
   "경찰청 행사 현수막"
  ],
  "queries": [
   [
    "경찰서 대강당 행사 현수막 2001년 2017년",
    "지방경찰청 강당 행사 현수막 2001년 2017년",
    "경찰청 행사 현수막 대강당 2001년 2017년"
   ],
   [
    "site:police.go.kr \"대강당\" \"현수막\" 경찰",
    "site:go.kr \"지방경찰청 대강당\" 행사",
    "\"경찰서 대강당\" \"현수막\"",
    "\"지방경찰청 강당\" \"현수막\""
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://www.police.ac.kr/pds/1745998371010.JPG",
  "picked_reason_ko": "후보 중 유일하게 대한민국 경찰기관 대강당의 실제 무대 공간과 시대에 맞는 마감재·막·영상 및 음향 설비를 읽을 수 있는 사진이다.",
  "sha256": "1b892b52d41e3ea8adf9ec504d5741e36747670e4155b946c7c9dc2ee76f67b5",
  "file": "eraref_ce0d416ddde31fcd.png"
 },
 "S76sh2::bgfirst_bg": {
  "input_fingerprint": "61c4ca55755cfd29",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 단상 아래 객석에 뻣뻣한 자세로 나란히 앉아 앞을 주시하는 나주경찰서장과 전택수의 전경.\n\nLOCATION (lock): In the front seating section of the regional police assembly hall, below the stage and banner.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the front aisle at seated chest height, frame the 나주경찰서장 and 전택수 in a medium oblique two-shot as the track settles into the audience row; they occupy the center and right midground while keeping their attention on the off-screen platform. Their similarly rigid seated silhouettes are distinguished by slightly different shoulder and head angles, with the surrounding uniformed officials held in staggered depth rather than appearing mechanically identical.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 단상 현수막 (Hung above the platform) — The printed front faces the audience and carries the field-meeting slogan stated in the scene; used as Provides the formal institutional backdrop behind the two principal seated figures; 객석 좌석 (Occupied by uniformed police officials) — Seat fronts share the platform-facing axis while the camera views them obliquely from the aisle; used as Creates receding rows around and behind the two principal figures; 경찰 간부 십여 명 (Seated rigidly and watching the platform); used as Forms a restrained institutional background with individual variations in shoulder angle, head position, and spacing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 지방경찰청 대강당 (2001-2017): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 단상 아래 객석에 뻣뻣한 자세로 나란히 앉아 앞을 주시하는 나주경찰서장과 전택수의 전경.\n\nLOCATION (lock): In the front seating section of the regional police assembly hall, below the stage and banner.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From the front aisle at seated chest height, frame the 나주경찰서장 and 전택수 in a medium oblique two-shot as the track settles into the audience row; they occupy the center and right midground while keeping their attention on the off-screen platform. Their similarly rigid seated silhouettes are distinguished by slightly different shoulder and head angles, with the surrounding uniformed officials held in staggered depth rather than appearing mechanically identical.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 단상 현수막 (Hung above the platform) — The printed front faces the audience and carries the field-meeting slogan stated in the scene; used as Provides the formal institutional backdrop behind the two principal seated figures; 객석 좌석 (Occupied by uniformed police officials) — Seat fronts share the platform-facing axis while the camera views them obliquely from the aisle; used as Creates receding rows around and behind the two principal figures; 경찰 간부 십여 명 (Seated rigidly and watching the platform); used as Forms a restrained institutional background with individual variations in shoulder angle, head position, and spacing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 지방경찰청 대강당 (2001-2017): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S76sh2__bgfirst_bg.png",
  "asset_id": "0cd18333-9524-4b75-b407-d8b2549cad1b",
  "input_asset_ids": [
   "9866e99a-83a7-483c-86c4-3504065b27de",
   "30b6cdb3-9609-40af-a1fd-291cc2c968d8"
  ],
  "era_research": {
   "subject": "대한민국 지방경찰청 대강당 (2001-2017)",
   "queries": [
    [
     "경찰서 대강당 행사 현수막 2001년 2017년",
     "지방경찰청 강당 행사 현수막 2001년 2017년",
     "경찰청 행사 현수막 대강당 2001년 2017년"
    ],
    [
     "site:police.go.kr \"대강당\" \"현수막\" 경찰",
     "site:go.kr \"지방경찰청 대강당\" 행사",
     "\"경찰서 대강당\" \"현수막\"",
     "\"지방경찰청 강당\" \"현수막\""
    ]
   ],
   "picked_url": "https://www.police.ac.kr/pds/1745998371010.JPG",
   "sha256": "1b892b52d41e3ea8adf9ec504d5741e36747670e4155b946c7c9dc2ee76f67b5",
   "file": "eraref_ce0d416ddde31fcd.png"
  }
 },
 "S76sh2": {
  "input_fingerprint": "b4daddd88150429a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단상 아래 객석에 뻣뻣한 자세로 나란히 앉아 앞을 주시하는 나주경찰서장과 전택수의 전경.\n\nLOCATION (lock): In the front seating section of the regional police assembly hall, below the stage and banner. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the front aisle at seated chest height, frame the 나주경찰서장 and 전택수 in a medium oblique two-shot as the track settles into the audience row; they occupy the center and right midground while keeping their attention on the off-screen platform. Their similarly rigid seated silhouettes are distinguished by slightly different shoulder and head angles, with the surrounding uniformed officials held in staggered depth rather than appearing mechanically identical.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 단상 현수막 (Hung above the platform) — The printed front faces the audience and carries the field-meeting slogan stated in the scene; used as Provides the formal institutional backdrop behind the two principal seated figures; 객석 좌석 (Occupied by uniformed police officials) — Seat fronts share the platform-facing axis while the camera views them obliquely from the aisle; used as Creates receding rows around and behind the two principal figures; 경찰 간부 십여 명 (Seated rigidly and watching the platform); used as Forms a restrained institutional background with individual variations in shoulder angle, head position, and spacing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's smartphone is in his pocket and begins vibrating during the meeting, immediately before he takes it out.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 단상 위 현수막: \"안전한 나주, 신뢰받는 나주경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단상 아래 객석에 뻣뻣한 자세로 나란히 앉아 앞을 주시하는 나주경찰서장과 전택수의 전경.\n\nLOCATION (lock): In the front seating section of the regional police assembly hall, below the stage and banner. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the front aisle at seated chest height, frame the 나주경찰서장 and 전택수 in a medium oblique two-shot as the track settles into the audience row; they occupy the center and right midground while keeping their attention on the off-screen platform. Their similarly rigid seated silhouettes are distinguished by slightly different shoulder and head angles, with the surrounding uniformed officials held in staggered depth rather than appearing mechanically identical.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 단상 현수막 (Hung above the platform) — The printed front faces the audience and carries the field-meeting slogan stated in the scene; used as Provides the formal institutional backdrop behind the two principal seated figures; 객석 좌석 (Occupied by uniformed police officials) — Seat fronts share the platform-facing axis while the camera views them obliquely from the aisle; used as Creates receding rows around and behind the two principal figures; 경찰 간부 십여 명 (Seated rigidly and watching the platform); used as Forms a restrained institutional background with individual variations in shoulder angle, head position, and spacing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's smartphone is in his pocket and begins vibrating during the meeting, immediately before he takes it out.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 단상 위 현수막: \"안전한 나주, 신뢰받는 나주경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 단상 아래 객석에 뻣뻣한 자세로 나란히 앉아 앞을 주시하는 나주경찰서장과 전택수의 전경.\n\nLOCATION (lock): In the front seating section of the regional police assembly hall, below the stage and banner. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the front aisle at seated chest height, frame the 나주경찰서장 and 전택수 in a medium oblique two-shot as the track settles into the audience row; they occupy the center and right midground while keeping their attention on the off-screen platform. Their similarly rigid seated silhouettes are distinguished by slightly different shoulder and head angles, with the surrounding uniformed officials held in staggered depth rather than appearing mechanically identical.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 단상 현수막 (Hung above the platform) — The printed front faces the audience and carries the field-meeting slogan stated in the scene; used as Provides the formal institutional backdrop behind the two principal seated figures; 객석 좌석 (Occupied by uniformed police officials) — Seat fronts share the platform-facing axis while the camera views them obliquely from the aisle; used as Creates receding rows around and behind the two principal figures; 경찰 간부 십여 명 (Seated rigidly and watching the platform); used as Forms a restrained institutional background with individual variations in shoulder angle, head position, and spacing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination with a restrained neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's smartphone is in his pocket and begins vibrating during the meeting, immediately before he takes it out.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 단상 위 현수막: \"안전한 나주, 신뢰받는 나주경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S76sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S76sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:905501>"
    },
    {
     "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:874875>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L63B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:905501>"
    },
    {
     "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:874875>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "요구된 사선 구도 대신 정면 구도를 렌더링하고 두 주연의 얼굴이 동일하게 생성되는 오류가 있으나, 치명적인 이미지 붕괴는 없어 선정되었습니다."
     },
     {
      "label": "B",
      "score": 1350,
      "verdict_ko": "지시된 구도와 인물 분리는 완벽하게 구현하였으나, 좌측 배경 인물의 얼굴이 기괴하게 뭉개져 실격 처리되었습니다.  ★위반: [gemini-pro] 화면 좌측 배경(앞에서 두 번째 줄)에 앉은 경찰의 얼굴이 이목구비가 전혀 없는 덩어리로 뭉개져 있음 (해부학적 불가능)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.571,
      "B": 1.6
     },
     "adjusted": {
      "A": 1.571,
      "B": 1.35
     },
     "violations": {
      "B": [
       "[gemini-pro] 화면 좌측 배경(앞에서 두 번째 줄)에 앉은 경찰의 얼굴이 이목구비가 전혀 없는 덩어리로 뭉개져 있음 (해부학적 불가능)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.429,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1571,
      "verdict_ko": "요구된 사선 구도 대신 정면 구도를 렌더링하고 두 주연의 얼굴이 동일하게 생성되는 오류가 있으나, 치명적인 이미지 붕괴는 없어 선정되었습니다."
     },
     {
      "label": "B",
      "score": 1350,
      "verdict_ko": "지시된 구도와 인물 분리는 완벽하게 구현하였으나, 좌측 배경 인물의 얼굴이 기괴하게 뭉개져 실격 처리되었습니다.  ★위반: [gemini-pro] 화면 좌측 배경(앞에서 두 번째 줄)에 앉은 경찰의 얼굴이 이목구비가 전혀 없는 덩어리로 뭉개져 있음 (해부학적 불가능)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "지시된 사선 앵글과 인물의 화면 배치 등 핵심 프레이밍을 훌륭히 구현했으나, 배경 인물의 심각한 안면 붕괴가 발생함."
     },
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "현수막 텍스트는 완벽하게 출력되었으나, 사선 구도 지시를 무시한 정면 앵글과 배경 인물들의 기계적 복제로 인해 전반적인 지시 충실도가 낮음."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 주요 인물을 포함한 객석의 인물들이 화면 우측 밖의 단상을 향해 시선을 고정함.",
      "built_space": "카메라가 통로 쪽에서 사선(oblique) 각도로 인물들을 포착하며, 좌석들은 단상을 향해 정렬됨.",
      "entities": "두 인물의 외형은 레퍼런스와 부합하고, 현수막의 '안전한 나주, 신뢰받는 나주경찰' 텍스트가 정확함. 좌측 배경의 인물 중 한 명의 얼굴이 심하게 뭉개져 있음.",
      "hard_violations": [
       "물리적으로 불가능한 해부학적 구조 (좌측 배경 인물의 완전히 뭉개진 얼굴 형상)"
      ],
      "physics": "모든 인물들이 의자에 엉덩이를 대고 정상적으로 착석하여 체중을 지탱함."
     },
     {
      "label": "B",
      "direction": "모든 인물이 카메라 앞쪽의 보이지 않는 단상을 향해 일제히 정면을 주시함.",
      "built_space": "카메라가 지시된 사선 구도가 아닌 완벽한 정면 대칭 구도에 위치하며, 공간의 우드 패널 마감이 레퍼런스와 전혀 다름.",
      "entities": "주요 인물 외형과 현수막의 텍스트는 완벽하게 일치함. 그러나 배경의 경찰 간부들이 지시문을 어기고 모두 기계적으로 동일한 얼굴을 하고 있음.",
      "hard_violations": [
       "중복되거나 여분의 신체 (배경 인물들의 얼굴이 완전히 똑같은 형태로 복제됨)"
      ],
      "physics": "인물들이 의자에 안정적으로 앉아 다리를 바닥에 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "지시된 사선 앵글과 인물의 화면 배치 등 핵심 프레이밍을 훌륭히 구현했으나, 배경 인물의 심각한 안면 붕괴가 발생함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "현수막 텍스트는 완벽하게 출력되었으나, 사선 구도 지시를 무시한 정면 앵글과 배경 인물들의 기계적 복제로 인해 전반적인 지시 충실도가 낮음."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "두 주요 인물을 포함한 객석의 인물들이 화면 우측 밖의 단상을 향해 시선을 고정함.",
      "built_space": "카메라가 통로 쪽에서 사선(oblique) 각도로 인물들을 포착하며, 좌석들은 단상을 향해 정렬됨.",
      "entities": "두 인물의 외형은 레퍼런스와 부합하고, 현수막의 '안전한 나주, 신뢰받는 나주경찰' 텍스트가 정확함. 좌측 배경의 인물 중 한 명의 얼굴이 심하게 뭉개져 있음.",
      "hard_violations": [
       "물리적으로 불가능한 해부학적 구조 (좌측 배경 인물의 완전히 뭉개진 얼굴 형상)"
      ],
      "physics": "모든 인물들이 의자에 엉덩이를 대고 정상적으로 착석하여 체중을 지탱함."
     },
     {
      "label": "A",
      "direction": "모든 인물이 카메라 앞쪽의 보이지 않는 단상을 향해 일제히 정면을 주시함.",
      "built_space": "카메라가 지시된 사선 구도가 아닌 완벽한 정면 대칭 구도에 위치하며, 공간의 우드 패널 마감이 레퍼런스와 전혀 다름.",
      "entities": "주요 인물 외형과 현수막의 텍스트는 완벽하게 일치함. 그러나 배경의 경찰 간부들이 지시문을 어기고 모두 기계적으로 동일한 얼굴을 하고 있음.",
      "hard_violations": [
       "중복되거나 여분의 신체 (배경 인물들의 얼굴이 완전히 똑같은 형태로 복제됨)"
      ],
      "physics": "인물들이 의자에 안정적으로 앉아 다리를 바닥에 지지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1575,
     "B": 1356
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   }
  },
  "totals": {
   "A": 1575,
   "B": 1356
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1571,
    "verdict_ko": "요구된 사선 구도 대신 정면 구도를 렌더링하고 두 주연의 얼굴이 동일하게 생성되는 오류가 있으나, 치명적인 이미지 붕괴는 없어 선정되었습니다."
   },
   {
    "label": "B",
    "score": 1350,
    "verdict_ko": "지시된 구도와 인물 분리는 완벽하게 구현하였으나, 좌측 배경 인물의 얼굴이 기괴하게 뭉개져 실격 처리되었습니다.  ★위반: [gemini-pro] 화면 좌측 배경(앞에서 두 번째 줄)에 앉은 경찰의 얼굴이 이목구비가 전혀 없는 덩어리로 뭉개져 있음 (해부학적 불가능)"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L63B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:905501>"
   },
   {
    "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:874875>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "제공된 배경 원본을 그대로 유지하라는 지시를 위반하여, 객석 의자(검은색에서 회색 팔걸이 의자로 변경), 벽면, 조명, 단상 구조 등 공간 전체가 완전히 다르게 렌더링됨.",
     "fix_en": "Change the audience chairs to black leather with metal legs and remove the window and curtains from the right wall, keeping the characters, their poses, and the overall room layout exactly as they are.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "레이아웃 스케치에 표시된 시선 방향(단상 쪽)을 무시하고, 전경의 두 주요 인물을 비롯한 배경의 경찰 간부들이 카메라 렌즈를 정면으로 응시하고 있음.",
     "fix_en": "Turn the heads and eyes of the two foreground characters and all background officials to look off-screen to the left towards the stage, ensuring no one is looking into the camera lens, while preserving all characters, their seated positions, uniforms, and the background environment.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "전경 앞줄에 나란히 앉은 두 인물의 얼굴이 서로 거의 동일하게 묘사되어, 제공된 레퍼런스의 나주경찰서장과 전택수로 명확히 구분되지 않음.",
     "fix_en": "Replace the face of the center character wearing black gloves with the broader face of the '나주경찰서장' reference so he is distinctly different from the character on the right, keeping his uniform, gloves, posture, the surrounding people, and the background completely unchanged.",
     "severity": "critical",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "제공된 배경 원본을 그대로 유지하라는 지시를 위반하여, 객석 의자(검은색에서 회색 팔걸이 의자로 변경), 벽면, 조명, 단상 구조 등 공간 전체가 완전히 다르게 렌더링됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "레이아웃 스케치에 표시된 시선 방향(단상 쪽)을 무시하고, 전경의 두 주요 인물을 비롯한 배경의 경찰 간부들이 카메라 렌즈를 정면으로 응시하고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "전경 앞줄에 나란히 앉은 두 인물의 얼굴이 서로 거의 동일하게 묘사되어, 제공된 레퍼런스의 나주경찰서장과 전택수로 명확히 구분되지 않음.",
     "severity": "critical"
    },
    {
     "issue_ko": "객석 의자가 참조 배경의 검은 가죽·크롬 다리가 아니라 회색 패브릭·나무 팔걸이로 바뀌어 있다",
     "severity": "major"
    },
    {
     "issue_ko": "좌우 벽 상단의 검은 스피커가 없고 참조에 없는 벽등이 달려 있다",
     "severity": "major"
    },
    {
     "issue_ko": "화면 오른쪽 벽에 참조 배경에 없던 커튼과 창이 있다",
     "severity": "major"
    },
    {
     "issue_ko": "단상 위 의자와 바닥·천장 등 홀의 고정 구조가 참조 배경과 다르다",
     "severity": "major"
    },
    {
     "issue_ko": "화면 가운데 나주경찰서장이 단상이 아닌 카메라를 정면으로 응시한다",
     "severity": "major"
    },
    {
     "issue_ko": "화면 오른쪽 전택수가 앞(단상)이 아니라 화면 오른쪽을 본다",
     "severity": "major"
    },
    {
     "issue_ko": "두 주인공과 주변 간부들이 렌즈를 향해 정렬되어 회의를 보는 순간이 아니라 단체 사진처럼 보인다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 7
   }
  },
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Change the audience chairs to black leather with metal legs and remove the window and curtains from the right wall, keeping the characters, their poses, and the overall room layout exactly as they are.\n- Turn the heads and eyes of the two foreground characters and all background officials to look off-screen to the left towards the stage, ensuring no one is looking into the camera lens, while preserving all characters, their seated positions, uniforms, and the background environment.\n- Replace the face of the center character wearing black gloves with the broader face of the '나주경찰서장' reference so he is distinctly different from the character on the right, keeping his uniform, gloves, posture, the surrounding people, and the background completely unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 현수막 텍스트와 경찰 정복을 잘 구현했으나, 두 인물의 얼굴이 복제되었고 배경 의자의 재질이 임의로 변경되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "배경에 스케치 가이드라인이 그대로 노출되었으며, 인물들의 얼굴과 의상(정장)이 레퍼런스와 전혀 일치하지 않습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 주요 인물의 시선이 정면 오프스크린 단상을 올바르게 향하고 있음.",
      "built_space": "객석 좌석들이 배열되어 있으나, 레퍼런스의 검은색 가죽 의자가 회색 패브릭 의자로 임의 변경됨.",
      "entities": "현수막에 '안전한 나주, 신뢰받는 나주경찰' 텍스트가 정확히 표기됨. 두 인물이 정복을 입었으나 얼굴이 동일하게 복제됨.",
      "hard_violations": [
       "두 주요 인물의 얼굴이 동일하게 복제됨 (duplicated identity)"
      ],
      "physics": "인물들의 엉덩이와 등이 의자에 닿아 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "두 주요 인물의 시선이 정면 단상을 향하고 있음.",
      "built_space": "객석 구조가 구현되었으나, 단상 뒤쪽 벽면에 레이아웃 스케치의 방사형 투시선이 지워지지 않고 남아 있음.",
      "entities": "단상 현수막 텍스트가 누락됨. 두 인물 모두 정복이 아닌 정장을 입고 있으며 얼굴이 레퍼런스와 다름.",
      "hard_violations": [
       "배경 벽면에 스케치 투시선 노출 (leaked markers/diagrams)",
       "레퍼런스와 완전히 다른 인물 및 의상 (invented people)"
      ],
      "physics": "인물들이 의자에 정상적으로 착석해 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 현수막 텍스트와 경찰 정복을 잘 구현했으나, 두 인물의 얼굴이 복제되었고 배경 의자의 재질이 임의로 변경되었습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "배경에 스케치 가이드라인이 그대로 노출되었으며, 인물들의 얼굴과 의상(정장)이 레퍼런스와 전혀 일치하지 않습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "두 주요 인물의 시선이 정면 오프스크린 단상을 올바르게 향하고 있음.",
      "built_space": "객석 좌석들이 배열되어 있으나, 레퍼런스의 검은색 가죽 의자가 회색 패브릭 의자로 임의 변경됨.",
      "entities": "현수막에 '안전한 나주, 신뢰받는 나주경찰' 텍스트가 정확히 표기됨. 두 인물이 정복을 입었으나 얼굴이 동일하게 복제됨.",
      "hard_violations": [
       "두 주요 인물의 얼굴이 동일하게 복제됨 (duplicated identity)"
      ],
      "physics": "인물들의 엉덩이와 등이 의자에 닿아 안정적으로 앉아 있음."
     },
     {
      "label": "B",
      "direction": "두 주요 인물의 시선이 정면 단상을 향하고 있음.",
      "built_space": "객석 구조가 구현되었으나, 단상 뒤쪽 벽면에 레이아웃 스케치의 방사형 투시선이 지워지지 않고 남아 있음.",
      "entities": "단상 현수막 텍스트가 누락됨. 두 인물 모두 정복이 아닌 정장을 입고 있으며 얼굴이 레퍼런스와 다름.",
      "hard_violations": [
       "배경 벽면에 스케치 투시선 노출 (leaked markers/diagrams)",
       "레퍼런스와 완전히 다른 인물 및 의상 (invented people)"
      ],
      "physics": "인물들이 의자에 정상적으로 착석해 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "객석 의자의 재질과 색상이 레퍼런스와 다르게 변경되었으나, 주요 인물의 신원과 경찰 정복(장갑 착용 여부 포함)을 정확히 묘사하고 현수막의 지정 텍스트를 완벽하게 구현하여 프롬프트 충실도가 높습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "구도와 자세는 지시를 따랐으나, 주요 인물들이 경찰 정복이 아닌 정장을 입고 있어 캐릭터 레퍼런스를 완전히 위반했으며 지정된 현수막 텍스트도 누락되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전경의 주요 인물들과 배경의 경찰 간부들 모두 오프스크린 단상 쪽을 똑바로 주시하고 있음.",
      "built_space": "지정된 회의실 구조와 카메라 앵글을 대체로 유지했으나, 객석 의자가 형태가 다른 버튼 장식의 가죽 의자로 변경됨. 단상 위 현수막에 텍스트가 없음.",
      "entities": "주요 인물인 전택수와 나주경찰서장이 레퍼런스에 명시된 정복과 모자 대신 검은색 정장을 입고 있어 신원 및 복장 지시를 심각하게 위반함(나주경찰서장의 장갑 착용만 유지됨). 현수막 텍스트 누락.",
      "hard_violations": [],
      "physics": "인물들이 의자에 바르게 앉아 하중이 좌석에 지지되고 있으며, 무릎 위에 놓인 손의 자세도 자연스러움."
     },
     {
      "label": "B",
      "direction": "전경의 두 주요 인물과 배경 인물들이 모두 뻣뻣한 자세로 정면 단상을 향해 시선을 고정하고 있음.",
      "built_space": "회의실의 전반적인 구조와 배열은 스케치를 따르나, 원본 배경 이미지의 검은색 의자가 회색 패브릭 의자로 변경됨. 현수막의 위치와 형태는 적절함.",
      "entities": "중앙의 나주경찰서장(검은 장갑 착용, 넓은 얼굴)과 우측의 전택수(맨손, 각진 얼굴)가 레퍼런스의 경찰 정복과 모자를 정확하게 착용하고 있음. 단상 위 현수막에 '안전한 나주, 신뢰받는 나주경찰'이라는 텍스트가 오타 없이 정확하게 렌더링됨.",
      "hard_violations": [],
      "physics": "착석한 인물들의 엉덩이와 등이 의자에 정상적으로 접촉되어 무게가 지지되며, 중력에 어긋나는 요소 없이 안정적인 자세를 보임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "객석 의자의 재질과 색상이 레퍼런스와 다르게 변경되었으나, 주요 인물의 신원과 경찰 정복(장갑 착용 여부 포함)을 정확히 묘사하고 현수막의 지정 텍스트를 완벽하게 구현하여 프롬프트 충실도가 높습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "구도와 자세는 지시를 따랐으나, 주요 인물들이 경찰 정복이 아닌 정장을 입고 있어 캐릭터 레퍼런스를 완전히 위반했으며 지정된 현수막 텍스트도 누락되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "전경의 주요 인물들과 배경의 경찰 간부들 모두 오프스크린 단상 쪽을 똑바로 주시하고 있음.",
      "built_space": "지정된 회의실 구조와 카메라 앵글을 대체로 유지했으나, 객석 의자가 형태가 다른 버튼 장식의 가죽 의자로 변경됨. 단상 위 현수막에 텍스트가 없음.",
      "entities": "주요 인물인 전택수와 나주경찰서장이 레퍼런스에 명시된 정복과 모자 대신 검은색 정장을 입고 있어 신원 및 복장 지시를 심각하게 위반함(나주경찰서장의 장갑 착용만 유지됨). 현수막 텍스트 누락.",
      "hard_violations": [],
      "physics": "인물들이 의자에 바르게 앉아 하중이 좌석에 지지되고 있으며, 무릎 위에 놓인 손의 자세도 자연스러움."
     },
     {
      "label": "A",
      "direction": "전경의 두 주요 인물과 배경 인물들이 모두 뻣뻣한 자세로 정면 단상을 향해 시선을 고정하고 있음.",
      "built_space": "회의실의 전반적인 구조와 배열은 스케치를 따르나, 원본 배경 이미지의 검은색 의자가 회색 패브릭 의자로 변경됨. 현수막의 위치와 형태는 적절함.",
      "entities": "중앙의 나주경찰서장(검은 장갑 착용, 넓은 얼굴)과 우측의 전택수(맨손, 각진 얼굴)가 레퍼런스의 경찰 정복과 모자를 정확하게 착용하고 있음. 단상 위 현수막에 '안전한 나주, 신뢰받는 나주경찰'이라는 텍스트가 오타 없이 정확하게 렌더링됨.",
      "hard_violations": [],
      "physics": "착석한 인물들의 엉덩이와 등이 의자에 정상적으로 접촉되어 무게가 지지되며, 중력에 어긋나는 요소 없이 안정적인 자세를 보임."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 11,
     "B": 5
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S76sh2__bgfirst_bg.png",
   "bg_asset_id": "0cd18333-9524-4b75-b407-d8b2549cad1b",
   "bg_record_key": "S76sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S76sh2::cine": {
  "applied": true,
  "fingerprint": "10e5ab42201a55b30571fbad17b6c47eca8bd5bef50b396ac5190972c6343e62",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S76sh2_sel.png",
  "source_sha256": "c4cbe9b05841a37f9cb0caf32863e52c56c64f9f434348081c5ee533ac7a0cfe",
  "file": "S76sh2_cine.png",
  "latency_ms": 14362
 },
 "S76sh3::signage": {
  "fp": "f11ed684b0a4e917",
  "inscriptions": [
   {
    "surface_native": "스마트폰 액정 화면",
    "text_native": "부재중 전화: 강형사",
    "reason_ko": "주머니에서 삐져나와 켜진 스마트폰 화면에 발신인을 표시하여 극적 긴장감을 주기 위해 필요합니다."
   }
  ]
 },
 "S76sh3": {
  "input_fingerprint": "c66dd72ad57a9778",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 주머니 바깥으로 삐져나와 화면이 켜진 스마트폰의 모서리 클로즈업.\n\nLOCATION (lock): At a seated place in the front audience section of the regional police assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hip height beside 전택수's seat, complete the dolly-in on the illuminated corner of the smartphone protruding from his pocket, framing it as an insert while retaining part of his uniformed hip, reaching hand, and the seat edge as scale references. The angle cuts across his body rather than following his forward axis, capturing the instant his fingers begin to retrieve the vibrating phone.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 스마트폰 (Screen on and vibrating) — Only the illuminated corner of the screen-facing side is visible as it protrudes from the pocket; used as Primary insert detail at 전택수's hip, kept below forty percent of the frame with his hand and seat visible around it; 객석 좌석 가장자리 (Occupied by 전택수) — Its aisle-side edge faces the low camera position; used as Provides immediate spatial reference beside the phone and body.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination remains subdued, with the phone screen providing only a localized brightness justified by its activated state.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's lit, vibrating smartphone remains partly protruding from his pocket as he starts to retrieve it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 스마트폰 액정 화면: \"부재중 전화: 강형사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 주머니 바깥으로 삐져나와 화면이 켜진 스마트폰의 모서리 클로즈업.\n\nLOCATION (lock): At a seated place in the front audience section of the regional police assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hip height beside 전택수's seat, complete the dolly-in on the illuminated corner of the smartphone protruding from his pocket, framing it as an insert while retaining part of his uniformed hip, reaching hand, and the seat edge as scale references. The angle cuts across his body rather than following his forward axis, capturing the instant his fingers begin to retrieve the vibrating phone.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 스마트폰 (Screen on and vibrating) — Only the illuminated corner of the screen-facing side is visible as it protrudes from the pocket; used as Primary insert detail at 전택수's hip, kept below forty percent of the frame with his hand and seat visible around it; 객석 좌석 가장자리 (Occupied by 전택수) — Its aisle-side edge faces the low camera position; used as Provides immediate spatial reference beside the phone and body.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination remains subdued, with the phone screen providing only a localized brightness justified by its activated state.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's lit, vibrating smartphone remains partly protruding from his pocket as he starts to retrieve it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 스마트폰 액정 화면: \"부재중 전화: 강형사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 주머니 바깥으로 삐져나와 화면이 켜진 스마트폰의 모서리 클로즈업.\n\nLOCATION (lock): At a seated place in the front audience section of the regional police assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At hip height beside 전택수's seat, complete the dolly-in on the illuminated corner of the smartphone protruding from his pocket, framing it as an insert while retaining part of his uniformed hip, reaching hand, and the seat edge as scale references. The angle cuts across his body rather than following his forward axis, capturing the instant his fingers begin to retrieve the vibrating phone.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: 스마트폰 (Screen on and vibrating) — Only the illuminated corner of the screen-facing side is visible as it protrudes from the pocket; used as Primary insert detail at 전택수's hip, kept below forty percent of the frame with his hand and seat visible around it; 객석 좌석 가장자리 (Occupied by 전택수) — Its aisle-side edge faces the low camera position; used as Provides immediate spatial reference beside the phone and body.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime interior illumination remains subdued, with the phone screen providing only a localized brightness justified by its activated state.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's lit, vibrating smartphone remains partly protruding from his pocket as he starts to retrieve it.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 스마트폰 액정 화면: \"부재중 전화: 강형사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.667,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.333,
    "B": 1.75
   },
   "adjusted": {
    "A": 0.833,
    "B": 1.25
   },
   "violations": {
    "A": [
     "[gemini-pro] 프롬프트가 명시적으로 금지한 추가 인물(배경의 경찰들) 포함",
     "[openrouter:x-ai/grok-4.6] 샷 텍스트가 허용하지 않은 다른 경찰 인물들이 배경에 다수 등장"
    ],
    "B": [
     "[gemini-pro] 명시된 통로 쪽 위치가 아닌 반대편에 카메라를 배치한 불가능한 스테이징",
     "[gemini-pro] 원근감을 완전히 무시한 텍스트 그래픽 오버레이"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 833,
   "B": 1250
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 833,
    "verdict_ko": "카메라 위치와 디스플레이 텍스트의 렌더링은 매우 훌륭하나, 프롬프트가 엄격히 금지한 추가 인물들을 배경에 배치하여 치명적인 오류를 범했습니다.  ★위반: [gemini-pro] 프롬프트가 명시적으로 금지한 추가 인물(배경의 경찰들) 포함 / [openrouter:x-ai/grok-4.6] 샷 텍스트가 허용하지 않은 다른 경찰 인물들이 배경에 다수 등장"
   },
   {
    "label": "B",
    "score": 1250,
    "verdict_ko": "레퍼런스의 공간 구조(통로 위치)를 무시하고 반대편에 카메라를 배치했으며, 스마트폰 텍스트가 기기의 원근을 무시한 그래픽 오버레이로 나타나 탈락입니다.  ★위반: [gemini-pro] 명시된 통로 쪽 위치가 아닌 반대편에 카메라를 배치한 불가능한 스테이징 / [gemini-pro] 원근감을 완전히 무시한 텍스트 그래픽 오버레이"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S76sh2_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "스마트폰 화면의 텍스트가 기기의 기울기와 원근감을 무시한 채 화면 위에 수평으로 덧씌워진 평면적인 그래픽 오버레이처럼 렌더링되었습니다.",
     "fix_en": "Redraw the illuminated interface on the smartphone screen so that it maps correctly to the 3D perspective, tilt, and depth of the glass surface, looking like a real glowing display rather than a flat graphical sticker. Preserve the hand, glove, uniform, phone placement, and out-of-focus background exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "주머니에서 스마트폰 대부분이 나와 화면 전체가 보여 켜진 모서리만 보이라는 인서트 클로즈업을 어긴다.",
     "fix_en": "Lower the smartphone deeper into the pocket so that only its top corner protrudes, concealing the rest of the body and screen inside the fabric. Preserve the character's hand, glove, uniform, and the background.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "스마트폰 화면의 텍스트가 기기의 기울기와 원근감을 무시한 채 화면 위에 수평으로 덧씌워진 평면적인 그래픽 오버레이처럼 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "주머니에서 스마트폰 대부분이 나와 화면 전체가 보여 켜진 모서리만 보이라는 인서트 클로즈업을 어긴다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "주머니에서 스마트폰 대부분이 나와 화면 전체가 보여 켜진 모서리만 보이라는 인서트 클로즈업을 어긴다.",
    "fix_en": "Lower the smartphone deeper into the pocket so that only its top corner protrudes, concealing the rest of the body and screen inside the fabric. Preserve the character's hand, glove, uniform, and the background.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the illuminated interface on the smartphone screen so that it maps correctly to the 3D perspective, tilt, and depth of the glass surface, looking like a real glowing display rather than a flat graphical sticker. Preserve the hand, glove, uniform, phone placement, and out-of-focus background exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "스마트폰 화면의 지정된 텍스트가 정확하고 선명하게 렌더링되었으며, 인서트 샷의 프레이밍과 조명 요구사항을 훌륭하게 충족함."
     },
     {
      "label": "A",
      "score": 1750,
      "verdict_ko": "전반적인 구도와 프레이밍은 동일하게 우수하나, 스마트폰 화면의 텍스트 획이 다소 뭉개지고 형태가 왜곡됨."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.857
     },
     "adjusted": {
      "A": 1.75,
      "B": 1.857
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.143,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "스마트폰 화면의 지정된 텍스트가 정확하고 선명하게 렌더링되었으며, 인서트 샷의 프레이밍과 조명 요구사항을 훌륭하게 충족함."
     },
     {
      "label": "A",
      "score": 1750,
      "verdict_ko": "전반적인 구도와 프레이밍은 동일하게 우수하나, 스마트폰 화면의 텍스트 획이 다소 뭉개지고 형태가 왜곡됨."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "화면의 텍스트가 물리적 깊이를 존중하여 손가락 뒤의 액정 표면에 올바르게 렌더링되었으며, 프롬프트의 지시사항을 잘 이행함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "화면의 텍스트가 전경에 있는 장갑 낀 손가락 위로 겹쳐서 렌더링되는 치명적인 그래픽 오버레이 오류가 발생함."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "카메라는 좌석 옆 엉덩이 높이에 위치하며, 손은 주머니에서 삐져나온 스마트폰을 향하고 있음.",
      "built_space": "경찰 강당의 배경과 앞쪽 객석 좌석의 가장자리가 올바르게 배치되어 있음.",
      "entities": "제복 입은 인물의 엉덩이와 검은 장갑 낀 손이 레퍼런스와 일치하며, 스마트폰 화면에 지정된 텍스트가 정확히 표출됨.",
      "hard_violations": [],
      "physics": "스마트폰은 주머니와 손에 의해 지지되고 있으며, 손가락이 폰을 쥐는 자세가 자연스러움."
     },
     {
      "label": "A",
      "direction": "카메라는 좌석 옆 엉덩이 높이에 위치하며, 손은 주머니에서 삐져나온 스마트폰을 향하고 있음.",
      "built_space": "경찰 강당의 배경과 앞쪽 객석 좌석의 가장자리가 올바르게 배치되어 있음.",
      "entities": "제복과 검은 장갑은 레퍼런스와 일치하나, 화면의 텍스트가 물리적 형태를 무시하고 렌더링됨.",
      "hard_violations": [
       "스마트폰 화면의 텍스트가 폰을 쥐고 있는 손가락 위로 겹쳐서 떠 있는 평면 스티커처럼 렌더링됨(물리적 깊이 및 재질 현실성 위반)."
      ],
      "physics": "스마트폰은 주머니와 손에 의해 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "화면의 텍스트가 물리적 깊이를 존중하여 손가락 뒤의 액정 표면에 올바르게 렌더링되었으며, 프롬프트의 지시사항을 잘 이행함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "화면의 텍스트가 전경에 있는 장갑 낀 손가락 위로 겹쳐서 렌더링되는 치명적인 그래픽 오버레이 오류가 발생함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 좌석 옆 엉덩이 높이에 위치하며, 손은 주머니에서 삐져나온 스마트폰을 향하고 있음.",
      "built_space": "경찰 강당의 배경과 앞쪽 객석 좌석의 가장자리가 올바르게 배치되어 있음.",
      "entities": "제복 입은 인물의 엉덩이와 검은 장갑 낀 손이 레퍼런스와 일치하며, 스마트폰 화면에 지정된 텍스트가 정확히 표출됨.",
      "hard_violations": [],
      "physics": "스마트폰은 주머니와 손에 의해 지지되고 있으며, 손가락이 폰을 쥐는 자세가 자연스러움."
     },
     {
      "label": "B",
      "direction": "카메라는 좌석 옆 엉덩이 높이에 위치하며, 손은 주머니에서 삐져나온 스마트폰을 향하고 있음.",
      "built_space": "경찰 강당의 배경과 앞쪽 객석 좌석의 가장자리가 올바르게 배치되어 있음.",
      "entities": "제복과 검은 장갑은 레퍼런스와 일치하나, 화면의 텍스트가 물리적 형태를 무시하고 렌더링됨.",
      "hard_violations": [
       "스마트폰 화면의 텍스트가 폰을 쥐고 있는 손가락 위로 겹쳐서 떠 있는 평면 스티커처럼 렌더링됨(물리적 깊이 및 재질 현실성 위반)."
      ],
      "physics": "스마트폰은 주머니와 손에 의해 지지됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 1757,
     "B": 1860
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S76sh2"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S76sh3::cine": {
  "applied": true,
  "fingerprint": "56620e036a970bb505e02a72fbdb4e5d58ddaf4da218beb23b9e1d1f7ecd903a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S76sh3_sel.png",
  "source_sha256": "ebb50e5b6c754ad3c5cb5acd59885b6865a960bcc96fc744a842b91a80a177c4",
  "file": "S76sh3_cine.png",
  "latency_ms": 16153
 },
 "S77sh2::signage": {
  "fp": "9d0c6eb9b2f862fa",
  "inscriptions": []
 },
 "era_assess::20c407ae397fe53c": {
  "subjects": [
   {
    "subject_native": "대한민국 지방경찰청 강당 앞 복도 (2001년-2017년)",
    "search_terms_native": [
     "경찰서 복도",
     "지방경찰청 복도",
     "경찰청 강당 입구",
     "경찰서 내부 인테리어"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 경찰청 특유의 벽면 배색, 참수리 마크, 안내 표지판 및 홍보물 구성은 일반적인 해외 경찰서 내부 전경과 뚜렷하게 구분됩니다."
   }
  ]
 },
 "era_ref::9b1d1249a889245d": {
  "subject": "대한민국 지방경찰청 강당 앞 복도 (2001년-2017년)",
  "terms": [
   "경찰서 복도",
   "지방경찰청 복도",
   "경찰청 강당 입구",
   "경찰서 내부 인테리어"
  ],
  "queries": [
   [
    "대한민국 경찰서 복도 지방경찰청 복도 2000년대",
    "경찰청 강당 입구 경찰서 내부 인테리어 2001년 2017년"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://ojsfile.ohmynews.com/STD_IMG_FILE/2015/0407/IE001816959_STD.jpg",
  "picked_reason_ko": "1번은 해당 시대 한국 경찰 관서의 일상적인 복도를 주피사체로 가장 명확히 보여주며 재료·색상·조명·문과 표지판까지 읽기 좋다.",
  "sha256": "885b61e72504f4f5e271dab9a5cedd6ecded3e28ecc2e46af889c6183a8ec880",
  "file": "eraref_9b1d1249a889245d.png"
 },
 "S77sh2::bgfirst_bg": {
  "input_fingerprint": "1ec4ff2fc39b4298",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 귀에 댄 스마트폰을 향해 눈을 크게 뜬 채 당황한 표정으로 입술을 뗀 전택수의 얼굴 클로즈업.\n\nLOCATION (lock): In the corridor directly outside the regional police assembly hall.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just above 전택수's eye level, finish the dolly-in at a close three-quarter angle beside the smartphone so his widened eyes and newly parted lips carry the frame while the phone remains against the near side of his face. His head is held still in mid-call, attention absorbed by the unheard caller as the corridor falls soft behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — Its side edge and back are seen beside 전택수's cheek; the screen is turned away from the camera; used as Held against his face as the narrative source of the alarming information; 전남청 대회의실 앞 복도 (Visible behind 전택수); used as Soft environmental context behind the close reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination with restrained neutral-to-cool color and realistic facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 지방경찰청 강당 앞 복도 (2001년-2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 귀에 댄 스마트폰을 향해 눈을 크게 뜬 채 당황한 표정으로 입술을 뗀 전택수의 얼굴 클로즈업.\n\nLOCATION (lock): In the corridor directly outside the regional police assembly hall.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Just above 전택수's eye level, finish the dolly-in at a close three-quarter angle beside the smartphone so his widened eyes and newly parted lips carry the frame while the phone remains against the near side of his face. His head is held still in mid-call, attention absorbed by the unheard caller as the corridor falls soft behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — Its side edge and back are seen beside 전택수's cheek; the screen is turned away from the camera; used as Held against his face as the narrative source of the alarming information; 전남청 대회의실 앞 복도 (Visible behind 전택수); used as Soft environmental context behind the close reaction.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination with restrained neutral-to-cool color and realistic facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 대한민국 지방경찰청 강당 앞 복도 (2001년-2017년): a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S77sh2__bgfirst_bg.png",
  "asset_id": "e6cd2cbb-4e5d-4d52-b2a1-27b0e5ddcd51",
  "input_asset_ids": [
   "5bd31b1f-0bcf-487a-a18e-ec8a8360ab5a",
   "9654d086-78f6-423b-b300-a99c776b32c7"
  ],
  "era_research": {
   "subject": "대한민국 지방경찰청 강당 앞 복도 (2001년-2017년)",
   "queries": [
    [
     "대한민국 경찰서 복도 지방경찰청 복도 2000년대",
     "경찰청 강당 입구 경찰서 내부 인테리어 2001년 2017년"
    ]
   ],
   "picked_url": "https://ojsfile.ohmynews.com/STD_IMG_FILE/2015/0407/IE001816959_STD.jpg",
   "sha256": "885b61e72504f4f5e271dab9a5cedd6ecded3e28ecc2e46af889c6183a8ec880",
   "file": "eraref_9b1d1249a889245d.png"
  }
 },
 "S77sh2": {
  "input_fingerprint": "acb03279568613a0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 귀에 댄 스마트폰을 향해 눈을 크게 뜬 채 당황한 표정으로 입술을 뗀 전택수의 얼굴 클로즈업.\n\nLOCATION (lock): In the corridor directly outside the regional police assembly hall. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above 전택수's eye level, finish the dolly-in at a close three-quarter angle beside the smartphone so his widened eyes and newly parted lips carry the frame while the phone remains against the near side of his face. His head is held still in mid-call, attention absorbed by the unheard caller as the corridor falls soft behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — Its side edge and back are seen beside 전택수's cheek; the screen is turned away from the camera; used as Held against his face as the narrative source of the alarming information; 전남청 대회의실 앞 복도 (Visible behind 전택수); used as Soft environmental context behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination with restrained neutral-to-cool color and realistic facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu has taken the same smartphone from his pocket and holds it to his ear during Wonseop's call.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 귀에 댄 스마트폰을 향해 눈을 크게 뜬 채 당황한 표정으로 입술을 뗀 전택수의 얼굴 클로즈업.\n\nLOCATION (lock): In the corridor directly outside the regional police assembly hall. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above 전택수's eye level, finish the dolly-in at a close three-quarter angle beside the smartphone so his widened eyes and newly parted lips carry the frame while the phone remains against the near side of his face. His head is held still in mid-call, attention absorbed by the unheard caller as the corridor falls soft behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — Its side edge and back are seen beside 전택수's cheek; the screen is turned away from the camera; used as Held against his face as the narrative source of the alarming information; 전남청 대회의실 앞 복도 (Visible behind 전택수); used as Soft environmental context behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination with restrained neutral-to-cool color and realistic facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu has taken the same smartphone from his pocket and holds it to his ear during Wonseop's call.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 귀에 댄 스마트폰을 향해 눈을 크게 뜬 채 당황한 표정으로 입술을 뗀 전택수의 얼굴 클로즈업.\n\nLOCATION (lock): In the corridor directly outside the regional police assembly hall. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just above 전택수's eye level, finish the dolly-in at a close three-quarter angle beside the smartphone so his widened eyes and newly parted lips carry the frame while the phone remains against the near side of his face. His head is held still in mid-call, attention absorbed by the unheard caller as the corridor falls soft behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — Its side edge and back are seen beside 전택수's cheek; the screen is turned away from the camera; used as Held against his face as the narrative source of the alarming information; 전남청 대회의실 앞 복도 (Visible behind 전택수); used as Soft environmental context behind the close reaction.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination with restrained neutral-to-cool color and realistic facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu has taken the same smartphone from his pocket and holds it to his ear during Wonseop's call.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S77sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S77sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:905501>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L64B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:905501>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트가 명시한 '짧게 정돈된 검은색 머리'와 당황한 표정을 잘 살렸으며, 기준 이미지의 장소 디테일(대강당 표지판 등)을 매우 정확하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "당황한 표정과 전화하는 포즈는 적절하나, 레퍼런스의 모자를 그대로 차용하여 텍스트의 머리스타일 지시를 반영하지 못했고 배경 표지판 글자가 왜곡되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "피사체는 화면 우측 밖을 응시하고 있음.",
      "built_space": "복도 배경, 좌측에 열린 나무 문과 '대강당' 표지판, 복도 끝 정면에 유리문이 보임.",
      "entities": "전택수(50대 중반 한국인 남성, 짧은 검은 머리, 경찰 정복), 귀에 밀착된 스마트폰.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 쥐고 귀에 댄 상태로 안정적으로 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "피사체는 화면 우측 밖을 응시하고 있음.",
      "built_space": "복도 배경, 좌측에 열린 나무 문과 글씨가 변형된 표지판, 정면에 유리문이 보임.",
      "entities": "전택수(경찰 정복과 모자를 착용), 귀에 밀착된 스마트폰.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 쥐고 귀에 댄 포즈를 유지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "single_gemini-pro",
     "models": [
      "gemini-pro"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "프롬프트가 명시한 '짧게 정돈된 검은색 머리'와 당황한 표정을 잘 살렸으며, 기준 이미지의 장소 디테일(대강당 표지판 등)을 매우 정확하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "당황한 표정과 전화하는 포즈는 적절하나, 레퍼런스의 모자를 그대로 차용하여 텍스트의 머리스타일 지시를 반영하지 못했고 배경 표지판 글자가 왜곡되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "피사체는 화면 우측 밖을 응시하고 있음.",
      "built_space": "복도 배경, 좌측에 열린 나무 문과 '대강당' 표지판, 복도 끝 정면에 유리문이 보임.",
      "entities": "전택수(50대 중반 한국인 남성, 짧은 검은 머리, 경찰 정복), 귀에 밀착된 스마트폰.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 쥐고 귀에 댄 상태로 안정적으로 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "피사체는 화면 우측 밖을 응시하고 있음.",
      "built_space": "복도 배경, 좌측에 열린 나무 문과 글씨가 변형된 표지판, 정면에 유리문이 보임.",
      "entities": "전택수(경찰 정복과 모자를 착용), 귀에 밀착된 스마트폰.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 쥐고 귀에 댄 포즈를 유지하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 당황한 표정과 입술을 뗀 모습, 그리고 스마트폰을 향한 시선을 매우 훌륭하게 구현했으며, 명시된 머리 스타일과 배경의 디테일(대강당 표지판 등)이 레퍼런스와 완벽하게 일치합니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "표정과 포즈의 기본 지시사항은 따랐으나, 프롬프트에 명시된 짧은 검은색 머리가 모자에 가려져 보이지 않으며 배경의 텍스트와 디테일 재현력이 상대적으로 떨어집니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "인물의 시선이 화면 우측, 귀에 댄 스마트폰 쪽을 향하고 있습니다.",
      "built_space": "경찰청 대회의실 앞 복도가 레퍼런스 사진과 동일한 구조로 정확히 구현되었으며, 열린 문과 '대강당' 표지판의 위치가 일치합니다.",
      "entities": "전택수(50대 남성, 각진 얼굴, 짧은 검은색 머리)가 지시된 외모와 경찰 제복을 입고 등장하며 스마트폰을 귀에 대고 있습니다.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 안정적으로 쥐고 귀에 밀착시켜 받치고 있습니다."
     },
     {
      "label": "A",
      "direction": "인물의 시선이 스마트폰을 향하지 않고 다소 정면 좌측 허공을 향하고 있습니다.",
      "built_space": "레퍼런스와 유사한 복도 구조이나, 대강당 표지판의 글자가 변형되어 있고 디테일이 일부 생략되었습니다.",
      "entities": "전택수의 외모적 특징이 일부 반영되었으나, 지시문에 명시된 머리 스타일이 모자에 가려져 확인되지 않습니다.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 쥐고 귀에 대고 있는 자세가 자연스럽게 지탱되고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 당황한 표정과 입술을 뗀 모습, 그리고 스마트폰을 향한 시선을 매우 훌륭하게 구현했으며, 명시된 머리 스타일과 배경의 디테일(대강당 표지판 등)이 레퍼런스와 완벽하게 일치합니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "표정과 포즈의 기본 지시사항은 따랐으나, 프롬프트에 명시된 짧은 검은색 머리가 모자에 가려져 보이지 않으며 배경의 텍스트와 디테일 재현력이 상대적으로 떨어집니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선이 화면 우측, 귀에 댄 스마트폰 쪽을 향하고 있습니다.",
      "built_space": "경찰청 대회의실 앞 복도가 레퍼런스 사진과 동일한 구조로 정확히 구현되었으며, 열린 문과 '대강당' 표지판의 위치가 일치합니다.",
      "entities": "전택수(50대 남성, 각진 얼굴, 짧은 검은색 머리)가 지시된 외모와 경찰 제복을 입고 등장하며 스마트폰을 귀에 대고 있습니다.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 안정적으로 쥐고 귀에 밀착시켜 받치고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선이 스마트폰을 향하지 않고 다소 정면 좌측 허공을 향하고 있습니다.",
      "built_space": "레퍼런스와 유사한 복도 구조이나, 대강당 표지판의 글자가 변형되어 있고 디테일이 일부 생략되었습니다.",
      "entities": "전택수의 외모적 특징이 일부 반영되었으나, 지시문에 명시된 머리 스타일이 모자에 가려져 확인되지 않습니다.",
      "hard_violations": [],
      "physics": "오른손이 스마트폰을 쥐고 귀에 대고 있는 자세가 자연스럽게 지탱되고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 10
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "피사체는 화면 우측 밖을 응시하고 있음.",
    "built_space": "복도 배경, 좌측에 열린 나무 문과 '대강당' 표지판, 복도 끝 정면에 유리문이 보임.",
    "entities": "전택수(50대 중반 한국인 남성, 짧은 검은 머리, 경찰 정복), 귀에 밀착된 스마트폰.",
    "hard_violations": [],
    "physics": "오른손이 스마트폰을 쥐고 귀에 댄 상태로 안정적으로 지지하고 있음."
   },
   {
    "label": "B",
    "direction": "피사체는 화면 우측 밖을 응시하고 있음.",
    "built_space": "복도 배경, 좌측에 열린 나무 문과 글씨가 변형된 표지판, 정면에 유리문이 보임.",
    "entities": "전택수(경찰 정복과 모자를 착용), 귀에 밀착된 스마트폰.",
    "hard_violations": [],
    "physics": "오른손이 스마트폰을 쥐고 귀에 댄 포즈를 유지하고 있음."
   }
  ],
  "totals": {
   "A": 14,
   "B": 10
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "프롬프트가 명시한 '짧게 정돈된 검은색 머리'와 당황한 표정을 잘 살렸으며, 기준 이미지의 장소 디테일(대강당 표지판 등)을 매우 정확하게 재현했습니다."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "당황한 표정과 전화하는 포즈는 적절하나, 레퍼런스의 모자를 그대로 차용하여 텍스트의 머리스타일 지시를 반영하지 못했고 배경 표지판 글자가 왜곡되었습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L64B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:905501>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "스마트폰의 뒷면(카메라 렌즈 부분)을 귀와 뺨에 대고 있어 기기의 앞뒤 방향이 거꾸로 되어 있습니다.",
     "fix_en": "Reverse the smartphone so its flat screen is pressed against the ear and cheek, and its back with the camera lenses faces outward to the left. Preserve the character's pose, facial expression, uniform, hand position, and the exact background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "얼굴이 화면을 채우는 타이트한 클로즈업을 지시한 레이아웃 스케치 및 텍스트와 달리, 가슴과 어깨 전체가 보이는 넓은 프레임으로 렌더링되었습니다.",
     "fix_en": "Push the camera in to a tight close-up so the man's face fills the frame, matching the layout sketch scale.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "원본 배경을 그대로 유지하라는 지시와 달리, 우측 벽면의 소화전이 사라지고 문 위쪽 간판의 영문 텍스트가 뭉개지는 등 배경이 임의로 변형되었습니다.",
     "fix_en": "Restore the fire extinguisher cabinet on the right wall and correct the sign text to match the background reference exactly.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "캐릭터 레퍼런스에 있는 단색 넥타이와 제복 견장 디자인이 스트라이프 넥타이와 별 모양 견장으로 임의 변경되었습니다.",
     "fix_en": "Change the tie to a solid dark color and remove the stars from the shoulder epaulettes to match the reference uniform. Preserve the face, pose, phone, and background.",
     "severity": "minor",
     "observation_index": 3
    },
    {
     "issue_ko": "중앙 전택수의 얼굴·이목구비가 캐릭터 레퍼런스 인물과 일치하지 않는다.",
     "fix_en": "Adjust the facial features to match the provided character reference. Preserve the panicked expression, pose, uniform, and background.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "전택수 고개가 스케치의 3/4 측면보다 정면에 가깝고 시선이 오른쪽 화살표가 아니라 거의 카메라를 향한다.",
     "fix_en": "Shift the eyes to look sharply to the right, following the sketch arrow. Preserve the facial features, expression, uniform, and background.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "스마트폰의 뒷면(카메라 렌즈 부분)을 귀와 뺨에 대고 있어 기기의 앞뒤 방향이 거꾸로 되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "얼굴이 화면을 채우는 타이트한 클로즈업을 지시한 레이아웃 스케치 및 텍스트와 달리, 가슴과 어깨 전체가 보이는 넓은 프레임으로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "원본 배경을 그대로 유지하라는 지시와 달리, 우측 벽면의 소화전이 사라지고 문 위쪽 간판의 영문 텍스트가 뭉개지는 등 배경이 임의로 변형되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "캐릭터 레퍼런스에 있는 단색 넥타이와 제복 견장 디자인이 스트라이프 넥타이와 별 모양 견장으로 임의 변경되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "중앙 전택수의 얼굴·이목구비가 캐릭터 레퍼런스 인물과 일치하지 않는다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수 고개가 스케치의 3/4 측면보다 정면에 가깝고 시선이 오른쪽 화살표가 아니라 거의 카메라를 향한다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 복도 원근·오른쪽 벽 회색 패널·끝 유리문 밖 풍경이 제공된 빈 배경 사진과 다르다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "얼굴이 화면을 채우는 타이트한 클로즈업을 지시한 레이아웃 스케치 및 텍스트와 달리, 가슴과 어깨 전체가 보이는 넓은 프레임으로 렌더링되었습니다.",
    "fix_en": "Push the camera in to a tight close-up so the man's face fills the frame, matching the layout sketch scale.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "원본 배경을 그대로 유지하라는 지시와 달리, 우측 벽면의 소화전이 사라지고 문 위쪽 간판의 영문 텍스트가 뭉개지는 등 배경이 임의로 변형되었습니다.",
    "fix_en": "Restore the fire extinguisher cabinet on the right wall and correct the sign text to match the background reference exactly.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "캐릭터 레퍼런스에 있는 단색 넥타이와 제복 견장 디자인이 스트라이프 넥타이와 별 모양 견장으로 임의 변경되었습니다.",
    "fix_en": "Change the tie to a solid dark color and remove the stars from the shoulder epaulettes to match the reference uniform. Preserve the face, pose, phone, and background.",
    "severity": "minor",
    "observation_index": 3
   },
   {
    "issue_ko": "중앙 전택수의 얼굴·이목구비가 캐릭터 레퍼런스 인물과 일치하지 않는다.",
    "fix_en": "Adjust the facial features to match the provided character reference. Preserve the panicked expression, pose, uniform, and background.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "전택수 고개가 스케치의 3/4 측면보다 정면에 가깝고 시선이 오른쪽 화살표가 아니라 거의 카메라를 향한다.",
    "fix_en": "Shift the eyes to look sharply to the right, following the sketch arrow. Preserve the facial features, expression, uniform, and background.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Reverse the smartphone so its flat screen is pressed against the ear and cheek, and its back with the camera lenses faces outward to the left. Preserve the character's pose, facial expression, uniform, hand position, and the exact background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "전택수의 얼굴과 정체성을 정확히 유지했으나, 지시된 클로즈업 프레이밍과 당황한 표정을 무시하고 넓은 샷으로 렌더링한 점이 아쉽습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시된 프레이밍과 표정은 잘 따랐으나, 캐릭터 레퍼런스와 완전히 다른 얼굴을 생성하여 정체성 불가협 원칙에 따라 탈락합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "시선은 화면 우측 허공을 향함.",
      "built_space": "복도 공간으로, 좌측에 열린 나무 문과 정면에 유리문이 보임.",
      "entities": "스마트폰을 들고 있는 경찰 제복의 남성이지만, 전택수 레퍼런스와는 완전히 다른 얼굴의 인물임.",
      "hard_violations": [
       "레퍼런스와 완전히 다른 가상 인물 생성 (정체성 상실)"
      ],
      "physics": "손가락으로 스마트폰을 쥐어 귀 옆에 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "시선은 화면 우측 허공을 향함.",
      "built_space": "배경 원본과 정확히 일치하는 복도와 문 구조가 보임.",
      "entities": "전택수의 얼굴과 제복(모자 포함)이 캐릭터 레퍼런스와 완벽히 일치함. 스마트폰을 귀에 대고 있음.",
      "hard_violations": [],
      "physics": "두 발로 바닥에 서서 한 손으로 스마트폰을 귀 옆에 지탱함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "전택수의 얼굴과 정체성을 정확히 유지했으나, 지시된 클로즈업 프레이밍과 당황한 표정을 무시하고 넓은 샷으로 렌더링한 점이 아쉽습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지시된 프레이밍과 표정은 잘 따랐으나, 캐릭터 레퍼런스와 완전히 다른 얼굴을 생성하여 정체성 불가협 원칙에 따라 탈락합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 화면 우측 허공을 향함.",
      "built_space": "복도 공간으로, 좌측에 열린 나무 문과 정면에 유리문이 보임.",
      "entities": "스마트폰을 들고 있는 경찰 제복의 남성이지만, 전택수 레퍼런스와는 완전히 다른 얼굴의 인물임.",
      "hard_violations": [
       "레퍼런스와 완전히 다른 가상 인물 생성 (정체성 상실)"
      ],
      "physics": "손가락으로 스마트폰을 쥐어 귀 옆에 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "시선은 화면 우측 허공을 향함.",
      "built_space": "배경 원본과 정확히 일치하는 복도와 문 구조가 보임.",
      "entities": "전택수의 얼굴과 제복(모자 포함)이 캐릭터 레퍼런스와 완벽히 일치함. 스마트폰을 귀에 대고 있음.",
      "hard_violations": [],
      "physics": "두 발로 바닥에 서서 한 손으로 스마트폰을 귀 옆에 지탱함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1083,
      "verdict_ko": "지정된 클로즈업 샷 크기와 스케치를 완벽하게 준수했으며, 당황하여 눈을 크게 뜨고 입술을 뗀 표정과 캐릭터의 외모를 훌륭하게 재현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 캐릭터 참조의 전택수가 아닌 다른 얼굴·헤어의 인물을 배치함"
     },
     {
      "label": "A",
      "score": 1125,
      "verdict_ko": "요구된 얼굴 클로즈업 프레이밍을 무시하고 전신에 가까운 미디엄 샷으로 연출하였으며, 당황한 표정 연기도 부족합니다.  ★위반: [gemini-pro] 프레이밍 위반: 명시된 '얼굴 클로즈업'과 레이아웃 스케치를 무시하고 미디엄 샷으로 렌더링함."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.375,
      "B": 1.333
     },
     "adjusted": {
      "A": 1.125,
      "B": 1.083
     },
     "violations": {
      "A": [
       "[gemini-pro] 프레이밍 위반: 명시된 '얼굴 클로즈업'과 레이아웃 스케치를 무시하고 미디엄 샷으로 렌더링함."
      ],
      "B": [
       "[openrouter:x-ai/grok-4.6] 캐릭터 참조의 전택수가 아닌 다른 얼굴·헤어의 인물을 배치함"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.667,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1083,
      "verdict_ko": "지정된 클로즈업 샷 크기와 스케치를 완벽하게 준수했으며, 당황하여 눈을 크게 뜨고 입술을 뗀 표정과 캐릭터의 외모를 훌륭하게 재현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 캐릭터 참조의 전택수가 아닌 다른 얼굴·헤어의 인물을 배치함"
     },
     {
      "label": "B",
      "score": 1125,
      "verdict_ko": "요구된 얼굴 클로즈업 프레이밍을 무시하고 전신에 가까운 미디엄 샷으로 연출하였으며, 당황한 표정 연기도 부족합니다.  ★위반: [gemini-pro] 프레이밍 위반: 명시된 '얼굴 클로즈업'과 레이아웃 스케치를 무시하고 미디엄 샷으로 렌더링함."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1086,
     "B": 1129
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S77sh2__bgfirst_bg.png",
   "bg_asset_id": "e6cd2cbb-4e5d-4d52-b2a1-27b0e5ddcd51",
   "bg_record_key": "S77sh2::bgfirst_bg",
   "chain_winner": true,
   "authority": "plate"
  },
  "ref_mode": "재투영 배경+콘티+엔티티 (2택1: 체인 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S77sh2::cine": {
  "applied": true,
  "fingerprint": "e8c3f2009610d06c062c12aea8cb3cb0a968e6fa8c2a4a3b77cf593e41cac17e",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S77sh2_sel.png",
  "source_sha256": "1f3913e1af71d6f08693f7d4acde6fa48d0ad312ceb5e4d659e475858fda6c37",
  "file": "S77sh2_cine.png",
  "latency_ms": 19387
 },
 "S77sh4::signage": {
  "fp": "19dc000b1e5aed1a",
  "inscriptions": [
   {
    "surface_native": "강당 입구 안내판",
    "text_native": "대강당",
    "reason_ko": "지방경찰청 강당 외곽 복도라는 공간적 배경의 사실감을 높이기 위해 복도 벽면에 설치된 대강당 안내판이 필요합니다."
   }
  ]
 },
 "S77sh4": {
  "input_fingerprint": "4d220b0e063f76c2",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 허공을 향해 결연한 눈빛으로 입을 벌린 전택수의 상체.\n\nLOCATION (lock): Along the interior corridor outside the regional police assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height beside 전택수, complete the upward tilt into a medium three-quarter view as his chin rises and his shoulders gather, shifting emphasis from lowered contemplation to resolve. He occupies one side of the frame with modest open corridor space in the direction of his lifted, unfixed gaze, the phone still held to his ear as he begins his answer.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — The back and side face the camera while the phone remains at his ear; used as Keeps the ongoing call legible without displacing the emphasis from his resolved upper body; 전남청 대회의실 앞 복도 (Visible around and behind 전택수); used as Supplies uncluttered negative space along his lifted gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination remains neutral and understated, preserving a realistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu continues holding the smartphone during the call as he agrees to meet Mi-gyeong.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 강당 입구 안내판: \"대강당\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 허공을 향해 결연한 눈빛으로 입을 벌린 전택수의 상체.\n\nLOCATION (lock): Along the interior corridor outside the regional police assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height beside 전택수, complete the upward tilt into a medium three-quarter view as his chin rises and his shoulders gather, shifting emphasis from lowered contemplation to resolve. He occupies one side of the frame with modest open corridor space in the direction of his lifted, unfixed gaze, the phone still held to his ear as he begins his answer.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — The back and side face the camera while the phone remains at his ear; used as Keeps the ongoing call legible without displacing the emphasis from his resolved upper body; 전남청 대회의실 앞 복도 (Visible around and behind 전택수); used as Supplies uncluttered negative space along his lifted gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination remains neutral and understated, preserving a realistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu continues holding the smartphone during the call as he agrees to meet Mi-gyeong.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 강당 입구 안내판: \"대강당\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 고개를 든 채 허공을 향해 결연한 눈빛으로 입을 벌린 전택수의 상체.\n\nLOCATION (lock): Along the interior corridor outside the regional police assembly hall. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From chest height beside 전택수, complete the upward tilt into a medium three-quarter view as his chin rises and his shoulders gather, shifting emphasis from lowered contemplation to resolve. He occupies one side of the frame with modest open corridor space in the direction of his lifted, unfixed gaze, the phone still held to his ear as he begins his answer.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 스마트폰 (In active use during the call) — The back and side face the camera while the phone remains at his ear; used as Keeps the ongoing call legible without displacing the emphasis from his resolved upper body; 전남청 대회의실 앞 복도 (Visible around and behind 전택수); used as Supplies uncluttered negative space along his lifted gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime corridor illumination remains neutral and understated, preserving a realistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu continues holding the smartphone during the call as he agrees to meet Mi-gyeong.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 강당 입구 안내판: \"대강당\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.875
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.875
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1875
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "가슴 높이에서 올려다보는 앵글과 고개를 들고 입을 벌린 결연한 표정이 프롬프트의 지시를 매우 충실하게 구현했으며, 스마트폰을 쥔 손의 묘사도 자연스럽습니다."
   },
   {
    "label": "B",
    "score": 1875,
    "verdict_ko": "프롬프트의 전반적인 상황과 인물의 표정은 잘 반영되었으나, 스마트폰을 쥔 손가락의 형태가 다소 어색하고 카메라의 앙각 느낌이 A에 비해 약합니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S77sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:905501>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "인물이 스마트폰의 화면(앞면)이 뺨이 아닌 바깥(카메라 방향)을 향하도록 기기를 뒤집어 쥐고 있습니다.",
     "fix_en": "Redraw the smartphone to face the correct way, with its solid back and side visible to the camera and the screen against his ear, preserving the man's pose, uniform, hand position, lighting, and the corridor background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "제복 우측 가슴(화면 좌측)에 위치한 명찰에 의미를 알 수 없는 글자가 왜곡된 형태로 생성되었습니다.",
     "fix_en": "Remove the readable text from the man's name tag, leaving a blank metallic surface, preserving the uniform, pose, lighting, and background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "안내판에 허용되지 않은 영문 'Grand Auditorium'이 선명히 읽힌다.",
     "fix_en": "Erase the English text 'Grand Auditorium' from the entrance sign, leaving only '대강당', preserving all surrounding architecture, lighting, and the man.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "인물이 스마트폰의 화면(앞면)이 뺨이 아닌 바깥(카메라 방향)을 향하도록 기기를 뒤집어 쥐고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "제복 우측 가슴(화면 좌측)에 위치한 명찰에 의미를 알 수 없는 글자가 왜곡된 형태로 생성되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 허공을 향해 결연한 눈빛으로 올려다보지 않고 카메라 쪽을 정면으로 응시한다.",
     "severity": "major"
    },
    {
     "issue_ko": "안내판에 허용되지 않은 영문 'Grand Auditorium'이 선명히 읽힌다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "제복 우측 가슴(화면 좌측)에 위치한 명찰에 의미를 알 수 없는 글자가 왜곡된 형태로 생성되었습니다.",
    "fix_en": "Remove the readable text from the man's name tag, leaving a blank metallic surface, preserving the uniform, pose, lighting, and background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "안내판에 허용되지 않은 영문 'Grand Auditorium'이 선명히 읽힌다.",
    "fix_en": "Erase the English text 'Grand Auditorium' from the entrance sign, leaving only '대강당', preserving all surrounding architecture, lighting, and the man.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the smartphone to face the correct way, with its solid back and side visible to the camera and the screen against his ear, preserving the man's pose, uniform, hand position, lighting, and the corridor background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1875,
      "verdict_ko": "요구된 '결연한 눈빛'과 '고개를 든 채 입을 벌린 상체'를 잘 표현했으며, 제복과 흉장의 디테일이 레퍼런스에 더 가깝습니다."
     },
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "지시된 포즈와 배경은 잘 따랐으나, 표정이 다소 덜 결연해 보이며 가슴 흉장의 형태가 일그러져 아쉽습니다."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.875,
      "B": 1.857
     },
     "adjusted": {
      "A": 1.875,
      "B": 1.857
     },
     "violations": {},
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1875,
      "verdict_ko": "요구된 '결연한 눈빛'과 '고개를 든 채 입을 벌린 상체'를 잘 표현했으며, 제복과 흉장의 디테일이 레퍼런스에 더 가깝습니다."
     },
     {
      "label": "B",
      "score": 1857,
      "verdict_ko": "지시된 포즈와 배경은 잘 따랐으나, 표정이 다소 덜 결연해 보이며 가슴 흉장의 형태가 일그러져 아쉽습니다."
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1875,
      "verdict_ko": "지정된 앵글과 상향된 시선, 결연한 표정을 잘 연출했으며, '대강당' 텍스트와 손의 형태가 자연스러워 A보다 우수함."
     },
     {
      "label": "A",
      "score": 1464,
      "verdict_ko": "안내판 텍스트의 자간이 벌어져 있고, 스마트폰을 쥔 오른손 손가락의 해부학적 묘사가 부자연스러워 감점됨.  ★위반: [gemini-pro] 스마트폰을 쥐고 있는 오른손 새끼손가락 부근의 물리적으로 불가능한 해부학적 기형"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 1.875
     },
     "adjusted": {
      "A": 1.464,
      "B": 1.875
     },
     "violations": {
      "A": [
       "[gemini-pro] 스마트폰을 쥐고 있는 오른손 새끼손가락 부근의 물리적으로 불가능한 해부학적 기형"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1875,
      "verdict_ko": "지정된 앵글과 상향된 시선, 결연한 표정을 잘 연출했으며, '대강당' 텍스트와 손의 형태가 자연스러워 A보다 우수함."
     },
     {
      "label": "B",
      "score": 1464,
      "verdict_ko": "안내판 텍스트의 자간이 벌어져 있고, 스마트폰을 쥔 오른손 손가락의 해부학적 묘사가 부자연스러워 감점됨.  ★위반: [gemini-pro] 스마트폰을 쥐고 있는 오른손 새끼손가락 부근의 물리적으로 불가능한 해부학적 기형"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 3750,
     "B": 3321
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S77sh2"
  }
 },
 "S77sh4::cine": {
  "applied": true,
  "fingerprint": "ee3b8bfb7fc93411a566083c44cb37451113e0922f847ea4aa0a7f55a01bc3a9",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S77sh4_sel.png",
  "source_sha256": "cc963dce8defbf67532a19e172739d0c0e1579b509b2cc2c40e9965d309bde48",
  "file": "S77sh4_cine.png",
  "latency_ms": 10817
 },
 "S78sh3::signage": {
  "fp": "a1c985bef68909c2",
  "inscriptions": []
 },
 "era_assess::ee865a88d72cd4fd": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 경찰 제복 및 순찰차",
    "search_terms_native": [
     "2000년대 한국 경찰",
     "2010년대 경찰차",
     "한국 경찰관 제복 2010년",
     "경찰 순찰차 소나타"
    ],
    "language_lock_native": "이 검색어는 반드시 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국 경찰의 제복 색상과 순찰차 도색 디자인은 시대별로 매우 구체적이어서 고증 없이 그리면 일반적인 서구식 경찰로 오인되기 쉽습니다."
   },
   {
    "subject_native": "2000년대~2010년대 한국 유치원 앞 어린이 보호구역",
    "search_terms_native": [
     "어린이 보호구역 도로",
     "유치원 정문 2000년대",
     "스쿨존 노면표시",
     "초등학교 유치원 앞 도로"
    ],
    "language_lock_native": "이 검색어는 반드시 한국어로만 검색해야 하며, 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "한국의 어린이 보호구역(스쿨존) 특유의 노란색·빨간색 도로 포장 및 한국어 안전 표지판 디자인은 일반 도로와 확연히 다릅니다."
   }
  ]
 },
 "era_ref::b2ad91b2555adccd": {
  "subject": "2000년대~2010년대 한국 경찰 제복 및 순찰차",
  "terms": [
   "2000년대 한국 경찰",
   "2010년대 경찰차",
   "한국 경찰관 제복 2010년",
   "경찰 순찰차 소나타"
  ],
  "queries": [
   [
    "2000년대 한국 경찰 2010년대 경찰차",
    "한국 경찰관 제복 2010년 경찰 순찰차 소나타"
   ],
   [
    "2010년대 경찰차 한국",
    "한국 경찰관 제복 2010년",
    "경찰 순찰차 소나타",
    "2000년대 한국 경찰"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://rasp.racers.co.kr/hdd/gallery1/street/2008/07/01/001_1214911501.jpg",
  "picked_reason_ko": "4번이 관광·전시용 연출 없이 2001~2017년 한국의 평범한 순찰차 외형과 실제 운용 도색을 가장 명확하게 보여 준다.",
  "sha256": "031df99f9ccbd1c0f35ea810097139f655b743004f6d61b11c124af6abc67489",
  "file": "eraref_b2ad91b2555adccd.png"
 },
 "S78sh3::bgfirst_bg": {
  "input_fingerprint": "76050576ede3dcb9",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용 쪽을 향해 고개를 돌린 채 미간을 잔뜩 찌푸린 이미경의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): On the road directly outside the kindergarten entrance, looking across toward the waiting officers.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 이미경's eye height, hold a static face close-up from a three-quarter angle offset from her line toward 서의용, placing her deeply furrowed brow and tightened mouth near the center while her turned cheek establishes the across-road gaze. Her halted upper body is only partially visible, keeping the single changed emphasis on the severity of her expression rather than altering the established sightline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward 서의용 across the road.\n- KEY BACKGROUND ELEMENTS: 유치원 앞 도로 (Visible in shallow background); used as Softly retained behind her to preserve the location without competing with her reaction; 유치원 입구 (Visible behind her) — The approach side lies behind 이미경 as her head turns away from it; used as Provides the destination she had been moving toward before noticing the men.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool grading and humane, non-sensational facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 경찰 제복 및 순찰차: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 서의용 쪽을 향해 고개를 돌린 채 미간을 잔뜩 찌푸린 이미경의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): On the road directly outside the kindergarten entrance, looking across toward the waiting officers.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: At 이미경's eye height, hold a static face close-up from a three-quarter angle offset from her line toward 서의용, placing her deeply furrowed brow and tightened mouth near the center while her turned cheek establishes the across-road gaze. Her halted upper body is only partially visible, keeping the single changed emphasis on the severity of her expression rather than altering the established sightline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward 서의용 across the road.\n- KEY BACKGROUND ELEMENTS: 유치원 앞 도로 (Visible in shallow background); used as Softly retained behind her to preserve the location without competing with her reaction; 유치원 입구 (Visible behind her) — The approach side lies behind 이미경 as her head turns away from it; used as Provides the destination she had been moving toward before noticing the men.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool grading and humane, non-sensational facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 경찰 제복 및 순찰차: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S78sh3__bgfirst_bg.png",
  "asset_id": "8ef788ad-2a71-42ea-a3ab-7b20d84d7594",
  "input_asset_ids": [
   "21c6b359-d69f-449c-a342-15b29ff98e91",
   "55866c0f-c417-4479-924b-a6cb9411273a"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국 경찰 제복 및 순찰차",
   "queries": [
    [
     "2000년대 한국 경찰 2010년대 경찰차",
     "한국 경찰관 제복 2010년 경찰 순찰차 소나타"
    ],
    [
     "2010년대 경찰차 한국",
     "한국 경찰관 제복 2010년",
     "경찰 순찰차 소나타",
     "2000년대 한국 경찰"
    ]
   ],
   "picked_url": "https://rasp.racers.co.kr/hdd/gallery1/street/2008/07/01/001_1214911501.jpg",
   "sha256": "031df99f9ccbd1c0f35ea810097139f655b743004f6d61b11c124af6abc67489",
   "file": "eraref_b2ad91b2555adccd.png"
  }
 },
 "S78sh3": {
  "input_fingerprint": "d34fd9c269ccb00b",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽을 향해 고개를 돌린 채 미간을 잔뜩 찌푸린 이미경의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): On the road directly outside the kindergarten entrance, looking across toward the waiting officers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 이미경's eye height, hold a static face close-up from a three-quarter angle offset from her line toward 서의용, placing her deeply furrowed brow and tightened mouth near the center while her turned cheek establishes the across-road gaze. Her halted upper body is only partially visible, keeping the single changed emphasis on the severity of her expression rather than altering the established sightline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward 서의용 across the road.\n- KEY BACKGROUND ELEMENTS: 유치원 앞 도로 (Visible in shallow background); used as Softly retained behind her to preserve the location without competing with her reaction; 유치원 입구 (Visible behind her) — The approach side lies behind 이미경 as her head turns away from it; used as Provides the destination she had been moving toward before noticing the men.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool grading and humane, non-sensational facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽을 향해 고개를 돌린 채 미간을 잔뜩 찌푸린 이미경의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): On the road directly outside the kindergarten entrance, looking across toward the waiting officers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 이미경's eye height, hold a static face close-up from a three-quarter angle offset from her line toward 서의용, placing her deeply furrowed brow and tightened mouth near the center while her turned cheek establishes the across-road gaze. Her halted upper body is only partially visible, keeping the single changed emphasis on the severity of her expression rather than altering the established sightline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward 서의용 across the road.\n- KEY BACKGROUND ELEMENTS: 유치원 앞 도로 (Visible in shallow background); used as Softly retained behind her to preserve the location without competing with her reaction; 유치원 입구 (Visible behind her) — The approach side lies behind 이미경 as her head turns away from it; used as Provides the destination she had been moving toward before noticing the men.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool grading and humane, non-sensational facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서의용 쪽을 향해 고개를 돌린 채 미간을 잔뜩 찌푸린 이미경의 굳은 얼굴 클로즈업.\n\nLOCATION (lock): On the road directly outside the kindergarten entrance, looking across toward the waiting officers. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 이미경's eye height, hold a static face close-up from a three-quarter angle offset from her line toward 서의용, placing her deeply furrowed brow and tightened mouth near the center while her turned cheek establishes the across-road gaze. Her halted upper body is only partially visible, keeping the single changed emphasis on the severity of her expression rather than altering the established sightline.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground, looks toward 서의용 across the road.\n- KEY BACKGROUND ELEMENTS: 유치원 앞 도로 (Visible in shallow background); used as Softly retained behind her to preserve the location without competing with her reaction; 유치원 입구 (Visible behind her) — The approach side lies behind 이미경 as her head turns away from it; used as Provides the destination she had been moving toward before noticing the men.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool grading and humane, non-sensational facial contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S78sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S78sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L65B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "불필요한 인물 추가 없이 지시된 클로즈업 구도와 피사체의 찌푸린 표정을 훌륭하게 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트가 허용하지 않은 인물(배경의 경찰관들)을 임의로 추가하여 지침을 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경은 화면 오른쪽 밖으로 시선을 두고 있으며, 배경에 있는 경찰관들을 직접 향하지 않음.",
      "built_space": "유치원 건물과 울타리, 그 앞의 도로가 배경에 바르게 배치됨.",
      "entities": "이미경의 인물 레퍼런스와 의상은 일치하나, 샷 텍스트에 포함되지 않은 경찰관 두 명이 배경에 등장함.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(경찰관 두 명) 임의 추가"
      ],
      "physics": "물리적으로 어긋난 부분 없이 안정적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "이미경은 화면 오른쪽 밖을 향해 시선을 두고 있음.",
      "built_space": "유치원 건물, 입구, 그리고 앞 도로가 배경으로 바르게 배치됨.",
      "entities": "이미경 단독으로 등장하며 레퍼런스의 외모와 의상을 정확히 반영함.",
      "hard_violations": [],
      "physics": "물리적으로 어긋난 부분 없이 화면 내에 자연스럽게 위치함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "불필요한 인물 추가 없이 지시된 클로즈업 구도와 피사체의 찌푸린 표정을 훌륭하게 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "프롬프트가 허용하지 않은 인물(배경의 경찰관들)을 임의로 추가하여 지침을 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "이미경은 화면 오른쪽 밖으로 시선을 두고 있으며, 배경에 있는 경찰관들을 직접 향하지 않음.",
      "built_space": "유치원 건물과 울타리, 그 앞의 도로가 배경에 바르게 배치됨.",
      "entities": "이미경의 인물 레퍼런스와 의상은 일치하나, 샷 텍스트에 포함되지 않은 경찰관 두 명이 배경에 등장함.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(경찰관 두 명) 임의 추가"
      ],
      "physics": "물리적으로 어긋난 부분 없이 안정적으로 서 있음."
     },
     {
      "label": "B",
      "direction": "이미경은 화면 오른쪽 밖을 향해 시선을 두고 있음.",
      "built_space": "유치원 건물, 입구, 그리고 앞 도로가 배경으로 바르게 배치됨.",
      "entities": "이미경 단독으로 등장하며 레퍼런스의 외모와 의상을 정확히 반영함.",
      "hard_violations": [],
      "physics": "물리적으로 어긋난 부분 없이 화면 내에 자연스럽게 위치함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 앵글과 인물의 찌푸린 표정을 정확히 묘사했으며, 허용되지 않은 외부 인물을 추가하지 않아 프롬프트의 엄격한 제한을 잘 준수했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "인물의 표정과 구도는 적절하나, 샷 텍스트에 묘사되지 않은 인물(경찰관 2명)을 배경에 임의로 추가하여 심각한 지시 위반을 범했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물의 시선은 화면 우측 프레임 밖의 대상을 향하고 있습니다.",
      "built_space": "카메라는 인물의 눈높이에 맞춰져 있으며, 배경에는 지시된 유치원 건물과 앞 도로가 정확하게 배치되어 있습니다.",
      "entities": "레퍼런스와 일치하는 이미경 단 한 명만이 프레임에 존재합니다.",
      "hard_violations": [],
      "physics": "인물은 두 발로 지면에 서서 자세를 유지하고 있습니다."
     },
     {
      "label": "B",
      "direction": "인물의 시선은 우측 화면 밖을 향하고 있습니다.",
      "built_space": "눈높이 앵글이며 유치원 건물과 도로 배경이 지시된 대로 배치되어 있습니다.",
      "entities": "이미경이 전경에 있으나, 배경에 프롬프트가 허용하지 않은 남성 경찰관 두 명이 추가되어 있습니다.",
      "hard_violations": [
       "샷 텍스트에 명시되지 않은 허용 외 인물(경찰관 2명) 추가"
      ],
      "physics": "전경의 이미경과 배경의 경찰관들 모두 지면에 서서 자세를 지탱하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 앵글과 인물의 찌푸린 표정을 정확히 묘사했으며, 허용되지 않은 외부 인물을 추가하지 않아 프롬프트의 엄격한 제한을 잘 준수했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "인물의 표정과 구도는 적절하나, 샷 텍스트에 묘사되지 않은 인물(경찰관 2명)을 배경에 임의로 추가하여 심각한 지시 위반을 범했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "인물의 시선은 화면 우측 프레임 밖의 대상을 향하고 있습니다.",
      "built_space": "카메라는 인물의 눈높이에 맞춰져 있으며, 배경에는 지시된 유치원 건물과 앞 도로가 정확하게 배치되어 있습니다.",
      "entities": "레퍼런스와 일치하는 이미경 단 한 명만이 프레임에 존재합니다.",
      "hard_violations": [],
      "physics": "인물은 두 발로 지면에 서서 자세를 유지하고 있습니다."
     },
     {
      "label": "A",
      "direction": "인물의 시선은 우측 화면 밖을 향하고 있습니다.",
      "built_space": "눈높이 앵글이며 유치원 건물과 도로 배경이 지시된 대로 배치되어 있습니다.",
      "entities": "이미경이 전경에 있으나, 배경에 프롬프트가 허용하지 않은 남성 경찰관 두 명이 추가되어 있습니다.",
      "hard_violations": [
       "샷 텍스트에 명시되지 않은 허용 외 인물(경찰관 2명) 추가"
      ],
      "physics": "전경의 이미경과 배경의 경찰관들 모두 지면에 서서 자세를 지탱하고 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "이미경은 화면 오른쪽 밖으로 시선을 두고 있으며, 배경에 있는 경찰관들을 직접 향하지 않음.",
    "built_space": "유치원 건물과 울타리, 그 앞의 도로가 배경에 바르게 배치됨.",
    "entities": "이미경의 인물 레퍼런스와 의상은 일치하나, 샷 텍스트에 포함되지 않은 경찰관 두 명이 배경에 등장함.",
    "hard_violations": [
     "샷 텍스트에 없는 인물(경찰관 두 명) 임의 추가"
    ],
    "physics": "물리적으로 어긋난 부분 없이 안정적으로 서 있음."
   },
   {
    "label": "B",
    "direction": "이미경은 화면 오른쪽 밖을 향해 시선을 두고 있음.",
    "built_space": "유치원 건물, 입구, 그리고 앞 도로가 배경으로 바르게 배치됨.",
    "entities": "이미경 단독으로 등장하며 레퍼런스의 외모와 의상을 정확히 반영함.",
    "hard_violations": [],
    "physics": "물리적으로 어긋난 부분 없이 화면 내에 자연스럽게 위치함."
   }
  ],
  "totals": {
   "A": 6,
   "B": 14
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "불필요한 인물 추가 없이 지시된 클로즈업 구도와 피사체의 찌푸린 표정을 훌륭하게 구현함."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "프롬프트가 허용하지 않은 인물(배경의 경찰관들)을 임의로 추가하여 지침을 위반함."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L65B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "배경에 보이는 유치원 현관의 계단 및 구조물 형태가 비현실적으로 뭉개지고 왜곡됨.",
     "fix_en": "Redraw the kindergarten entrance stairs and canopy in the background to form distinct, realistic architectural steps and lines. Maintain the character, her expression, clothing, foreground framing, and lighting.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "인물 뒤로 도로 전체가 가로놓여 있어, 유치원 입구 앞이라는 프롬프트의 위치 설정과 불일치함.",
     "severity": "major"
    },
    {
     "issue_ko": "배경에 보이는 유치원 현관의 계단 및 구조물 형태가 비현실적으로 뭉개지고 왜곡됨.",
     "severity": "major"
    },
    {
     "issue_ko": "레퍼런스 이미지에 존재하는 가슴 부위의 흰색 직사각형 명찰이 누락됨.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "배경에 보이는 유치원 현관의 계단 및 구조물 형태가 비현실적으로 뭉개지고 왜곡됨.",
    "fix_en": "Redraw the kindergarten entrance stairs and canopy in the background to form distinct, realistic architectural steps and lines. Maintain the character, her expression, clothing, foreground framing, and lighting.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S78sh3__bgfirst_bg.png",
   "bg_asset_id": "8ef788ad-2a71-42ea-a3ab-7b20d84d7594",
   "bg_record_key": "S78sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S78sh3::cine": {
  "applied": true,
  "fingerprint": "e818fe3536073957dfbc7cd5b9b1e74137ba45685cd17fa51f62078aabe0a3f2",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S78sh3_sel.png",
  "source_sha256": "c25abf280731d1b3621928239c977e3632742827b4e19988102f77c13c46f2ee",
  "file": "S78sh3_cine.png",
  "latency_ms": 10310
 },
 "S79sh3::signage": {
  "fp": "6a25f77806cf2f31",
  "inscriptions": []
 },
 "era_assess::d0d8d26f5e499824": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 유치원 놀이터",
    "search_terms_native": [
     "유치원 야외 놀이터 조합놀이대",
     "어린이집 놀이터 바닥 우레탄",
     "한국 2000년대 놀이터"
    ],
    "language_lock_native": "모든 검색어는 오직 한국어로만 작성되어야 하며, 영어 등 다른 언어로 번역하거나 혼용하지 마십시오.",
    "reason_ko": "2001~2017년 기간의 한국 유치원 놀이터는 특유의 알록달록한 플라스틱 조합놀이대와 우레탄 바닥재, 주변 펜스 구조 등 고유의 시각적 형태가 있어 일반적인 서구식 놀이터와 명확히 구분됩니다."
   }
  ]
 },
 "era_ref::afe8ccd67181857f": {
  "subject": "2000년대~2010년대 한국 유치원 놀이터",
  "terms": [
   "유치원 야외 놀이터 조합놀이대",
   "어린이집 놀이터 바닥 우레탄",
   "한국 2000년대 놀이터"
  ],
  "queries": [
   [
    "유치원 야외 놀이터 조합놀이대 한국 2000년대",
    "어린이집 놀이터 바닥 우레탄 한국 2000년대 2010년대"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://file.daesoon.org/webzine/Daesoon_148_%EC%A2%85%EB%8B%A8%EC%86%8C%EC%8B%9D5-2.jpg",
  "picked_reason_ko": "1번은 2000~2010년대 한국 유치원 놀이터의 전형적인 복합 놀이시설과 안전바닥을 가장 크고 명료하게 보여 주는 일상적 사례다.",
  "sha256": "b1c7619cadc128f2043aa7c1c6cfc7c2af860ba68ea7fe6f866ac19ce32cc7e7",
  "file": "eraref_afe8ccd67181857f.png"
 },
 "S79sh3::bgfirst_bg": {
  "input_fingerprint": "a4eb3876a2c6550c",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 전택수의 스마트폰 화면 속 사진을 뚫어지게 내려다보며 눈을 동그랗게 뜬 이미경의 충격받은 얼굴 클로즈업.\n\nLOCATION (lock): At one side of the kindergarten's outdoor playground, apart from the playing children.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just above 이미경's eye line and beside 전택수, finish the dolly-in on her shocked face while angling downward across the smartphone retained at the lower frame edge. 이미경 occupies most of the close frame with widened eyes fixed downward on the photograph; 전택수's near shoulder and extended hand remain peripheral so the evidence and its presenter share the same physical axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 스마트폰 화면 in the lower-center of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward 스마트폰 화면.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Screen on with the photograph displayed) — The screen side is angled toward 이미경 and partially toward the camera, showing the 2001 alibi photograph of 지국현 with 이미경; used as Retained along the lower edge as the immediate cause of 이미경's shock, occupying a limited portion of the frame; 유치원 놀이터 (Visible out of focus behind the subjects); used as Soft background context separating the close confrontation from the wider playground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool color and moderate contrast across her startled features.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 유치원 놀이터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 전택수의 스마트폰 화면 속 사진을 뚫어지게 내려다보며 눈을 동그랗게 뜬 이미경의 충격받은 얼굴 클로즈업.\n\nLOCATION (lock): At one side of the kindergarten's outdoor playground, apart from the playing children.\n\nTIME OF DAY (lock): day.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From just above 이미경's eye line and beside 전택수, finish the dolly-in on her shocked face while angling downward across the smartphone retained at the lower frame edge. 이미경 occupies most of the close frame with widened eyes fixed downward on the photograph; 전택수's near shoulder and extended hand remain peripheral so the evidence and its presenter share the same physical axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 스마트폰 화면 in the lower-center of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward 스마트폰 화면.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Screen on with the photograph displayed) — The screen side is angled toward 이미경 and partially toward the camera, showing the 2001 alibi photograph of 지국현 with 이미경; used as Retained along the lower edge as the immediate cause of 이미경's shock, occupying a limited portion of the frame; 유치원 놀이터 (Visible out of focus behind the subjects); used as Soft background context separating the close confrontation from the wider playground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool color and moderate contrast across her startled features.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 유치원 놀이터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S79sh3__bgfirst_bg.png",
  "asset_id": "515fdbd4-bdb5-4ac5-9ccd-013cc2b6af64",
  "input_asset_ids": [
   "5e076583-97a6-463c-aa0d-8c9d6a8a9e6f",
   "ac608831-07b9-417d-a635-68184c0284f6"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국 유치원 놀이터",
   "queries": [
    [
     "유치원 야외 놀이터 조합놀이대 한국 2000년대",
     "어린이집 놀이터 바닥 우레탄 한국 2000년대 2010년대"
    ]
   ],
   "picked_url": "https://file.daesoon.org/webzine/Daesoon_148_%EC%A2%85%EB%8B%A8%EC%86%8C%EC%8B%9D5-2.jpg",
   "sha256": "b1c7619cadc128f2043aa7c1c6cfc7c2af860ba68ea7fe6f866ac19ce32cc7e7",
   "file": "eraref_afe8ccd67181857f.png"
  }
 },
 "S79sh3": {
  "input_fingerprint": "2fddd8e8e1f0ec0a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 스마트폰 화면 속 사진을 뚫어지게 내려다보며 눈을 동그랗게 뜬 이미경의 충격받은 얼굴 클로즈업.\n\nLOCATION (lock): At one side of the kindergarten's outdoor playground, apart from the playing children. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above 이미경's eye line and beside 전택수, finish the dolly-in on her shocked face while angling downward across the smartphone retained at the lower frame edge. 이미경 occupies most of the close frame with widened eyes fixed downward on the photograph; 전택수's near shoulder and extended hand remain peripheral so the evidence and its presenter share the same physical axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 스마트폰 화면 in the lower-center of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward 스마트폰 화면.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Screen on with the photograph displayed) — The screen side is angled toward 이미경 and partially toward the camera, showing the 2001 alibi photograph of 지국현 with 이미경; used as Retained along the lower edge as the immediate cause of 이미경's shock, occupying a limited portion of the frame; 유치원 놀이터 (Visible out of focus behind the subjects); used as Soft background context separating the close confrontation from the wider playground.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool color and moderate contrast across her startled features.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong is holding Taeksu's smartphone, which displays the 2001 alibi photograph of her with Ji Gukhyeon.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 스마트폰 화면 속 사진을 뚫어지게 내려다보며 눈을 동그랗게 뜬 이미경의 충격받은 얼굴 클로즈업.\n\nLOCATION (lock): At one side of the kindergarten's outdoor playground, apart from the playing children. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above 이미경's eye line and beside 전택수, finish the dolly-in on her shocked face while angling downward across the smartphone retained at the lower frame edge. 이미경 occupies most of the close frame with widened eyes fixed downward on the photograph; 전택수's near shoulder and extended hand remain peripheral so the evidence and its presenter share the same physical axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 스마트폰 화면 in the lower-center of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward 스마트폰 화면.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Screen on with the photograph displayed) — The screen side is angled toward 이미경 and partially toward the camera, showing the 2001 alibi photograph of 지국현 with 이미경; used as Retained along the lower edge as the immediate cause of 이미경's shock, occupying a limited portion of the frame; 유치원 놀이터 (Visible out of focus behind the subjects); used as Soft background context separating the close confrontation from the wider playground.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool color and moderate contrast across her startled features.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong is holding Taeksu's smartphone, which displays the 2001 alibi photograph of her with Ji Gukhyeon.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수의 스마트폰 화면 속 사진을 뚫어지게 내려다보며 눈을 동그랗게 뜬 이미경의 충격받은 얼굴 클로즈업.\n\nLOCATION (lock): At one side of the kindergarten's outdoor playground, apart from the playing children. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From just above 이미경's eye line and beside 전택수, finish the dolly-in on her shocked face while angling downward across the smartphone retained at the lower frame edge. 이미경 occupies most of the close frame with widened eyes fixed downward on the photograph; 전택수's near shoulder and extended hand remain peripheral so the evidence and its presenter share the same physical axis.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 스마트폰 화면 in the lower-center of the frame, foreground; 이미경 in the middle-center of the frame, midground, looks toward 스마트폰 화면.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Screen on with the photograph displayed) — The screen side is angled toward 이미경 and partially toward the camera, showing the 2001 alibi photograph of 지국현 with 이미경; used as Retained along the lower edge as the immediate cause of 이미경's shock, occupying a limited portion of the frame; 유치원 놀이터 (Visible out of focus behind the subjects); used as Soft background context separating the close confrontation from the wider playground.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light with restrained neutral-to-cool color and moderate contrast across her startled features.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong is holding Taeksu's smartphone, which displays the 2001 alibi photograph of her with Ji Gukhyeon.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S79sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S79sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L66B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "전면 왼쪽 인물과 연결되지 않는 방향에서 나타난 기형적인 손이 스마트폰을 들고 있어 물리적 구도에서 치명적인 오류가 발생했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "제시된 샷의 구도, 인물의 시선 및 충격받은 표정, 스마트폰을 쥔 손의 물리적 연결이 프롬프트의 지시와 정확하게 일치합니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경의 시선은 아래쪽의 스마트폰 화면을 향하고 있습니다.",
      "built_space": "유치원 놀이터의 배경 건물과 놀이기구가 레퍼런스와 일치하게 배치되어 있습니다.",
      "entities": "이미경의 인상과 의상이 레퍼런스와 일치하며, 스마트폰 화면에 두 사람의 사진이 표시되어 있습니다.",
      "hard_violations": [
       "물리적으로 불가능한 구도 및 해부학 (전면 좌측의 전택수와 연결될 수 없는 방향에서 기형적인 손이 나타나 스마트폰을 쥐고 있음)"
      ],
      "physics": "스마트폰을 받치는 손의 관절과 손가락 형태가 해부학적으로 무너져 있으며 구조적으로 불안정합니다."
     },
     {
      "label": "B",
      "direction": "이미경이 놀란 눈으로 스마트폰 화면을 정확히 내려다보고 있습니다.",
      "built_space": "배경의 유치원 건물과 놀이터 구성이 레퍼런스의 공간을 충실히 반영합니다.",
      "entities": "이미경의 외모와 복장이 레퍼런스와 완벽히 일치하며, 화면 하단에 스마트폰과 전택수의 어깨가 올바르게 배치되었습니다.",
      "hard_violations": [],
      "physics": "화면 좌측 전경의 인물(전택수)이 왼손으로 스마트폰을 자연스럽게 쥐고 있어 지지가 확실합니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "전면 왼쪽 인물과 연결되지 않는 방향에서 나타난 기형적인 손이 스마트폰을 들고 있어 물리적 구도에서 치명적인 오류가 발생했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "제시된 샷의 구도, 인물의 시선 및 충격받은 표정, 스마트폰을 쥔 손의 물리적 연결이 프롬프트의 지시와 정확하게 일치합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "이미경의 시선은 아래쪽의 스마트폰 화면을 향하고 있습니다.",
      "built_space": "유치원 놀이터의 배경 건물과 놀이기구가 레퍼런스와 일치하게 배치되어 있습니다.",
      "entities": "이미경의 인상과 의상이 레퍼런스와 일치하며, 스마트폰 화면에 두 사람의 사진이 표시되어 있습니다.",
      "hard_violations": [
       "물리적으로 불가능한 구도 및 해부학 (전면 좌측의 전택수와 연결될 수 없는 방향에서 기형적인 손이 나타나 스마트폰을 쥐고 있음)"
      ],
      "physics": "스마트폰을 받치는 손의 관절과 손가락 형태가 해부학적으로 무너져 있으며 구조적으로 불안정합니다."
     },
     {
      "label": "B",
      "direction": "이미경이 놀란 눈으로 스마트폰 화면을 정확히 내려다보고 있습니다.",
      "built_space": "배경의 유치원 건물과 놀이터 구성이 레퍼런스의 공간을 충실히 반영합니다.",
      "entities": "이미경의 외모와 복장이 레퍼런스와 완벽히 일치하며, 화면 하단에 스마트폰과 전택수의 어깨가 올바르게 배치되었습니다.",
      "hard_violations": [],
      "physics": "화면 좌측 전경의 인물(전택수)이 왼손으로 스마트폰을 자연스럽게 쥐고 있어 지지가 확실합니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "해부학적 왜곡 없이 충격받은 표정을 자연스럽게 연출했으며, 요구된 카메라 구도와 배경 요소를 충실히 반영함."
     },
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "눈동자를 비정상적으로 확대하여 '해부학적 변형 불가' 지침을 심각하게 위반함.  ★위반: [gemini-pro] 해부학적으로 불가능할 정도로 확대된 눈동자 (표정 연기 지침 위반)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.714,
      "B": 1.429
     },
     "adjusted": {
      "A": 1.714,
      "B": 1.179
     },
     "violations": {
      "B": [
       "[gemini-pro] 해부학적으로 불가능할 정도로 확대된 눈동자 (표정 연기 지침 위반)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "해부학적 왜곡 없이 충격받은 표정을 자연스럽게 연출했으며, 요구된 카메라 구도와 배경 요소를 충실히 반영함."
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "눈동자를 비정상적으로 확대하여 '해부학적 변형 불가' 지침을 심각하게 위반함.  ★위반: [gemini-pro] 해부학적으로 불가능할 정도로 확대된 눈동자 (표정 연기 지침 위반)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1182,
     "B": 1721
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "이미경의 시선은 아래쪽의 스마트폰 화면을 향하고 있습니다.",
    "built_space": "유치원 놀이터의 배경 건물과 놀이기구가 레퍼런스와 일치하게 배치되어 있습니다.",
    "entities": "이미경의 인상과 의상이 레퍼런스와 일치하며, 스마트폰 화면에 두 사람의 사진이 표시되어 있습니다.",
    "hard_violations": [
     "물리적으로 불가능한 구도 및 해부학 (전면 좌측의 전택수와 연결될 수 없는 방향에서 기형적인 손이 나타나 스마트폰을 쥐고 있음)"
    ],
    "physics": "스마트폰을 받치는 손의 관절과 손가락 형태가 해부학적으로 무너져 있으며 구조적으로 불안정합니다."
   },
   {
    "label": "B",
    "direction": "이미경이 놀란 눈으로 스마트폰 화면을 정확히 내려다보고 있습니다.",
    "built_space": "배경의 유치원 건물과 놀이터 구성이 레퍼런스의 공간을 충실히 반영합니다.",
    "entities": "이미경의 외모와 복장이 레퍼런스와 완벽히 일치하며, 화면 하단에 스마트폰과 전택수의 어깨가 올바르게 배치되었습니다.",
    "hard_violations": [],
    "physics": "화면 좌측 전경의 인물(전택수)이 왼손으로 스마트폰을 자연스럽게 쥐고 있어 지지가 확실합니다."
   }
  ],
  "totals": {
   "A": 1182,
   "B": 1721
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "전면 왼쪽 인물과 연결되지 않는 방향에서 나타난 기형적인 손이 스마트폰을 들고 있어 물리적 구도에서 치명적인 오류가 발생했습니다."
   },
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "제시된 샷의 구도, 인물의 시선 및 충격받은 표정, 스마트폰을 쥔 손의 물리적 연결이 프롬프트의 지시와 정확하게 일치합니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L66B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트의 CARRIED STATE는 이미경이 스마트폰을 들고 있다고 명시했으나, 이미지에서는 화면 왼쪽 전경의 인물(전택수)이 스마트폰을 쥐고 있습니다.",
     "fix_en": "Replace the hand holding the smartphone with Mi-gyeong's hand reaching up from the bottom right to grip the device, completely removing the male hand and arm extending from the left, while strictly preserving Mi-gyeong's face, her pink cardigan, the smartphone's position and screen content, the left character's head and shoulder, and the playground background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "캐릭터 기준 이미지에 있는 이미경의 왼쪽 가슴 명찰이 생성된 이미지의 이미경 겉옷(분홍색 가디건)에는 누락되어 있습니다.",
     "fix_en": "Add a small rectangular yellow name tag to the left chest area of Mi-gyeong's pink cardigan, preserving her face, clothing, the smartphone, and the background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "스마트폰 화면 속 사진을 보면 두 인물 사이(여성의 허리 부근)에 피부색을 띤 정체불명의 형태가 옷과 어색하게 융합되어 묘사되었습니다.",
     "fix_en": "Remove the fleshy artifact between the two figures on the smartphone screen by painting over it with their respective clothing, preserving the rest of the screen content, the holding hand, Mi-gyeong's face, and the background.",
     "severity": "minor",
     "observation_index": 2
    },
    {
     "issue_ko": "전택수의 머리 뒷모습이 왼쪽 전경을 크게 차지해 주변부 어깨·손이 아니다.",
     "fix_en": "Reduce the size of the left character's head and shoulder so they occupy less of the foreground edge, preserving Mi-gyeong's face, clothing, the smartphone, and the background.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "이미경이 프레임 대부분·중앙을 차지하는 얼굴 클로즈업이 아니라 오른쪽에 치우친 투샷이다.",
     "fix_en": "Shift the framing so Mi-gyeong's face occupies the middle-center of the frame as a close-up, preserving her expression, clothing, the smartphone, and the background.",
     "severity": "major",
     "observation_index": 6,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트의 CARRIED STATE는 이미경이 스마트폰을 들고 있다고 명시했으나, 이미지에서는 화면 왼쪽 전경의 인물(전택수)이 스마트폰을 쥐고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "캐릭터 기준 이미지에 있는 이미경의 왼쪽 가슴 명찰이 생성된 이미지의 이미경 겉옷(분홍색 가디건)에는 누락되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "스마트폰 화면 속 사진을 보면 두 인물 사이(여성의 허리 부근)에 피부색을 띤 정체불명의 형태가 옷과 어색하게 융합되어 묘사되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "이미경이 스마트폰을 쥐고 있지 않고 왼쪽 인물이 들고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 시선이 화면을 내려다보지 않고 왼쪽 인물 쪽을 향한다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수의 머리 뒷모습이 왼쪽 전경을 크게 차지해 주변부 어깨·손이 아니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경이 프레임 대부분·중앙을 차지하는 얼굴 클로즈업이 아니라 오른쪽에 치우친 투샷이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "캐릭터 기준 이미지에 있는 이미경의 왼쪽 가슴 명찰이 생성된 이미지의 이미경 겉옷(분홍색 가디건)에는 누락되어 있습니다.",
    "fix_en": "Add a small rectangular yellow name tag to the left chest area of Mi-gyeong's pink cardigan, preserving her face, clothing, the smartphone, and the background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "스마트폰 화면 속 사진을 보면 두 인물 사이(여성의 허리 부근)에 피부색을 띤 정체불명의 형태가 옷과 어색하게 융합되어 묘사되었습니다.",
    "fix_en": "Remove the fleshy artifact between the two figures on the smartphone screen by painting over it with their respective clothing, preserving the rest of the screen content, the holding hand, Mi-gyeong's face, and the background.",
    "severity": "minor",
    "observation_index": 2
   },
   {
    "issue_ko": "전택수의 머리 뒷모습이 왼쪽 전경을 크게 차지해 주변부 어깨·손이 아니다.",
    "fix_en": "Reduce the size of the left character's head and shoulder so they occupy less of the foreground edge, preserving Mi-gyeong's face, clothing, the smartphone, and the background.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "이미경이 프레임 대부분·중앙을 차지하는 얼굴 클로즈업이 아니라 오른쪽에 치우친 투샷이다.",
    "fix_en": "Shift the framing so Mi-gyeong's face occupies the middle-center of the frame as a close-up, preserving her expression, clothing, the smartphone, and the background.",
    "severity": "major",
    "observation_index": 6,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the hand holding the smartphone with Mi-gyeong's hand reaching up from the bottom right to grip the device, completely removing the male hand and arm extending from the left, while strictly preserving Mi-gyeong's face, her pink cardigan, the smartphone's position and screen content, the left character's head and shoulder, and the playground background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "주요 피사체의 안구를 비정상적으로 크게 왜곡하여 '해부학적 변형 금지'라는 프롬프트의 명시적 지시를 정면으로 위반했으며, 폰을 쥔 손에도 다지가 묘사되는 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (인물의 안구가 만화처럼 비정상적으로 확대되어 프롬프트 금지 조항 위반) / [gemini-pro] 물리적으로 불가능한 해부학 (스마트폰을 쥔 손의 오른쪽 아래 끄트머리에 추가적인 엄지손가락 형태가 묘사됨)"
     },
     {
      "label": "B",
      "score": 1625,
      "verdict_ko": "이미경의 기준 이미지와 표정 지시(해부학적 변형 없는 놀란 연기)를 훌륭하게 구현했으나, 전택수의 오른팔에 왼손이 달려 있는 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (전택수의 오른쪽 어깨에서 이어진 전경의 팔 끝에, 엄지가 오른쪽에 위치한 '왼손'이 달려 폰을 쥐고 있음)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.875
     },
     "adjusted": {
      "A": 1.0,
      "B": 1.625
     },
     "violations": {
      "A": [
       "[gemini-pro] 물리적으로 불가능한 해부학 (인물의 안구가 만화처럼 비정상적으로 확대되어 프롬프트 금지 조항 위반)",
       "[gemini-pro] 물리적으로 불가능한 해부학 (스마트폰을 쥔 손의 오른쪽 아래 끄트머리에 추가적인 엄지손가락 형태가 묘사됨)"
      ],
      "B": [
       "[gemini-pro] 물리적으로 불가능한 해부학 (전택수의 오른쪽 어깨에서 이어진 전경의 팔 끝에, 엄지가 오른쪽에 위치한 '왼손'이 달려 폰을 쥐고 있음)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.125,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "주요 피사체의 안구를 비정상적으로 크게 왜곡하여 '해부학적 변형 금지'라는 프롬프트의 명시적 지시를 정면으로 위반했으며, 폰을 쥔 손에도 다지가 묘사되는 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (인물의 안구가 만화처럼 비정상적으로 확대되어 프롬프트 금지 조항 위반) / [gemini-pro] 물리적으로 불가능한 해부학 (스마트폰을 쥔 손의 오른쪽 아래 끄트머리에 추가적인 엄지손가락 형태가 묘사됨)"
     },
     {
      "label": "B",
      "score": 1625,
      "verdict_ko": "이미경의 기준 이미지와 표정 지시(해부학적 변형 없는 놀란 연기)를 훌륭하게 구현했으나, 전택수의 오른팔에 왼손이 달려 있는 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 해부학 (전택수의 오른쪽 어깨에서 이어진 전경의 팔 끝에, 엄지가 오른쪽에 위치한 '왼손'이 달려 폰을 쥐고 있음)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "지정된 구도와 이미경의 충격받은 표정을 훌륭하게 구현했으며, 시선 처리와 스마트폰을 쥐고 있는 손의 묘사도 물리적으로 자연스럽습니다."
     },
     {
      "label": "A",
      "score": 1194,
      "verdict_ko": "인물의 표정과 배경의 분위기는 좋으나, 스마트폰을 쥐고 있는 손가락이 기괴하게 분열되어 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 스마트폰을 쥐고 있는 손의 해부학적 구조가 불가능함 (화면 양쪽과 하단에서 손가락이 비정상적인 각도로 돋아나 있음)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.444,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.194,
      "B": 1.714
     },
     "violations": {
      "A": [
       "[gemini-pro] 스마트폰을 쥐고 있는 손의 해부학적 구조가 불가능함 (화면 양쪽과 하단에서 손가락이 비정상적인 각도로 돋아나 있음)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "지정된 구도와 이미경의 충격받은 표정을 훌륭하게 구현했으며, 시선 처리와 스마트폰을 쥐고 있는 손의 묘사도 물리적으로 자연스럽습니다."
     },
     {
      "label": "B",
      "score": 1194,
      "verdict_ko": "인물의 표정과 배경의 분위기는 좋으나, 스마트폰을 쥐고 있는 손가락이 기괴하게 분열되어 치명적인 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 스마트폰을 쥐고 있는 손의 해부학적 구조가 불가능함 (화면 양쪽과 하단에서 손가락이 비정상적인 각도로 돋아나 있음)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2714,
     "B": 2819
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S79sh3__bgfirst_bg.png",
   "bg_asset_id": "515fdbd4-bdb5-4ac5-9ccd-013cc2b6af64",
   "bg_record_key": "S79sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S79sh3::cine": {
  "applied": true,
  "fingerprint": "4a1b496dbb72938452e46ea274ce076c6d392b75cc6496ef0d9bbac6432af0be",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S79sh3_sel.png",
  "source_sha256": "49a2497a833b3dbad10cfd1bc180f29737ea72d07e1584b9e4e3f7eb819065aa",
  "file": "S79sh3_cine.png",
  "latency_ms": 9931
 },
 "S79sh4::signage": {
  "fp": "8e7642ce1fa60161",
  "inscriptions": [
   {
    "surface_native": "놀이터 이용안내판",
    "text_native": "어린이 놀이터 이용안내",
    "reason_ko": "유치원 놀이터라는 공간적 배경을 사실적으로 나타내기 위해 배경의 안내판에 들어갈 표기입니다."
   }
  ]
 },
 "S79sh4": {
  "input_fingerprint": "9f9541159a168e71",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손바닥으로 전택수의 스마트폰을 밀어낸 채 차가운 눈빛으로 쏘아보는 이미경의 측면.\n\nLOCATION (lock): At one side of the kindergarten playground during the officers' conversation. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside 이미경, continue the lateral track into a slightly opened profile frame that holds her cold stare, outstretched palm, and the phone being pressed back toward 전택수. 이미경 occupies the near side while 전택수 remains across the narrow midground gap, his hand yielding around the phone as both focus on each other rather than the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이미경 in the middle-left of the frame, foreground, reaches for 전택수's phone hand; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Held between their opposing gestures) — The screen side remains turned generally toward 이미경 while her palm pushes it back; used as Small contact point between their hands, preserving the refusal action without overwhelming either figure; 유치원 놀이터 (Visible behind 이미경 and 전택수); used as Provides restrained location context behind the interpersonal standoff.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light rendered in a subdued neutral-to-cool palette with realistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong pushes Taeksu's smartphone back toward him; the alibi photograph remains displayed on its screen.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 이미경 right now — 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 놀이터 이용안내판: \"어린이 놀이터 이용안내\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손바닥으로 전택수의 스마트폰을 밀어낸 채 차가운 눈빛으로 쏘아보는 이미경의 측면.\n\nLOCATION (lock): At one side of the kindergarten playground during the officers' conversation. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside 이미경, continue the lateral track into a slightly opened profile frame that holds her cold stare, outstretched palm, and the phone being pressed back toward 전택수. 이미경 occupies the near side while 전택수 remains across the narrow midground gap, his hand yielding around the phone as both focus on each other rather than the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이미경 in the middle-left of the frame, foreground, reaches for 전택수's phone hand; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Held between their opposing gestures) — The screen side remains turned generally toward 이미경 while her palm pushes it back; used as Small contact point between their hands, preserving the refusal action without overwhelming either figure; 유치원 놀이터 (Visible behind 이미경 and 전택수); used as Provides restrained location context behind the interpersonal standoff.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light rendered in a subdued neutral-to-cool palette with realistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong pushes Taeksu's smartphone back toward him; the alibi photograph remains displayed on its screen.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 이미경 right now — 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 놀이터 이용안내판: \"어린이 놀이터 이용안내\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 손바닥으로 전택수의 스마트폰을 밀어낸 채 차가운 눈빛으로 쏘아보는 이미경의 측면.\n\nLOCATION (lock): At one side of the kindergarten playground during the officers' conversation. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside 이미경, continue the lateral track into a slightly opened profile frame that holds her cold stare, outstretched palm, and the phone being pressed back toward 전택수. 이미경 occupies the near side while 전택수 remains across the narrow midground gap, his hand yielding around the phone as both focus on each other rather than the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이미경 in the middle-left of the frame, foreground, reaches for 전택수's phone hand; 전택수 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 전택수의 스마트폰 (Held between their opposing gestures) — The screen side remains turned generally toward 이미경 while her palm pushes it back; used as Small contact point between their hands, preserving the refusal action without overwhelming either figure; 유치원 놀이터 (Visible behind 이미경 and 전택수); used as Provides restrained location context behind the interpersonal standoff.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light rendered in a subdued neutral-to-cool palette with realistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong pushes Taeksu's smartphone back toward him; the alibi photograph remains displayed on its screen.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 이미경 right now — 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 놀이터 이용안내판: \"어린이 놀이터 이용안내\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.778
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.778
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 1778
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "요청된 '약간 열린 측면' 앵글과 차가운 눈빛을 완벽하게 구현했으며, 이전 샷과 전택수의 의상(베이지 재킷) 연속성까지 잘 유지함."
   },
   {
    "label": "B",
    "score": 1778,
    "verdict_ko": "지시된 앵글보다 완전한 측면에 가깝게 렌더링되었으며, 이전 샷에 단서가 있던 전택수의 의상이 임의로 변경되어 연속성이 다소 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S79sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이미경의 손바닥이 스마트폰에 닿지 않고 허공에 떠 있어, 프롬프트에 명시된 두 사람 손 사이의 접촉(Small contact point)과 스마트폰을 밀어내는 동작이 전혀 구현되지 않음.",
     "fix_en": "Redraw the woman's raised right hand and forearm to extend further right so her open palm firmly contacts and presses against the left edge of the smartphone, physically pushing it. Preserve the woman's face, hair, and clothing, the man's face, pose, and clothing, the smartphone's screen content, and the playground background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "이미경의 표정이 차가운 눈빛이 아니라 눈을 크게 뜨고 입을 벌린 당황·걱정 표정이다",
     "fix_en": "Change the woman's expression by closing her mouth and narrowing her eyes into a cold, piercing stare. Preserve her identity, hair, clothing, the man, the smartphone, and the background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "스마트폰 화면이 이미경 쪽이 아니라 카메라 정면을 향하고 있다",
     "fix_en": "Rotate the smartphone so its screen faces generally toward the woman on the left, showing the display to her at an oblique angle rather than flat to the camera. Preserve the people, their clothing, the hand gripping the phone, and the playground background.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이미경의 손바닥이 스마트폰에 닿지 않고 허공에 떠 있어, 프롬프트에 명시된 두 사람 손 사이의 접촉(Small contact point)과 스마트폰을 밀어내는 동작이 전혀 구현되지 않음.",
     "severity": "critical"
    },
    {
     "issue_ko": "이미경의 손바닥이 스마트폰에 닿지 않고 떨어져 밀어내는 접촉이 없다",
     "severity": "critical"
    },
    {
     "issue_ko": "이미경의 표정이 차가운 눈빛이 아니라 눈을 크게 뜨고 입을 벌린 당황·걱정 표정이다",
     "severity": "major"
    },
    {
     "issue_ko": "스마트폰 화면이 이미경 쪽이 아니라 카메라 정면을 향하고 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "이미경의 표정이 차가운 눈빛이 아니라 눈을 크게 뜨고 입을 벌린 당황·걱정 표정이다",
    "fix_en": "Change the woman's expression by closing her mouth and narrowing her eyes into a cold, piercing stare. Preserve her identity, hair, clothing, the man, the smartphone, and the background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "스마트폰 화면이 이미경 쪽이 아니라 카메라 정면을 향하고 있다",
    "fix_en": "Rotate the smartphone so its screen faces generally toward the woman on the left, showing the display to her at an oblique angle rather than flat to the camera. Preserve the people, their clothing, the hand gripping the phone, and the playground background.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the woman's raised right hand and forearm to extend further right so her open palm firmly contacts and presses against the left edge of the smartphone, physically pushing it. Preserve the woman's face, hair, and clothing, the man's face, pose, and clothing, the smartphone's screen content, and the playground background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "스마트폰을 손바닥으로 밀어내는 물리적 접촉 지점을 정확히 연출하여 프롬프트의 핵심 동작을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "이미경의 손과 스마트폰 사이에 공간이 있어 '스마트폰을 밀어낸 채'라는 핵심 접촉 동작 연출에 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경은 전택수를 응시하며 손바닥을 스마트폰 쪽으로 향하고 있으나 목표물에 닿지 않음.",
      "built_space": "지정된 유치원 놀이터 배경과 '어린이 놀이터 이용안내' 표지판이 올바르게 배치됨.",
      "entities": "이미경의 외모 및 의상 레퍼런스 일치, 전택수 형태 유지. 스마트폰 화면에 이전 샷의 사진이 표시됨.",
      "hard_violations": [],
      "physics": "이미경의 손이 스마트폰에 닿지 않고 허공에 떠 있어 밀어내는 힘이 작용하지 않음. 전택수가 스마트폰을 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "이미경이 전택수를 차갑게 응시하며 손바닥을 스마트폰에 정확히 밀착시키고 있음.",
      "built_space": "지정된 유치원 놀이터 배경과 '어린이 놀이터 이용안내' 표지판이 올바르게 배치됨.",
      "entities": "이미경의 외모 및 의상 레퍼런스 일치, 전택수 형태 유지. 스마트폰 화면에 이전 샷의 사진이 표시됨.",
      "hard_violations": [],
      "physics": "이미경의 손바닥이 스마트폰 모서리에 물리적으로 맞닿아 밀어내는 동작의 지지점과 접촉이 성립함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "스마트폰을 손바닥으로 밀어내는 물리적 접촉 지점을 정확히 연출하여 프롬프트의 핵심 동작을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "이미경의 손과 스마트폰 사이에 공간이 있어 '스마트폰을 밀어낸 채'라는 핵심 접촉 동작 연출에 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "이미경은 전택수를 응시하며 손바닥을 스마트폰 쪽으로 향하고 있으나 목표물에 닿지 않음.",
      "built_space": "지정된 유치원 놀이터 배경과 '어린이 놀이터 이용안내' 표지판이 올바르게 배치됨.",
      "entities": "이미경의 외모 및 의상 레퍼런스 일치, 전택수 형태 유지. 스마트폰 화면에 이전 샷의 사진이 표시됨.",
      "hard_violations": [],
      "physics": "이미경의 손이 스마트폰에 닿지 않고 허공에 떠 있어 밀어내는 힘이 작용하지 않음. 전택수가 스마트폰을 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "이미경이 전택수를 차갑게 응시하며 손바닥을 스마트폰에 정확히 밀착시키고 있음.",
      "built_space": "지정된 유치원 놀이터 배경과 '어린이 놀이터 이용안내' 표지판이 올바르게 배치됨.",
      "entities": "이미경의 외모 및 의상 레퍼런스 일치, 전택수 형태 유지. 스마트폰 화면에 이전 샷의 사진이 표시됨.",
      "hard_violations": [],
      "physics": "이미경의 손바닥이 스마트폰 모서리에 물리적으로 맞닿아 밀어내는 동작의 지지점과 접촉이 성립함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "이미경의 손이 스마트폰에 닿지 않아 '밀어내는' 동작이 물리적으로 구현되지 않았으나, 인물의 외형, 텍스트, 배경 등은 지시사항을 잘 반영했습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "A와 동일하게 손 접촉이 누락되었으며, 남성 캐릭터의 머리카락 윗부분이 부자연스럽게 검게 덧칠되어 시각적 품질이 저하되었습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경과 전택수는 서로의 얼굴을 응시함. 스마트폰 화면은 이미경을 향함.",
      "built_space": "유치원 놀이터 배경에 슬라이드와 놀이기구가 보이며, 안내판이 정확한 텍스트로 올바르게 위치함.",
      "entities": "이미경은 레퍼런스 인상 및 의상과 일치함. 전택수는 회색 셔츠를 입고 있음. 안내판에 '어린이 놀이터 이용안내' 텍스트가 정확함. 화면에 두 사람의 사진이 보임.",
      "hard_violations": [],
      "physics": "전택수의 손이 스마트폰을 쥐어 지지하고 있으나, 이미경의 손은 스마트폰에 닿지 않은 채 공중에 떠 있어 밀어내는 접촉이 발생하지 않음."
     },
     {
      "label": "B",
      "direction": "이미경과 전택수는 서로의 얼굴을 응시함. 스마트폰 화면은 이미경을 향함.",
      "built_space": "유치원 놀이터 배경 및 안내판 위치가 A와 동일하게 적절함.",
      "entities": "이미경은 레퍼런스와 일치함. 전택수와 스마트폰 화면, 안내판 텍스트 모두 A와 동일하나 전택수의 머리카락 텍스처가 뭉개져 있음.",
      "hard_violations": [],
      "physics": "전택수의 손이 스마트폰을 지지하고 있으며, 이미경의 손은 스마트폰과 접촉하지 않고 공중에 뻗어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "이미경의 손이 스마트폰에 닿지 않아 '밀어내는' 동작이 물리적으로 구현되지 않았으나, 인물의 외형, 텍스트, 배경 등은 지시사항을 잘 반영했습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "A와 동일하게 손 접촉이 누락되었으며, 남성 캐릭터의 머리카락 윗부분이 부자연스럽게 검게 덧칠되어 시각적 품질이 저하되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "이미경과 전택수는 서로의 얼굴을 응시함. 스마트폰 화면은 이미경을 향함.",
      "built_space": "유치원 놀이터 배경에 슬라이드와 놀이기구가 보이며, 안내판이 정확한 텍스트로 올바르게 위치함.",
      "entities": "이미경은 레퍼런스 인상 및 의상과 일치함. 전택수는 회색 셔츠를 입고 있음. 안내판에 '어린이 놀이터 이용안내' 텍스트가 정확함. 화면에 두 사람의 사진이 보임.",
      "hard_violations": [],
      "physics": "전택수의 손이 스마트폰을 쥐어 지지하고 있으나, 이미경의 손은 스마트폰에 닿지 않은 채 공중에 떠 있어 밀어내는 접촉이 발생하지 않음."
     },
     {
      "label": "A",
      "direction": "이미경과 전택수는 서로의 얼굴을 응시함. 스마트폰 화면은 이미경을 향함.",
      "built_space": "유치원 놀이터 배경 및 안내판 위치가 A와 동일하게 적절함.",
      "entities": "이미경은 레퍼런스와 일치함. 전택수와 스마트폰 화면, 안내판 텍스트 모두 A와 동일하나 전택수의 머리카락 텍스처가 뭉개져 있음.",
      "hard_violations": [],
      "physics": "전택수의 손이 스마트폰을 지지하고 있으며, 이미경의 손은 스마트폰과 접촉하지 않고 공중에 뻗어 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 10,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S79sh3"
  }
 },
 "S79sh4::cine": {
  "applied": true,
  "fingerprint": "6dcf4c350a32b7bcfc7a6df52842e5cece0f68494310cd363fd9a791f50e6b20",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S79sh4_sel.png",
  "source_sha256": "d3234743d27a88e7b90e4d6c1a1dd7d0ee546fe88c419dc2bb1c38f6e4943fab",
  "file": "S79sh4_cine.png",
  "latency_ms": 12122
 },
 "S79sh8::signage": {
  "fp": "b62b64fbaae08838",
  "inscriptions": []
 },
 "S79sh8": {
  "input_fingerprint": "0cbed93cbe3755de",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 손으로 얼굴을 감싸 쥔 채 어깨를 잔뜩 움츠린 이미경의 상체.\n\nLOCATION (lock): In a quiet corner of the outdoor kindergarten playground, with children playing farther away. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly above 이미경's head level, complete the intimate dolly-in at a front-side three-quarter angle while widening only enough to contain both hands, her covered face, and the full contraction of her shoulders. Her folded upper body occupies the central foreground, with no other person intruding as she concentrates entirely on shielding her face while crying.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 유치원 놀이터 (Visible in shallow background); used as Held softly behind her as minimal spatial context before the cut to the distant children.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light remains restrained and neutral-to-cool, treating her breakdown with soft, non-sensational realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 손으로 얼굴을 감싸 쥔 채 어깨를 잔뜩 움츠린 이미경의 상체.\n\nLOCATION (lock): In a quiet corner of the outdoor kindergarten playground, with children playing farther away. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly above 이미경's head level, complete the intimate dolly-in at a front-side three-quarter angle while widening only enough to contain both hands, her covered face, and the full contraction of her shoulders. Her folded upper body occupies the central foreground, with no other person intruding as she concentrates entirely on shielding her face while crying.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 유치원 놀이터 (Visible in shallow background); used as Held softly behind her as minimal spatial context before the cut to the distant children.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light remains restrained and neutral-to-cool, treating her breakdown with soft, non-sensational realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 두 손으로 얼굴을 감싸 쥔 채 어깨를 잔뜩 움츠린 이미경의 상체.\n\nLOCATION (lock): In a quiet corner of the outdoor kindergarten playground, with children playing farther away. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly above 이미경's head level, complete the intimate dolly-in at a front-side three-quarter angle while widening only enough to contain both hands, her covered face, and the full contraction of her shoulders. Her folded upper body occupies the central foreground, with no other person intruding as she concentrates entirely on shielding her face while crying.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 이미경 in the middle-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 유치원 놀이터 (Visible in shallow background); used as Held softly behind her as minimal spatial context before the cut to the distant children.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime exterior light remains restrained and neutral-to-cool, treating her breakdown with soft, non-sensational realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "얼굴을 양손으로 완전히 감싼 채 시선이 아래를 향하고 있음.",
    "built_space": "이전 샷에 등장한 안내 표지판, 미끄럼틀, 유치원 건물이 올바른 위치와 비례로 배경에 배치됨.",
    "entities": "이미경은 지정된 분홍색 가디건과 흰색 블라우스를 입고 있으며, 배경에 흐릿한 아이들이 보임.",
    "hard_violations": [],
    "physics": "바닥에 선 채로 상체를 숙여 두 손으로 얼굴을 지탱하는 자세가 자연스럽고 안정적임."
   },
   {
    "label": "B",
    "direction": "양손으로 얼굴을 덮고 있으며 약간 왼쪽을 향해 있음.",
    "built_space": "놀이터 배경은 존재하나 카메라가 지시된 '머리 위'가 아닌 가슴 높이에 위치함.",
    "entities": "이미경은 참조 의상을 착용했으나, 좌측 배경의 아이가 주인공과 완전히 동일한 색상의 가디건과 치마를 입고 있음.",
    "hard_violations": [
     "물리적으로 불가능한 형태의 눈물 (딱딱한 유리줄기 같은 액체 묘사)",
     "주인공의 고유 의상이 배경 인물에게 복제됨 (leaked markers)"
    ],
    "physics": "앉아있는 자세는 지시를 벗어났으며, 뺨을 타고 흐르는 눈물이 물리적으로 불가능한 고체 형태로 떠 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "지시된 머리 위 카메라 앵글과 웅크린 상체의 포즈를 훌륭하게 구현했으며, 이전 샷의 장소와 자연스럽게 연결됩니다."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "물리적으로 불가능한 형태의 눈물이 묘사되었고, 배경 인물에게 주인공의 의상이 복제되는 치명적인 오류가 있습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S79sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 두 손으로 얼굴을 완전히 덮어 가린 상태를 지시했으나, 생성된 이미지에서는 눈과 코를 비롯한 얼굴 대부분이 그대로 노출되어 있습니다.",
     "fix_en": "Raise her hands to completely cover her eyes, nose, and mouth. Preserve her hair, pink cardigan, mint skirt, the playground background, the sign, and the distant children.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "눈 밑에 흐르는 눈물이 실제 액체가 아니라 피부 위에 그려진 흰색 점선(그래픽 오버레이)처럼 잘못 렌더링되었습니다.",
     "fix_en": "Remove the white dotted graphic marks on her cheeks, replacing them with natural, glossy liquid tear streaks. Maintain her face, hands, clothing, and all background elements unchanged.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "레퍼런스 이미지의 핑크색 가디건 왼쪽 가슴에 있던 명찰이 이 샷에서는 누락되었습니다.",
     "fix_en": "Add the reference's small rectangular name tag to the left chest of her pink cardigan.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "이미경이 프레임 정중앙이 아니라 왼쪽에 치우쳐 있다",
     "fix_en": "Crop the right side of the frame to center the character.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "양손·가린 얼굴·움츠린 어깨만 담을 타이트한 숏이 아니라 놀이터 배경이 과하게 넓다",
     "fix_en": "Crop the frame tighter to focus only on her hands, covered face, and shoulders.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 두 손으로 얼굴을 완전히 덮어 가린 상태를 지시했으나, 생성된 이미지에서는 눈과 코를 비롯한 얼굴 대부분이 그대로 노출되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "눈 밑에 흐르는 눈물이 실제 액체가 아니라 피부 위에 그려진 흰색 점선(그래픽 오버레이)처럼 잘못 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "레퍼런스 이미지의 핑크색 가디건 왼쪽 가슴에 있던 명찰이 이 샷에서는 누락되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경이 프레임 정중앙이 아니라 왼쪽에 치우쳐 있다",
     "severity": "major"
    },
    {
     "issue_ko": "양손·가린 얼굴·움츠린 어깨만 담을 타이트한 숏이 아니라 놀이터 배경이 과하게 넓다",
     "severity": "major"
    },
    {
     "issue_ko": "배경에 뛰어노는 아이 두 명이 보여 다른 사람이 없어야 할 숏을 어겼다",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "레퍼런스 이미지의 핑크색 가디건 왼쪽 가슴에 있던 명찰이 이 샷에서는 누락되었습니다.",
    "fix_en": "Add the reference's small rectangular name tag to the left chest of her pink cardigan.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "이미경이 프레임 정중앙이 아니라 왼쪽에 치우쳐 있다",
    "fix_en": "Crop the right side of the frame to center the character.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "양손·가린 얼굴·움츠린 어깨만 담을 타이트한 숏이 아니라 놀이터 배경이 과하게 넓다",
    "fix_en": "Crop the frame tighter to focus only on her hands, covered face, and shoulders.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Raise her hands to completely cover her eyes, nose, and mouth. Preserve her hair, pink cardigan, mint skirt, the playground background, the sign, and the distant children.\n- Remove the white dotted graphic marks on her cheeks, replacing them with natural, glossy liquid tear streaks. Maintain her face, hands, clothing, and all background elements unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1833,
      "verdict_ko": "요구된 샷 크기, 카메라 앵글, 얼굴을 감싸고 움츠린 자세를 완벽히 구현했으며, 이전 샷의 배경 연속성도 훌륭하게 유지했습니다."
     },
     {
      "label": "A",
      "score": 1150,
      "verdict_ko": "지정된 구도와 배경의 연속성은 잘 따랐으나, 화면 우측에 위치한 인물의 손가락 개수가 비정상적으로 묘사되는 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 화면 우측에 보이는 인물의 왼손 손가락이 기형적으로 많이 그려져 있음 (physically impossible anatomy)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.4,
      "B": 1.833
     },
     "adjusted": {
      "A": 1.15,
      "B": 1.833
     },
     "violations": {
      "A": [
       "[gemini-pro] 화면 우측에 보이는 인물의 왼손 손가락이 기형적으로 많이 그려져 있음 (physically impossible anatomy)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.167,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1833,
      "verdict_ko": "요구된 샷 크기, 카메라 앵글, 얼굴을 감싸고 움츠린 자세를 완벽히 구현했으며, 이전 샷의 배경 연속성도 훌륭하게 유지했습니다."
     },
     {
      "label": "A",
      "score": 1150,
      "verdict_ko": "지정된 구도와 배경의 연속성은 잘 따랐으나, 화면 우측에 위치한 인물의 손가락 개수가 비정상적으로 묘사되는 치명적인 해부학적 오류가 있습니다.  ★위반: [gemini-pro] 화면 우측에 보이는 인물의 왼손 손가락이 기형적으로 많이 그려져 있음 (physically impossible anatomy)"
     }
    ],
    "all_candidates_fail": false
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1714,
      "verdict_ko": "가슴의 명찰이 누락되고 왼손에 지시되지 않은 반지가 추가되었으나, 해부학적 오류 없이 웅크려 우는 인물의 감정과 지정된 앵글을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 1125,
      "verdict_ko": "화면 우측에 보이는 왼손의 손가락이 6개로 묘사되는 치명적인 해부학적 오류가 있어 실격입니다.  ★위반: [gemini-pro] 화면 우측에 위치한 인물의 왼손에 엄지를 제외하고도 바깥으로 보이는 손가락이 5개로 그려져, 총 6개의 손가락을 가진 불가능한 해부학적 묘사가 발생함."
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.375,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.125,
      "B": 1.714
     },
     "violations": {
      "A": [
       "[gemini-pro] 화면 우측에 위치한 인물의 왼손에 엄지를 제외하고도 바깥으로 보이는 손가락이 5개로 그려져, 총 6개의 손가락을 가진 불가능한 해부학적 묘사가 발생함."
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.286,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1714,
      "verdict_ko": "가슴의 명찰이 누락되고 왼손에 지시되지 않은 반지가 추가되었으나, 해부학적 오류 없이 웅크려 우는 인물의 감정과 지정된 앵글을 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 1125,
      "verdict_ko": "화면 우측에 보이는 왼손의 손가락이 6개로 묘사되는 치명적인 해부학적 오류가 있어 실격입니다.  ★위반: [gemini-pro] 화면 우측에 위치한 인물의 왼손에 엄지를 제외하고도 바깥으로 보이는 손가락이 5개로 그려져, 총 6개의 손가락을 가진 불가능한 해부학적 묘사가 발생함."
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2864,
     "B": 2958
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S79sh4"
  }
 },
 "S79sh8::cine": {
  "applied": true,
  "fingerprint": "5e5b22baed093d91948beac0d4264ee866ebbb81c901ecaf950f1b01d2341b46",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S79sh8_sel.png",
  "source_sha256": "35e03627bc1801cefa721e752b69ed2dfb276e2d4962eda9bd76fbf6aae28fc2",
  "file": "S79sh8_cine.png",
  "latency_ms": 11266
 },
 "S80sh4::signage": {
  "fp": "c40c78fe9e0112f8",
  "inscriptions": [
   {
    "surface_native": "프레젠테이션 스크린",
    "text_native": "공소사실 요지",
    "reason_ko": "법정 내 검사 발표용 화면에 공소사실의 핵심 내용을 설명하는 제목이 표시되어 재판 상황의 사실성을 높입니다."
   }
  ]
 },
 "S80sh4": {
  "input_fingerprint": "9529a165aeffeaae",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 쪽을 향해 손바닥을 펼친 채 확신에 찬 표정으로 입을 크게 벌린 장원섭의 상체.\n\nLOCATION (lock): At the prosecution presentation area inside the courtroom, beside the document camera and evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track into a medium view from slightly below 장원섭's eye level on his screen-side three-quarter angle, placing him in the right midground and retaining part of the screen on the left. His torso turns toward the court while his open palm extends toward the displayed satellite photograph, and his emphatic gaze follows the argument toward the screen rather than meeting the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-right of the frame, midground, points to courtroom screen; courtroom screen in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: 법정 스크린 (Displaying the satellite photograph) — Its display face is visible at an oblique angle, showing the satellite photograph with the case site, 지국현's house, the grandmother's house, and their marked distances; used as Visible beside 장원섭 as the evidentiary anchor for his open-palm explanation; 검사석 (Occupied by 장원섭 during his statement) — The side nearest the screen faces the camera at an oblique angle; used as Grounds 장원섭's formal position in the courtroom without obscuring his gesture.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime courtroom illumination with a subdued neutral-to-cool grade, moderate contrast, and procedural clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness area, daylight, and prosecutor's formal clothing. Exclude the defense attorney and alibi photograph, showing the prosecutor presenting the route map.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The satellite photograph remains projected on the courtroom screen, marked with the crime scene, Ji Gukhyeon's home, and his grandmother's home.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 프레젠테이션 스크린: \"공소사실 요지\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 쪽을 향해 손바닥을 펼친 채 확신에 찬 표정으로 입을 크게 벌린 장원섭의 상체.\n\nLOCATION (lock): At the prosecution presentation area inside the courtroom, beside the document camera and evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track into a medium view from slightly below 장원섭's eye level on his screen-side three-quarter angle, placing him in the right midground and retaining part of the screen on the left. His torso turns toward the court while his open palm extends toward the displayed satellite photograph, and his emphatic gaze follows the argument toward the screen rather than meeting the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-right of the frame, midground, points to courtroom screen; courtroom screen in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: 법정 스크린 (Displaying the satellite photograph) — Its display face is visible at an oblique angle, showing the satellite photograph with the case site, 지국현's house, the grandmother's house, and their marked distances; used as Visible beside 장원섭 as the evidentiary anchor for his open-palm explanation; 검사석 (Occupied by 장원섭 during his statement) — The side nearest the screen faces the camera at an oblique angle; used as Grounds 장원섭's formal position in the courtroom without obscuring his gesture.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime courtroom illumination with a subdued neutral-to-cool grade, moderate contrast, and procedural clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness area, daylight, and prosecutor's formal clothing. Exclude the defense attorney and alibi photograph, showing the prosecutor presenting the route map.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The satellite photograph remains projected on the courtroom screen, marked with the crime scene, Ji Gukhyeon's home, and his grandmother's home.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 프레젠테이션 스크린: \"공소사실 요지\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 스크린 쪽을 향해 손바닥을 펼친 채 확신에 찬 표정으로 입을 크게 벌린 장원섭의 상체.\n\nLOCATION (lock): At the prosecution presentation area inside the courtroom, beside the document camera and evidence screen. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track into a medium view from slightly below 장원섭's eye level on his screen-side three-quarter angle, placing him in the right midground and retaining part of the screen on the left. His torso turns toward the court while his open palm extends toward the displayed satellite photograph, and his emphatic gaze follows the argument toward the screen rather than meeting the lens.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-right of the frame, midground, points to courtroom screen; courtroom screen in the middle-left of the frame, background.\n- KEY BACKGROUND ELEMENTS: 법정 스크린 (Displaying the satellite photograph) — Its display face is visible at an oblique angle, showing the satellite photograph with the case site, 지국현's house, the grandmother's house, and their marked distances; used as Visible beside 장원섭 as the evidentiary anchor for his open-palm explanation; 검사석 (Occupied by 장원섭 during his statement) — The side nearest the screen faces the camera at an oblique angle; used as Grounds 장원섭's formal position in the courtroom without obscuring his gesture.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime courtroom illumination with a subdued neutral-to-cool grade, moderate contrast, and procedural clarity.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, large screen, witness area, daylight, and prosecutor's formal clothing. Exclude the defense attorney and alibi photograph, showing the prosecutor presenting the route map.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The satellite photograph remains projected on the courtroom screen, marked with the crime scene, Ji Gukhyeon's home, and his grandmother's home.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 프레젠테이션 스크린: \"공소사실 요지\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "장원섭의 시선과 길게 뻗은 오른손 바닥이 화면 좌측의 법정 스크린(위성 사진)을 정확히 향하고 있음.",
    "built_space": "법정 내부. 화면 좌측 중경에 대형 스크린이 있고, 좌측 하단에 '검사석' 명패가 놓인 책상이 위치함.",
    "entities": "장원섭의 인상착의(짧은 검은 머리, 40대 남성, 정장)가 캐릭터 레퍼런스와 일치함. 스크린 상단에 '공소사실 요지' 텍스트와 지도 위 마커들이 렌더링됨.",
    "hard_violations": [],
    "physics": "인물이 몸을 바로 세운 채 서 있으며, 허공으로 뻗은 오른팔의 자세와 근육 긴장이 자연스럽게 유지됨."
   },
   {
    "label": "A",
    "direction": "장원섭의 시선과 펼친 왼손 바닥이 좌측의 법정 스크린을 향하고 있음.",
    "built_space": "법정 내부. 좌측에 위성 사진 스크린, 전면에 '변호인석' 명패가 있는 책상이 배치됨.",
    "entities": "장원섭의 외모와 의상이 레퍼런스와 부합함. 스크린 텍스트 '공소사실 요지' 및 지도 마커가 식별됨.",
    "hard_violations": [],
    "physics": "오른손은 책상에 얹혀 몸을 지탱하고, 왼손은 스크린 쪽으로 뻗어 있어 포즈의 물리적 성립이 자연스러움."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "프롬프트가 요구한 '검사석' 명패를 정확히 반영하였고, 인물의 확신에 찬 표정과 스크린을 향한 제스처가 지시사항에 충실하게 구현되었습니다."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "인물의 제스처와 스크린 배치는 준수하나, 프롬프트에서 명시한 '검사석' 대신 레퍼런스의 '변호인석' 명패를 잘못 렌더링하여 우선순위에서 밀립니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S75sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:796050>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "좌측 스크린의 지도 라벨 중 '지국현 집'이 '지국현 켭'으로 오기되어 있으며, '할머니 집' 라벨이 불필요하게 두 번 중복 표기됨.",
     "fix_en": "Replace the map's small label text with illegible blurred marks to hide the typo, and remove the duplicate label, preserving the main title, map image, prosecutor's face, suit, pose, and set.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "우측 배경의 나무 단상 정면에 부착되어 있던 '증인석' 명패가 레퍼런스 이미지와 달리 누락됨.",
     "fix_en": "Add a black nameplate with unreadable blurred white text to the witness stand, preserving the prosecutor, screen, and all other courtroom details.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "오른쪽 인물이 캐릭터 레퍼런스 장원섭이 아니고 이전 샷과 같은 얼굴·체형이다",
     "fix_en": "Change the prosecutor's face and hair to match the Jang Won-seop reference, strictly preserving his expression, open-palm pose, dark suit, the screen, and the courtroom setting.",
     "severity": "critical",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "좌측 스크린의 지도 라벨 중 '지국현 집'이 '지국현 켭'으로 오기되어 있으며, '할머니 집' 라벨이 불필요하게 두 번 중복 표기됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "우측 배경의 나무 단상 정면에 부착되어 있던 '증인석' 명패가 레퍼런스 이미지와 달리 누락됨.",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 인물이 캐릭터 레퍼런스 장원섭이 아니고 이전 샷과 같은 얼굴·체형이다",
     "severity": "critical"
    },
    {
     "issue_ko": "장원섭이 레퍼런스의 검은 법복과 흰 jabot가 아닌 양복과 파란 넥타이를 입고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 스크린 위성지도에 할머니 집 표기가 두 곳으로 중복되어 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "우측 배경의 나무 단상 정면에 부착되어 있던 '증인석' 명패가 레퍼런스 이미지와 달리 누락됨.",
    "fix_en": "Add a black nameplate with unreadable blurred white text to the witness stand, preserving the prosecutor, screen, and all other courtroom details.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the map's small label text with illegible blurred marks to hide the typo, and remove the duplicate label, preserving the main title, map image, prosecutor's face, suit, pose, and set.\n- Change the prosecutor's face and hair to match the Jang Won-seop reference, strictly preserving his expression, open-palm pose, dark suit, the screen, and the courtroom setting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1167,
      "verdict_ko": "두 후보 모두 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사하는 치명적 오류를 범했으나, A는 스크린의 지도 마커 텍스트를 명확하게 렌더링하여 B보다 우수합니다.  ★위반: [gemini-pro] 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사함 / [openrouter:x-ai/grok-4.6] 스크린에 허용 목록 밖 가독 텍스트(사건 현장, 지국현 집, 할머니 집 중복) 누출"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "두 후보 모두 인물 신원 렌더링에 실패했으며, B는 스크린의 지도 마커 텍스트를 부자연스럽게 블러 처리하여 품질이 더 떨어집니다.  ★위반: [gemini-pro] 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사함"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.667,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.167,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[gemini-pro] 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사함",
       "[openrouter:x-ai/grok-4.6] 스크린에 허용 목록 밖 가독 텍스트(사건 현장, 지국현 집, 할머니 집 중복) 누출"
      ],
      "B": [
       "[gemini-pro] 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사함"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.333,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1167,
      "verdict_ko": "두 후보 모두 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사하는 치명적 오류를 범했으나, A는 스크린의 지도 마커 텍스트를 명확하게 렌더링하여 B보다 우수합니다.  ★위반: [gemini-pro] 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사함 / [openrouter:x-ai/grok-4.6] 스크린에 허용 목록 밖 가독 텍스트(사건 현장, 지국현 집, 할머니 집 중복) 누출"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "두 후보 모두 인물 신원 렌더링에 실패했으며, B는 스크린의 지도 마커 텍스트를 부자연스럽게 블러 처리하여 품질이 더 떨어집니다.  ★위반: [gemini-pro] 캐릭터 레퍼런스가 아닌 이전 샷 스틸의 인물 얼굴을 복사함"
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1100,
      "verdict_ko": "지정된 텍스트 외 추가 텍스트 금지 지침을 준수해 지도 마커를 흐리게 처리했으며, 인물의 제스처와 표정을 정확히 구현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 캐릭터 레퍼런스의 장원섭이 아닌 이전 샷 얼굴의 다른 인물 / [openrouter:x-ai/grok-4.6] 스크린 위 검열 블러 박스(그래픽 오버레이)"
     },
     {
      "label": "B",
      "score": 1464,
      "verdict_ko": "스크린 지도에 금지된 추가 텍스트('사건 현장' 등)를 임의로 삽입하여 지시사항을 위반했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 캐릭터 레퍼런스의 장원섭이 아닌 이전 샷 얼굴의 다른 인물"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.6,
      "B": 1.714
     },
     "adjusted": {
      "A": 1.1,
      "B": 1.464
     },
     "violations": {
      "A": [
       "[openrouter:x-ai/grok-4.6] 캐릭터 레퍼런스의 장원섭이 아닌 이전 샷 얼굴의 다른 인물",
       "[openrouter:x-ai/grok-4.6] 스크린 위 검열 블러 박스(그래픽 오버레이)"
      ],
      "B": [
       "[openrouter:x-ai/grok-4.6] 캐릭터 레퍼런스의 장원섭이 아닌 이전 샷 얼굴의 다른 인물"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.4,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1100,
      "verdict_ko": "지정된 텍스트 외 추가 텍스트 금지 지침을 준수해 지도 마커를 흐리게 처리했으며, 인물의 제스처와 표정을 정확히 구현했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 캐릭터 레퍼런스의 장원섭이 아닌 이전 샷 얼굴의 다른 인물 / [openrouter:x-ai/grok-4.6] 스크린 위 검열 블러 박스(그래픽 오버레이)"
     },
     {
      "label": "A",
      "score": 1464,
      "verdict_ko": "스크린 지도에 금지된 추가 텍스트('사건 현장' 등)를 임의로 삽입하여 지시사항을 위반했습니다.  ★위반: [openrouter:x-ai/grok-4.6] 캐릭터 레퍼런스의 장원섭이 아닌 이전 샷 얼굴의 다른 인물"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 2631,
     "B": 2600
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S75sh6"
  }
 },
 "S80sh4::cine": {
  "applied": true,
  "fingerprint": "64f36bfafabd074989739767f1110e15f08f1f2ed5d58db1c318c14b10cffd05",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S80sh4_sel.png",
  "source_sha256": "f57d00777acb3312d3224847b2fb89bd871f47714c4e0e2f4156ea9cb820d6e6",
  "file": "S80sh4_cine.png",
  "latency_ms": 11437
 },
 "S81sh1::signage": {
  "fp": "9cfc08471ebc8c57",
  "inscriptions": []
 },
 "S81sh1": {
  "input_fingerprint": "482bfab6a756de3a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 책이 빼곡한 연구실 창가에 서서 유리창에 비친 자신의 모습을 물끄러미 바라보는 전택수의 뒷모습\n\nLOCATION (lock): At the window inside a book-lined forensic institute research office at night, with the interior lit enough to reflect him in the glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a high position several meters behind 전택수, begin the slow dolly-in with his back low in the central midground and his reflected face held within the window ahead, preserving the research room as a wide field around his isolation. He leans subtly toward the glass and looks at his own reflection, while the nearby snow globe remains a small secondary shape beside the window in preparation for the descending arc.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-center of the frame, midground, looks toward his reflection in the window; window holding 전택수's reflection in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 연구실 유리창 (전택수's reflection is visible) — The surface reflects 전택수's face and upper body back toward the high rear camera; used as Primary visual anchor ahead of 전택수, carrying the reflected counterpart to his back-facing figure; 책이 빼곡한 연구실 책장 (Packed with books) — The book-filled faces of the shelving angle inward around the window area; used as Builds dense research-room context along the sides and depth of the frame; 스노우볼 (Resting beside the window) — The castle model inside the round glass globe is visible from the room side; used as Small secondary object near the window that will enter the camera's next viewing axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to a research room at night, kept subdued with a neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The snow globe containing a grand castle and a girl at its upper window remains beside the laboratory window, available for Taeksu to pick up.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 책이 빼곡한 연구실 창가에 서서 유리창에 비친 자신의 모습을 물끄러미 바라보는 전택수의 뒷모습\n\nLOCATION (lock): At the window inside a book-lined forensic institute research office at night, with the interior lit enough to reflect him in the glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a high position several meters behind 전택수, begin the slow dolly-in with his back low in the central midground and his reflected face held within the window ahead, preserving the research room as a wide field around his isolation. He leans subtly toward the glass and looks at his own reflection, while the nearby snow globe remains a small secondary shape beside the window in preparation for the descending arc.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-center of the frame, midground, looks toward his reflection in the window; window holding 전택수's reflection in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 연구실 유리창 (전택수's reflection is visible) — The surface reflects 전택수's face and upper body back toward the high rear camera; used as Primary visual anchor ahead of 전택수, carrying the reflected counterpart to his back-facing figure; 책이 빼곡한 연구실 책장 (Packed with books) — The book-filled faces of the shelving angle inward around the window area; used as Builds dense research-room context along the sides and depth of the frame; 스노우볼 (Resting beside the window) — The castle model inside the round glass globe is visible from the room side; used as Small secondary object near the window that will enter the camera's next viewing axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to a research room at night, kept subdued with a neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The snow globe containing a grand castle and a girl at its upper window remains beside the laboratory window, available for Taeksu to pick up.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 책이 빼곡한 연구실 창가에 서서 유리창에 비친 자신의 모습을 물끄러미 바라보는 전택수의 뒷모습\n\nLOCATION (lock): At the window inside a book-lined forensic institute research office at night, with the interior lit enough to reflect him in the glass. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From a high position several meters behind 전택수, begin the slow dolly-in with his back low in the central midground and his reflected face held within the window ahead, preserving the research room as a wide field around his isolation. He leans subtly toward the glass and looks at his own reflection, while the nearby snow globe remains a small secondary shape beside the window in preparation for the descending arc.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the lower-center of the frame, midground, looks toward his reflection in the window; window holding 전택수's reflection in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: 연구실 유리창 (전택수's reflection is visible) — The surface reflects 전택수's face and upper body back toward the high rear camera; used as Primary visual anchor ahead of 전택수, carrying the reflected counterpart to his back-facing figure; 책이 빼곡한 연구실 책장 (Packed with books) — The book-filled faces of the shelving angle inward around the window area; used as Builds dense research-room context along the sides and depth of the frame; 스노우볼 (Resting beside the window) — The castle model inside the round glass globe is visible from the room side; used as Small secondary object near the window that will enter the camera's next viewing axis.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to a research room at night, kept subdued with a neutral-to-cool grade and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The snow globe containing a grand castle and a girl at its upper window remains beside the laboratory window, available for Taeksu to pick up.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "피사체는 창문 쪽을 향해 시선을 두고 있으며, 창문에 비친 얼굴이 피사체를 마주보고 있음.",
    "built_space": "높은 앵글에서 피사체의 등 뒤를 촬영함. 양옆으로 책장이 있고, 왼쪽 하단에 이전 샷과 연결되는 서류가 놓인 책상이 보이며 중앙 창턱에 스노우볼이 위치함.",
    "entities": "전택수의 뒷모습(경찰 재킷, 수사 모자 착용), 반사된 전택수의 정면(얼굴 일치, 모자 착용), 스노우볼(성 모형 포함).",
    "hard_violations": [
     "불가능한 반사: 유리창에 비친 인물의 모자와 의상 텍스트가 좌우 반전되지 않고 정방향으로 읽힘"
    ],
    "physics": "인물은 바닥에 서서 지지됨. 그러나 창문에 비친 모습의 텍스트가 반전되지 않아 거울 반사 물리 법칙에 위배됨."
   },
   {
    "label": "B",
    "direction": "피사체가 창문을 바라보고 있으며, 반사된 얼굴이 피사체를 향함.",
    "built_space": "피사체 뒤에서 촬영된 구도로 양옆에 책장이, 오른쪽에 작은 수납장이 위치함.",
    "entities": "전택수의 뒷모습(모자 미착용), 반사된 전택수(모자 착용), 스노우볼.",
    "hard_violations": [
     "불가능한 반사: 실제 뒷모습에는 모자가 없으나 반사된 모습에는 모자가 씌워져 있음",
     "불가능한 반사: 반사된 의상과 모자의 텍스트가 좌우 반전되지 않음"
    ],
    "physics": "바닥을 딛고 서 있으나, 모자의 유무 불일치 및 텍스트 비반전으로 인해 물리적으로 불가능한 반사 상태임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 5,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "이전 샷의 위치적 특징(서류 등)과 전체적인 프레이밍은 잘 구현되었으나, 창문에 비친 글자가 좌우 반전되지 않은 물리적 오류로 인해 아쉬움."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "실제 인물의 뒷모습엔 모자가 없으나 반사된 이미지엔 모자가 존재하는 등 물리적으로 불가능한 반사 오류가 명백함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S29sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "왼쪽 책상 위 '부검소견서' 서류 표지 하단에 레퍼런스에 없는 만화적인 인물 그림이 임의로 추가됨.",
     "fix_en": "Remove the cartoon illustration from the document cover, replacing it with a plain blank paper surface. Preserve the document's printed text and all other items on the desk.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "스노우볼의 나무 받침대 정면에 의미를 알 수 없는 가상의 문자가 생성됨.",
     "fix_en": "Remove the illegible text from the wooden base of the snow globe, leaving a clean, polished dark wood surface. Preserve the snow globe's internal details and the surrounding window ledge.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "창문 속 인물이 서 있는 전택수와 좌우로 어긋나 있고 모자·재킷 글자가 거울처럼 반전되지 않아 광학적으로 불가능한 반사이다",
     "fix_en": "Replace the figure appearing outside the window with a true optical reflection of the foreground character rendered directly on the glass: align the reflected face with his line of sight, mirror-reverse the yellow text on the cap and the white text on the jacket, and make the image slightly semi-transparent against the dark night. Preserve the foreground character's pose and clothing, the books, the snow globe, and the desk exactly as they are.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "전택수가 화면 하단 중앙이 아니라 좌측으로 치우쳐 있다",
     "fix_en": "Move the foreground character from the left to the lower-center of the midground. Maintain his scale and posture.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "유리창 속 모습의 모자 글씨('수사')가 좌우 반전되지 않고 시선 각도가 맞지 않아 반사가 아닌 창밖의 사람처럼 보임.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 책상 위 '부검소견서' 서류 표지 하단에 레퍼런스에 없는 만화적인 인물 그림이 임의로 추가됨.",
     "severity": "major"
    },
    {
     "issue_ko": "스노우볼의 나무 받침대 정면에 의미를 알 수 없는 가상의 문자가 생성됨.",
     "severity": "major"
    },
    {
     "issue_ko": "창문 속 인물이 서 있는 전택수와 좌우로 어긋나 있고 모자·재킷 글자가 거울처럼 반전되지 않아 광학적으로 불가능한 반사이다",
     "severity": "critical"
    },
    {
     "issue_ko": "이전 스틸의 소파·창호·선반 등 연구실 고정 구조와 다른 공간으로 보인다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 화면 하단 중앙이 아니라 좌측으로 치우쳐 있다",
     "severity": "major"
    },
    {
     "issue_ko": "유리창 쪽으로 살짝 기대지 않고 양팔을 내린 채 곧게 서 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 책상 위 '부검소견서' 서류 표지 하단에 레퍼런스에 없는 만화적인 인물 그림이 임의로 추가됨.",
    "fix_en": "Remove the cartoon illustration from the document cover, replacing it with a plain blank paper surface. Preserve the document's printed text and all other items on the desk.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "스노우볼의 나무 받침대 정면에 의미를 알 수 없는 가상의 문자가 생성됨.",
    "fix_en": "Remove the illegible text from the wooden base of the snow globe, leaving a clean, polished dark wood surface. Preserve the snow globe's internal details and the surrounding window ledge.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "전택수가 화면 하단 중앙이 아니라 좌측으로 치우쳐 있다",
    "fix_en": "Move the foreground character from the left to the lower-center of the midground. Maintain his scale and posture.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the figure appearing outside the window with a true optical reflection of the foreground character rendered directly on the glass: align the reflected face with his line of sight, mirror-reverse the yellow text on the cap and the white text on the jacket, and make the image slightly semi-transparent against the dark night. Preserve the foreground character's pose and clothing, the books, the snow globe, and the desk exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "창문에 비친 모습의 텍스트(SALKS, POLICE)와 좌우 배치가 반전되지 않아 물리적으로 불가능한 반사 이미지를 생성하여 치명적 오류입니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "창문에 비친 모습의 텍스트와 좌우 배치를 물리 법칙에 맞게 정확히 반전시켜 프롬프트의 요구사항을 훌륭하게 구현했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수가 창문을 향해 서서 유리에 비친 자신의 얼굴을 바라봄.",
      "built_space": "책이 꽂힌 책장, 서류가 있는 책상, 창문이 있는 야간의 연구실.",
      "entities": "전택수의 뒷모습, 비친 얼굴, 스노우볼, 책. 의상과 인물 외형 일치함.",
      "hard_violations": [
       "창문에 비친 텍스트와 옷의 배치가 거울상으로 반전되지 않아 물리적으로 불가능한 반사 이미지임"
      ],
      "physics": "인물은 바닥에 서 있으나 반사된 이미지가 물리 법칙에 위배됨."
     },
     {
      "label": "B",
      "direction": "전택수가 창문을 향해 서서 유리에 비친 자신의 얼굴을 바라봄.",
      "built_space": "책이 꽂힌 책장, 서류가 있는 책상, 창문이 있는 야간의 연구실.",
      "entities": "전택수의 뒷모습, 반사된 얼굴(텍스트 반전됨), 스노우볼, 책. 의상과 인물 외형 일치함.",
      "hard_violations": [],
      "physics": "인물은 바닥에 서 있으며 반사된 이미지가 거울상 물리 법칙을 정확히 따름."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "창문에 비친 모습의 텍스트(SALKS, POLICE)와 좌우 배치가 반전되지 않아 물리적으로 불가능한 반사 이미지를 생성하여 치명적 오류입니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "창문에 비친 모습의 텍스트와 좌우 배치를 물리 법칙에 맞게 정확히 반전시켜 프롬프트의 요구사항을 훌륭하게 구현했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수가 창문을 향해 서서 유리에 비친 자신의 얼굴을 바라봄.",
      "built_space": "책이 꽂힌 책장, 서류가 있는 책상, 창문이 있는 야간의 연구실.",
      "entities": "전택수의 뒷모습, 비친 얼굴, 스노우볼, 책. 의상과 인물 외형 일치함.",
      "hard_violations": [
       "창문에 비친 텍스트와 옷의 배치가 거울상으로 반전되지 않아 물리적으로 불가능한 반사 이미지임"
      ],
      "physics": "인물은 바닥에 서 있으나 반사된 이미지가 물리 법칙에 위배됨."
     },
     {
      "label": "B",
      "direction": "전택수가 창문을 향해 서서 유리에 비친 자신의 얼굴을 바라봄.",
      "built_space": "책이 꽂힌 책장, 서류가 있는 책상, 창문이 있는 야간의 연구실.",
      "entities": "전택수의 뒷모습, 반사된 얼굴(텍스트 반전됨), 스노우볼, 책. 의상과 인물 외형 일치함.",
      "hard_violations": [],
      "physics": "인물은 바닥에 서 있으며 반사된 이미지가 거울상 물리 법칙을 정확히 따름."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 875,
      "verdict_ko": "점퍼의 텍스트가 거울상으로 정상 반전되어 물리 법칙을 준수했으며, 요구된 야간 연구실의 분위기, 인물 배치 및 구도를 매우 충실하게 구현함.  ★위반: [openrouter:x-ai/grok-4.6] 창 속 인물이 광학적 반사가 아닌 정면 의상 배치의 추가 인물로 읽힘 / [openrouter:x-ai/grok-4.6] 명시된 후방 카메라에서 성립할 수 없는 유리 반사"
     },
     {
      "label": "B",
      "score": 1179,
      "verdict_ko": "텍스트를 가독성 있게 수정하려다 유리에 비친 글자가 전혀 반전되지 않고 정방향으로 읽히는 치명적인 물리적 불가능 오류를 발생시킴.  ★위반: [gemini-pro] 유리에 비친 인물의 옷과 모자 텍스트가 거울상으로 반전되지 않고 정방향으로 읽히는 물리적 불가능 오류"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.375,
      "B": 1.429
     },
     "adjusted": {
      "A": 0.875,
      "B": 1.179
     },
     "violations": {
      "B": [
       "[gemini-pro] 유리에 비친 인물의 옷과 모자 텍스트가 거울상으로 반전되지 않고 정방향으로 읽히는 물리적 불가능 오류"
      ],
      "A": [
       "[openrouter:x-ai/grok-4.6] 창 속 인물이 광학적 반사가 아닌 정면 의상 배치의 추가 인물로 읽힘",
       "[openrouter:x-ai/grok-4.6] 명시된 후방 카메라에서 성립할 수 없는 유리 반사"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.625,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 875,
      "verdict_ko": "점퍼의 텍스트가 거울상으로 정상 반전되어 물리 법칙을 준수했으며, 요구된 야간 연구실의 분위기, 인물 배치 및 구도를 매우 충실하게 구현함.  ★위반: [openrouter:x-ai/grok-4.6] 창 속 인물이 광학적 반사가 아닌 정면 의상 배치의 추가 인물로 읽힘 / [openrouter:x-ai/grok-4.6] 명시된 후방 카메라에서 성립할 수 없는 유리 반사"
     },
     {
      "label": "A",
      "score": 1179,
      "verdict_ko": "텍스트를 가독성 있게 수정하려다 유리에 비친 글자가 전혀 반전되지 않고 정방향으로 읽히는 치명적인 물리적 불가능 오류를 발생시킴.  ★위반: [gemini-pro] 유리에 비친 인물의 옷과 모자 텍스트가 거울상으로 반전되지 않고 정방향으로 읽히는 물리적 불가능 오류"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1182,
     "B": 882
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": false,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S29sh4"
  }
 },
 "S81sh1::cine": {
  "applied": true,
  "fingerprint": "84a7c48ad52a435fccbc3dc60d0ca05a4524f5c883af5df5bf767c0deeab9866",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S81sh1_sel.png",
  "source_sha256": "0a50faa1d5d42ecf1f11dad503d01476ac2e3481ab0e927648ea228c722d9a3d",
  "file": "S81sh1_cine.png",
  "latency_ms": 12557
 },
 "S81sh6::signage": {
  "fp": "63a8d1ad9635881c",
  "inscriptions": []
 },
 "S81sh6": {
  "input_fingerprint": "9b8e65f23704b086",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 소파에 몸을 깊숙이 묻은 채 지친 기색으로 기대어 앉아 있는 조일남의 상체\n\nLOCATION (lock): In the sofa seating area of the book-lined forensic research office at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The track finishes slightly above 조일남's seated eye line at medium distance, holding his upper body in the center-right as he sinks deeply into the sofa. His shoulders collapse into the backrest and his lowered gaze remains on the floor, while the densely packed books give the frame a compressed, work-worn depth.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: sofa (Occupied by 조일남) — The seating side opens diagonally toward the camera and the unseen opposite seat; used as Supports 조일남's collapsed seated posture and anchors the conversational area; book-filled research room shelving (Densely filled with books) — The shelving faces the room and is seen obliquely behind 조일남; used as Forms a dense background behind the exhausted figure without drawing focus.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime laboratory illumination with a restrained neutral-to-cool palette and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same book-filled research room, window reflections, low night lighting, and sofa placement. Exclude the officer at the window and show the exhausted expert sunk into the sofa.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu continues holding the castle snow globe while seated opposite the exhausted Jo Ilnam.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 소파에 몸을 깊숙이 묻은 채 지친 기색으로 기대어 앉아 있는 조일남의 상체\n\nLOCATION (lock): In the sofa seating area of the book-lined forensic research office at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The track finishes slightly above 조일남's seated eye line at medium distance, holding his upper body in the center-right as he sinks deeply into the sofa. His shoulders collapse into the backrest and his lowered gaze remains on the floor, while the densely packed books give the frame a compressed, work-worn depth.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: sofa (Occupied by 조일남) — The seating side opens diagonally toward the camera and the unseen opposite seat; used as Supports 조일남's collapsed seated posture and anchors the conversational area; book-filled research room shelving (Densely filled with books) — The shelving faces the room and is seen obliquely behind 조일남; used as Forms a dense background behind the exhausted figure without drawing focus.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime laboratory illumination with a restrained neutral-to-cool palette and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same book-filled research room, window reflections, low night lighting, and sofa placement. Exclude the officer at the window and show the exhausted expert sunk into the sofa.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu continues holding the castle snow globe while seated opposite the exhausted Jo Ilnam.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 소파에 몸을 깊숙이 묻은 채 지친 기색으로 기대어 앉아 있는 조일남의 상체\n\nLOCATION (lock): In the sofa seating area of the book-lined forensic research office at night. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The track finishes slightly above 조일남's seated eye line at medium distance, holding his upper body in the center-right as he sinks deeply into the sofa. His shoulders collapse into the backrest and his lowered gaze remains on the floor, while the densely packed books give the frame a compressed, work-worn depth.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: sofa (Occupied by 조일남) — The seating side opens diagonally toward the camera and the unseen opposite seat; used as Supports 조일남's collapsed seated posture and anchors the conversational area; book-filled research room shelving (Densely filled with books) — The shelving faces the room and is seen obliquely behind 조일남; used as Forms a dense background behind the exhausted figure without drawing focus.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime laboratory illumination with a restrained neutral-to-cool palette and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same book-filled research room, window reflections, low night lighting, and sofa placement. Exclude the officer at the window and show the exhausted expert sunk into the sofa.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu continues holding the castle snow globe while seated opposite the exhausted Jo Ilnam.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.571,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.321,
    "B": 1.417
   },
   "violations": {
    "A": [
     "[openrouter:x-ai/grok-4.6] 샷 텍스트·PEOPLE가 허용하지 않은 추가 인물이 프레임에 들어옴"
    ],
    "B": [
     "[openrouter:x-ai/grok-4.6] 샷 텍스트·PEOPLE가 허용하지 않은 추가 인물이 프레임에 들어옴"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1417,
   "A": 1321
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1417,
    "verdict_ko": "조일남의 의상과 외모가 레퍼런스와 정확히 일치하며 지친 상체의 프레이밍과 분위기를 훌륭히 구현함.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트·PEOPLE가 허용하지 않은 추가 인물이 프레임에 들어옴"
   },
   {
    "label": "A",
    "score": 1321,
    "verdict_ko": "프레이밍과 소품은 적절하나, 조일남의 의상(흰 가운)이 레퍼런스와 완전히 불일치함.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트·PEOPLE가 허용하지 않은 추가 인물이 프레임에 들어옴"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S81sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 조일남: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878922>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트의 'PEOPLE' 목록과 숏 텍스트에 지정되지 않은 인물(스노우글로브를 든 사람)이 화면 왼쪽 앞에 렌즈를 등지고 등장함",
     "fix_en": "Remove the person and the snow globe in the left foreground entirely, filling the space with the out-of-focus wooden desk and empty air. Preserve Jo Ilnam, his pose, his clothing, the sofa, and the bookshelf background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "조일남의 가슴에 달린 신분증에 의미를 알 수 없는 왜곡된 글자가 임의로 적혀 있음",
     "fix_en": "Replace the legible, garbled text on the ID badge on Jo Ilnam's chest with indistinct, blurred markings that cannot be read. Preserve Jo Ilnam's face, his clothing, the sofa, and the background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "스노우글로브 명판에 레퍼런스에 없는 '유 원'이라는 글자가 임의로 추가됨",
     "fix_en": "Replace the clear Korean characters on the snow globe's wooden plaque with a dark, unreadable surface. Preserve the snow globe's glass and interior, the hands holding it, Jo Ilnam, and the background.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "레퍼런스의 이전 숏 스틸과 달리 조일남의 머리 바로 뒤쪽 창문에 굵은 수직 창틀이 추가됨",
     "fix_en": "Remove the thick vertical window mullion behind Jo Ilnam's head, replacing it with a single continuous pane of glass reflecting the night interior. Preserve Jo Ilnam, his clothing, the sofa, and the bookshelves.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "조일남의 시선이 바닥이 아니라 왼쪽의 스노우글로브 쪽으로 향해 있다",
     "fix_en": "Adjust Jo Ilnam's eyes so his gaze is directed downward toward the floor. Preserve his head position, facial features, clothing, and the surrounding environment.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트의 'PEOPLE' 목록과 숏 텍스트에 지정되지 않은 인물(스노우글로브를 든 사람)이 화면 왼쪽 앞에 렌즈를 등지고 등장함",
     "severity": "critical"
    },
    {
     "issue_ko": "조일남의 가슴에 달린 신분증에 의미를 알 수 없는 왜곡된 글자가 임의로 적혀 있음",
     "severity": "critical"
    },
    {
     "issue_ko": "스노우글로브 명판에 레퍼런스에 없는 '유 원'이라는 글자가 임의로 추가됨",
     "severity": "critical"
    },
    {
     "issue_ko": "레퍼런스의 이전 숏 스틸과 달리 조일남의 머리 바로 뒤쪽 창문에 굵은 수직 창틀이 추가됨",
     "severity": "major"
    },
    {
     "issue_ko": "샷 텍스트에 없는 인물이 왼쪽 전경에 상반신·손으로 등장해 스노우글로브를 들고 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "조일남의 시선이 바닥이 아니라 왼쪽의 스노우글로브 쪽으로 향해 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "레퍼런스의 이전 숏 스틸과 달리 조일남의 머리 바로 뒤쪽 창문에 굵은 수직 창틀이 추가됨",
    "fix_en": "Remove the thick vertical window mullion behind Jo Ilnam's head, replacing it with a single continuous pane of glass reflecting the night interior. Preserve Jo Ilnam, his clothing, the sofa, and the bookshelves.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "조일남의 시선이 바닥이 아니라 왼쪽의 스노우글로브 쪽으로 향해 있다",
    "fix_en": "Adjust Jo Ilnam's eyes so his gaze is directed downward toward the floor. Preserve his head position, facial features, clothing, and the surrounding environment.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the person and the snow globe in the left foreground entirely, filling the space with the out-of-focus wooden desk and empty air. Preserve Jo Ilnam, his pose, his clothing, the sofa, and the bookshelf background.\n- Replace the legible, garbled text on the ID badge on Jo Ilnam's chest with indistinct, blurred markings that cannot be read. Preserve Jo Ilnam's face, his clothing, the sofa, and the background.\n- Replace the clear Korean characters on the snow globe's wooden plaque with a dark, unreadable surface. Preserve the snow globe's glass and interior, the hands holding it, Jo Ilnam, and the background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지문이 요구하는 조일남의 미디엄 샷과 지친 자세를 정확히 구현했으며, 프레임에 허용되지 않은 인물을 올바르게 배제하여 제약을 훌륭히 준수했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지문과 PEOPLE 목록에 명시되지 않은 인물(스노우글로브를 든 사람)을 화면에 포함시켜 엄격한 등장인물 제한 규정을 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "조일남의 시선은 아래쪽 바닥을 자연스럽게 향하고 있습니다.",
      "built_space": "야간의 연구실 배경으로 창문, 책장, 소파가 참조 이미지와 동일하게 정확히 배치되었습니다.",
      "entities": "조일남의 얼굴과 복장이 캐릭터 참조와 일치하며, 지시대로 프레임 내에 다른 인물은 없습니다.",
      "hard_violations": [],
      "physics": "인물의 체중이 소파 등받이에 무겁게 기대어 안착된 상태입니다."
     },
     {
      "label": "A",
      "direction": "조일남은 바닥을 향해 시선을 두고 있으며, 왼쪽의 인물은 스노우글로브와 조일남을 향해 있습니다.",
      "built_space": "참조된 연구실 공간의 책장, 창문, 소파 배치가 적절하게 구현되었습니다.",
      "entities": "조일남의 외형은 일치하나, PEOPLE 목록에 없는 인물의 어깨와 손이 프레임 왼쪽에 등장합니다.",
      "hard_violations": [
       "PEOPLE 목록에 없는 인물 추가"
      ],
      "physics": "조일남은 소파에 앉아 있고, 추가된 인물의 두 손이 스노우글로브를 지탱하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지문이 요구하는 조일남의 미디엄 샷과 지친 자세를 정확히 구현했으며, 프레임에 허용되지 않은 인물을 올바르게 배제하여 제약을 훌륭히 준수했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "지문과 PEOPLE 목록에 명시되지 않은 인물(스노우글로브를 든 사람)을 화면에 포함시켜 엄격한 등장인물 제한 규정을 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "조일남의 시선은 아래쪽 바닥을 자연스럽게 향하고 있습니다.",
      "built_space": "야간의 연구실 배경으로 창문, 책장, 소파가 참조 이미지와 동일하게 정확히 배치되었습니다.",
      "entities": "조일남의 얼굴과 복장이 캐릭터 참조와 일치하며, 지시대로 프레임 내에 다른 인물은 없습니다.",
      "hard_violations": [],
      "physics": "인물의 체중이 소파 등받이에 무겁게 기대어 안착된 상태입니다."
     },
     {
      "label": "A",
      "direction": "조일남은 바닥을 향해 시선을 두고 있으며, 왼쪽의 인물은 스노우글로브와 조일남을 향해 있습니다.",
      "built_space": "참조된 연구실 공간의 책장, 창문, 소파 배치가 적절하게 구현되었습니다.",
      "entities": "조일남의 외형은 일치하나, PEOPLE 목록에 없는 인물의 어깨와 손이 프레임 왼쪽에 등장합니다.",
      "hard_violations": [
       "PEOPLE 목록에 없는 인물 추가"
      ],
      "physics": "조일남은 소파에 앉아 있고, 추가된 인물의 두 손이 스노우글로브를 지탱하고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "숏 텍스트에 명시된 조일남 단독 샷을 지친 자세와 억제된 톤으로 정확하게 구현함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "숏 텍스트가 지시하지 않은 추가 인물(택수)을 프레임에 등장시켜 명시적 규칙을 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "조일남의 시선이 바닥을 향함.",
      "built_space": "배경의 책장과 소파 배치가 정확하며 인물이 소파에 깊이 착석함.",
      "entities": "조일남의 외모와 복장이 레퍼런스와 일치하며 추가 인물 없음.",
      "hard_violations": [],
      "physics": "등과 엉덩이를 소파가 안정적으로 지탱함."
     },
     {
      "label": "B",
      "direction": "조일남은 바닥을, 앞쪽 인물은 조일남을 향함.",
      "built_space": "소파와 책장이 올바르게 배치됨.",
      "entities": "조일남은 일치하나 숏 텍스트에 없는 추가 인물과 스노우글로브가 전경에 등장함.",
      "hard_violations": [
       "숏 텍스트에 없는 인물 등장 (extra bodies/invented people)"
      ],
      "physics": "조일남은 소파가 지탱하고, 추가 인물의 손이 스노우글로브를 들고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "숏 텍스트에 명시된 조일남 단독 샷을 지친 자세와 억제된 톤으로 정확하게 구현함."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "숏 텍스트가 지시하지 않은 추가 인물(택수)을 프레임에 등장시켜 명시적 규칙을 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "조일남의 시선이 바닥을 향함.",
      "built_space": "배경의 책장과 소파 배치가 정확하며 인물이 소파에 깊이 착석함.",
      "entities": "조일남의 외모와 복장이 레퍼런스와 일치하며 추가 인물 없음.",
      "hard_violations": [],
      "physics": "등과 엉덩이를 소파가 안정적으로 지탱함."
     },
     {
      "label": "A",
      "direction": "조일남은 바닥을, 앞쪽 인물은 조일남을 향함.",
      "built_space": "소파와 책장이 올바르게 배치됨.",
      "entities": "조일남은 일치하나 숏 텍스트에 없는 추가 인물과 스노우글로브가 전경에 등장함.",
      "hard_violations": [
       "숏 텍스트에 없는 인물 등장 (extra bodies/invented people)"
      ],
      "physics": "조일남은 소파가 지탱하고, 추가 인물의 손이 스노우글로브를 들고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 14
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S81sh1"
  }
 },
 "S81sh6::cine": {
  "applied": true,
  "fingerprint": "4ce604995543d633a50ac2e119e7751bd54a6aa490b07522736a9e0cd628d143",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S81sh6_sel.png",
  "source_sha256": "62938d04553d9a953d6d088debb17e6f0f2dfb633d513a685825cf9b58d82362",
  "file": "S81sh6_cine.png",
  "latency_ms": 9462
 },
 "S81sh14::signage": {
  "fp": "0b29ffb9337f1146",
  "inscriptions": []
 },
 "S81sh14": {
  "input_fingerprint": "781d5fd6102e66ab",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 날카롭게 번뜩이는 눈빛으로 스노우볼을 꽉 쥔 채 조일남을 쏘아보는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): Beside the sofa and window inside the forensic research office, holding the snow globe under indoor light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 조일남's side of the conversational axis, the dolly-in ends slightly below 전택수's eye line on a tight three-quarter close-up. 전택수's face occupies most of the frame as he fixes his sharpened gaze on 조일남, while the snow globe remains at the lower edge beside his clenched hand without obscuring his expression.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: snow globe (Gripped tightly in 전택수's hand) — The camera looks through the round glass toward the castle model and settled artificial snow inside; used as Retained at the lower frame edge as the physical trigger for 전택수's realization; book-filled shelving (Densely filled with books) — Only an oblique, softly focused portion of the shelving faces the camera; used as Provides a subdued research-room context beyond the close focus plane.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime laboratory illumination preserves facial detail while keeping the palette restrained and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same book-lined room, dark window, night lighting, and the glass snow globe. Exclude the expert from the close frame and isolate the officer gripping the snow globe.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu grips the same castle snow globe after shaking and setting it down, using it as the key to his new idea.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 날카롭게 번뜩이는 눈빛으로 스노우볼을 꽉 쥔 채 조일남을 쏘아보는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): Beside the sofa and window inside the forensic research office, holding the snow globe under indoor light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 조일남's side of the conversational axis, the dolly-in ends slightly below 전택수's eye line on a tight three-quarter close-up. 전택수's face occupies most of the frame as he fixes his sharpened gaze on 조일남, while the snow globe remains at the lower edge beside his clenched hand without obscuring his expression.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: snow globe (Gripped tightly in 전택수's hand) — The camera looks through the round glass toward the castle model and settled artificial snow inside; used as Retained at the lower frame edge as the physical trigger for 전택수's realization; book-filled shelving (Densely filled with books) — Only an oblique, softly focused portion of the shelving faces the camera; used as Provides a subdued research-room context beyond the close focus plane.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime laboratory illumination preserves facial detail while keeping the palette restrained and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same book-lined room, dark window, night lighting, and the glass snow globe. Exclude the expert from the close frame and isolate the officer gripping the snow globe.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu grips the same castle snow globe after shaking and setting it down, using it as the key to his new idea.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): night.\n\nSHOT TEXT (authoritative, Korean): 날카롭게 번뜩이는 눈빛으로 스노우볼을 꽉 쥔 채 조일남을 쏘아보는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): Beside the sofa and window inside the forensic research office, holding the snow globe under indoor light. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From near 조일남's side of the conversational axis, the dolly-in ends slightly below 전택수's eye line on a tight three-quarter close-up. 전택수's face occupies most of the frame as he fixes his sharpened gaze on 조일남, while the snow globe remains at the lower edge beside his clenched hand without obscuring his expression.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: snow globe (Gripped tightly in 전택수's hand) — The camera looks through the round glass toward the castle model and settled artificial snow inside; used as Retained at the lower frame edge as the physical trigger for 전택수's realization; book-filled shelving (Densely filled with books) — Only an oblique, softly focused portion of the shelving faces the camera; used as Provides a subdued research-room context beyond the close focus plane.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Subdued nighttime laboratory illumination preserves facial detail while keeping the palette restrained and neutral-to-cool.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same book-lined room, dark window, night lighting, and the glass snow globe. Exclude the expert from the close frame and isolate the officer gripping the snow globe.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu grips the same castle snow globe after shaking and setting it down, using it as the key to his new idea.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.4,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.6,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.6,
    "B": 1.167
   },
   "violations": {
    "B": [
     "[gemini-pro] leaked markers/diagrams/text (재킷 가슴 부위에 의미 불명의 한글 텍스트 생성)",
     "[gemini-pro] physically impossible anatomy (스노우볼을 쥐고 있는 손가락들이 부자연스럽게 융합됨)"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1600,
   "B": 1167
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1600,
    "verdict_ko": "요구된 클로즈업 구도와 스노우볼을 쥔 동작을 잘 구현했으나, 금지된 지시와 달리 이전 컷 인물의 얼굴을 그대로 차용한 점이 아쉽습니다."
   },
   {
    "label": "B",
    "score": 1167,
    "verdict_ko": "재킷에 의미 없는 문자가 생성되었고 손의 해부학적 구조가 어색하며, A와 마찬가지로 이전 컷 인물의 얼굴을 잘못 가져왔습니다.  ★위반: [gemini-pro] leaked markers/diagrams/text (재킷 가슴 부위에 의미 불명의 한글 텍스트 생성) / [gemini-pro] physically impossible anatomy (스노우볼을 쥐고 있는 손가락들이 부자연스럽게 융합됨)"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S81sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "PROP REFERENCE — 성 모형이 든 스노우볼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1309901>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 속 얼굴이 전택수 레퍼런스가 아니라 이전 샷 전문가의 얼굴이다",
     "fix_en": "Replace the man's face to match the character reference, ensuring a squarer jaw and correct facial features, while preserving his head position, the cap, the jacket, the hands, the snow globe, and the library background.",
     "severity": "critical",
     "observation_index": 4
    },
    {
     "issue_ko": "전택수가 조일남을 쏘아보지 않고 왼손의 스노우볼을 내려다보고 있다",
     "fix_en": "Redraw the man's eyes so his pupils gaze straight forward and slightly up at an off-camera target, rather than looking down, while keeping his face shape, head tilt, cap, jacket, the snow globe, and the background exactly as they are.",
     "severity": "critical",
     "observation_index": 5
    },
    {
     "issue_ko": "얼굴이 프레임의 대부분을 차지하는 타이트한 클로즈업이 지정되었으나, 상반신과 팔이 넓게 보이는 미디엄 샷으로 연출됨",
     "fix_en": "Crop the frame to a tight close-up on the face, reducing the visible torso and pushing the snow globe to the lower edge, while maintaining the same character, clothing, and room environment.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "인물이 착용한 모자 정면에 캐릭터 레퍼런스의 '수사' 텍스트 대신 임의의 경찰 마크가 렌더링됨",
     "fix_en": "Replace the gold eagle emblem on the cap with indistinct horizontal gold lettering, keeping the cap's shape, the man's face, the jacket, the snow globe, and the library setting completely unchanged.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "자켓 왼쪽 팔과 가슴에 캐릭터 레퍼런스에 존재하지 않는 형태의 대형 경찰 마크와 패치가 추가됨",
     "fix_en": "Remove the large shield patch on the sleeve and the 'POLICE' text on the chest, filling those areas with plain dark blue jacket fabric, without altering the character's face, hands, the snow globe, or the background.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "인물의 시선이 프롬프트에 지시된 대로 화면 밖 대화 상대를 쏘아보지 않고 손에 든 스노우볼을 내려다보고 있음",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴이 프레임의 대부분을 차지하는 타이트한 클로즈업이 지정되었으나, 상반신과 팔이 넓게 보이는 미디엄 샷으로 연출됨",
     "severity": "major"
    },
    {
     "issue_ko": "인물이 착용한 모자 정면에 캐릭터 레퍼런스의 '수사' 텍스트 대신 임의의 경찰 마크가 렌더링됨",
     "severity": "major"
    },
    {
     "issue_ko": "자켓 왼쪽 팔과 가슴에 캐릭터 레퍼런스에 존재하지 않는 형태의 대형 경찰 마크와 패치가 추가됨",
     "severity": "major"
    },
    {
     "issue_ko": "화면 속 얼굴이 전택수 레퍼런스가 아니라 이전 샷 전문가의 얼굴이다",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수가 조일남을 쏘아보지 않고 왼손의 스노우볼을 내려다보고 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "얼굴이 화면 대부분을 차지하는 타이트 클로즈업이 아니고 상반신과 큰 스노우볼이 화면을 나눈다",
     "severity": "major"
    },
    {
     "issue_ko": "스노우볼이 프레임 하단 가장자리가 아니라 왼쪽 중경에 크게 들려 있다",
     "severity": "major"
    },
    {
     "issue_ko": "모자가 레퍼런스의 ‘수사’ 글자가 아닌 금색 날개 엠블럼이다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "얼굴이 프레임의 대부분을 차지하는 타이트한 클로즈업이 지정되었으나, 상반신과 팔이 넓게 보이는 미디엄 샷으로 연출됨",
    "fix_en": "Crop the frame to a tight close-up on the face, reducing the visible torso and pushing the snow globe to the lower edge, while maintaining the same character, clothing, and room environment.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "인물이 착용한 모자 정면에 캐릭터 레퍼런스의 '수사' 텍스트 대신 임의의 경찰 마크가 렌더링됨",
    "fix_en": "Replace the gold eagle emblem on the cap with indistinct horizontal gold lettering, keeping the cap's shape, the man's face, the jacket, the snow globe, and the library setting completely unchanged.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "자켓 왼쪽 팔과 가슴에 캐릭터 레퍼런스에 존재하지 않는 형태의 대형 경찰 마크와 패치가 추가됨",
    "fix_en": "Remove the large shield patch on the sleeve and the 'POLICE' text on the chest, filling those areas with plain dark blue jacket fabric, without altering the character's face, hands, the snow globe, or the background.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the man's face to match the character reference, ensuring a squarer jaw and correct facial features, while preserving his head position, the cap, the jacket, the hands, the snow globe, and the library background.\n- Redraw the man's eyes so his pupils gaze straight forward and slightly up at an off-camera target, rather than looking down, while keeping his face shape, head tilt, cap, jacket, the snow globe, and the background exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 타이트한 클로즈업 프레이밍과 스노우볼을 꽉 쥔 액션, 전택수의 경찰 의상을 정확히 구현하여 샷 텍스트의 의도를 훌륭하게 살렸습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍 지시를 어겼으며, 제외해야 할 이전 샷 인물의 의상과 신체에 얼굴만 합성했고 스노우볼을 쥐는 핵심 액션도 누락되어 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "화면 오른쪽 앵글 밖(조일남이 있는 위치)을 향해 날카로운 시선을 고정하고 있습니다.",
      "built_space": "야간의 연구실 내부로, 배경에 초점이 흐려진 책장과 창문이 올바르게 배치되어 있습니다.",
      "entities": "전택수(참조 이미지의 얼굴, 남색 경찰 점퍼 일치, 모자의 로고는 다소 변형됨), 스노우볼(내부 성 모형과 디테일이 참조와 일치함).",
      "hard_violations": [],
      "physics": "스노우볼은 전택수의 손에 의해 꽉 쥐어져 공중에 안정적으로 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "정면의 카메라 렌즈 방향을 무표정하게 응시하고 있습니다.",
      "built_space": "이전 샷과 동일한 위치의 소파, 책장, 창문이 있는 야간 연구실입니다.",
      "entities": "전택수(얼굴과 모자는 일치하나, 이전 샷 인물의 의상인 흰색 의사 가운과 넥타이를 입고 있음), 스노우볼(참조와 일치함).",
      "hard_violations": [
       "클로즈업 프레이밍을 무시하고 미디엄 샷으로 렌더링됨",
       "제외되어야 할 이전 샷 인물의 의상과 신체를 그대로 차용함",
       "스노우볼을 손에 쥐어야 하는 지시를 무시하고 소파 위에 방치함"
      ],
      "physics": "스노우볼은 손에 들려 있지 않고 소파 쿠션 위에 놓여 지탱되고 있으며, 두 손은 허벅지 위에 가만히 얹혀 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 타이트한 클로즈업 프레이밍과 스노우볼을 꽉 쥔 액션, 전택수의 경찰 의상을 정확히 구현하여 샷 텍스트의 의도를 훌륭하게 살렸습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍 지시를 어겼으며, 제외해야 할 이전 샷 인물의 의상과 신체에 얼굴만 합성했고 스노우볼을 쥐는 핵심 액션도 누락되어 실패했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "화면 오른쪽 앵글 밖(조일남이 있는 위치)을 향해 날카로운 시선을 고정하고 있습니다.",
      "built_space": "야간의 연구실 내부로, 배경에 초점이 흐려진 책장과 창문이 올바르게 배치되어 있습니다.",
      "entities": "전택수(참조 이미지의 얼굴, 남색 경찰 점퍼 일치, 모자의 로고는 다소 변형됨), 스노우볼(내부 성 모형과 디테일이 참조와 일치함).",
      "hard_violations": [],
      "physics": "스노우볼은 전택수의 손에 의해 꽉 쥐어져 공중에 안정적으로 지탱되고 있습니다."
     },
     {
      "label": "B",
      "direction": "정면의 카메라 렌즈 방향을 무표정하게 응시하고 있습니다.",
      "built_space": "이전 샷과 동일한 위치의 소파, 책장, 창문이 있는 야간 연구실입니다.",
      "entities": "전택수(얼굴과 모자는 일치하나, 이전 샷 인물의 의상인 흰색 의사 가운과 넥타이를 입고 있음), 스노우볼(참조와 일치함).",
      "hard_violations": [
       "클로즈업 프레이밍을 무시하고 미디엄 샷으로 렌더링됨",
       "제외되어야 할 이전 샷 인물의 의상과 신체를 그대로 차용함",
       "스노우볼을 손에 쥐어야 하는 지시를 무시하고 소파 위에 방치함"
      ],
      "physics": "스노우볼은 손에 들려 있지 않고 소파 쿠션 위에 놓여 지탱되고 있으며, 두 손은 허벅지 위에 가만히 얹혀 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지시된 타이트한 클로즈업 프레이밍, 인물의 고유 복장, 스노우볼을 쥔 동작 및 강렬한 시선을 모두 정확하게 구현하여 압도적으로 우수합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍을 완전히 무시했고, 절대로 가져오지 말라고 명시된 이전 샷의 의사 가운을 입고 있으며 스노우볼도 쥐고 있지 않습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선이 화면 밖의 대상을 향해 날카롭게 고정되어 있음.",
      "built_space": "흐릿하게 처리된 책장과 창밖 야경이 배경에 위치함.",
      "entities": "전택수의 외모와 경찰 복장이 레퍼런스와 일치하며, 스노우볼 역시 디테일이 정확함.",
      "hard_violations": [],
      "physics": "손이 스노우볼의 하단을 단단히 붙잡고 지지하고 있음."
     },
     {
      "label": "A",
      "direction": "시선이 정면을 멍하게 향하고 있어 특정 대상을 쏘아보지 않음.",
      "built_space": "이전 샷의 소파, 데스크, 책장, 창문이 동일한 구도로 배치됨.",
      "entities": "모자는 전택수의 것이나 복장은 이전 샷 의사의 가운과 넥타이를 그대로 입고 있음. 스노우볼은 소파 위에 방치됨.",
      "hard_violations": [
       "지시된 클로즈업 프레이밍 위반 (미디엄 샷)",
       "이전 샷 인물의 복장 차용 금지 조항 위반",
       "스노우볼을 손에 쥐고 있지 않음"
      ],
      "physics": "스노우볼이 손이 아닌 소파 방석 위에 지지되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 타이트한 클로즈업 프레이밍, 인물의 고유 복장, 스노우볼을 쥔 동작 및 강렬한 시선을 모두 정확하게 구현하여 압도적으로 우수합니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "클로즈업 프레이밍을 완전히 무시했고, 절대로 가져오지 말라고 명시된 이전 샷의 의사 가운을 입고 있으며 스노우볼도 쥐고 있지 않습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선이 화면 밖의 대상을 향해 날카롭게 고정되어 있음.",
      "built_space": "흐릿하게 처리된 책장과 창밖 야경이 배경에 위치함.",
      "entities": "전택수의 외모와 경찰 복장이 레퍼런스와 일치하며, 스노우볼 역시 디테일이 정확함.",
      "hard_violations": [],
      "physics": "손이 스노우볼의 하단을 단단히 붙잡고 지지하고 있음."
     },
     {
      "label": "B",
      "direction": "시선이 정면을 멍하게 향하고 있어 특정 대상을 쏘아보지 않음.",
      "built_space": "이전 샷의 소파, 데스크, 책장, 창문이 동일한 구도로 배치됨.",
      "entities": "모자는 전택수의 것이나 복장은 이전 샷 의사의 가운과 넥타이를 그대로 입고 있음. 스노우볼은 소파 위에 방치됨.",
      "hard_violations": [
       "지시된 클로즈업 프레이밍 위반 (미디엄 샷)",
       "이전 샷 인물의 복장 차용 금지 조항 위반",
       "스노우볼을 손에 쥐고 있지 않음"
      ],
      "physics": "스노우볼이 손이 아닌 소파 방석 위에 지지되어 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S81sh6"
  }
 },
 "S81sh14::cine": {
  "applied": true,
  "fingerprint": "16ba68ec921d4696e63e85a1495f5488c458e4885119e9564827dc5a45c7c368",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S81sh14_sel.png",
  "source_sha256": "570e0cc23591df36ee6ebac4fd43ee77937e3dee7f4c644a452a0c92990863a8",
  "file": "S81sh14_cine.png",
  "latency_ms": 13688
 },
 "S82sh2::signage": {
  "fp": "04341b8bdf7a6769",
  "inscriptions": [
   {
    "surface_native": "버스정류장 표지판",
    "text_native": "유치원 앞",
    "reason_ko": "유치원 근처 버스 정류장이라는 공간적 배경을 자연스럽게 나타내기 위해 정류장 표지판에 이름을 표기합니다."
   }
  ]
 },
 "era_assess::c586ca8bfc20a562": {
  "subjects": [
   {
    "subject_native": "2000년대-2010년대 한국의 버스 정류장 쉘터",
    "search_terms_native": [
     "2010년 버스정류장",
     "2000년대 버스 쉘터",
     "한국 버스정류소 승강장"
    ],
    "language_lock_native": "모든 검색어는 한국어로만 작성되어야 하며, 영어 단어나 타 언어 번역을 포함해서는 안 됩니다.",
    "reason_ko": "2001~2017년 기간의 한국 버스 정류장은 특정 형태의 유리 및 금속 프레임, 노선 표지판, 벤치 디자인을 지니고 있어 일반적인 이미지 생성기로는 왜곡되기 쉽습니다."
   }
  ]
 },
 "era_ref::234b507d11b86947": {
  "subject": "2000년대-2010년대 한국의 버스 정류장 쉘터",
  "terms": [
   "2010년 버스정류장",
   "2000년대 버스 쉘터",
   "한국 버스정류소 승강장"
  ],
  "queries": [
   [
    "2010년 한국 버스정류장",
    "2000년대 한국 버스 쉘터 버스정류소 승강장"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDAzMTJfMTQg/MDAxNzEwMjM2MjQ0NDUx.NcqHf14R2BZdf2Fum0oP0uV3bg1D4tAJ28PuOxK0A4sg.GYtgFSjAqy9Q0qjdJ6OVmTBcehMEwLXXuk_8Wn-1Y9Ag.JPEG/20240312_124131.jpg?type=w800",
  "picked_reason_ko": "1번은 해당 시대 한국의 평범한 버스 정류장 쉘터를 가장 크게 또렷하게 보여 주며, 골조·벽체·지붕·벤치와 각종 부착물까지 읽기 가장 쉽다.",
  "sha256": "ecdd146658c64ad708618d21b1e60c7ec58ca08f4d428e985cb5ced027276e59",
  "file": "eraref_234b507d11b86947.png"
 },
 "S82sh2::bgfirst_bg": {
  "input_fingerprint": "34c38e37d117d91d",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 정류장 지붕 아래 서서 난감한 표정으로 떨어지는 빗줄기를 올려다보는 이미경의 상체\n\nLOCATION (lock): Under the roof of a roadside bus shelter near the kindergarten, open to the falling rain.\n\nTIME OF DAY (lock): morning, overcast with rain showers.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Under the bus-stop roof, the dolly-in settles at chest height beside and slightly ahead of 이미경, viewing her in profile across a medium upper-body frame. She holds her weight back beneath the shelter and tilts her face toward the falling rain outside, with the roof edge separating her protected position from the shower beyond.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bus-stop roof (In use as shelter from the shower) — Its outer edge crosses the upper frame, with the street visible beyond it; used as Creates an overhead boundary that frames 이미경 beneath shelter; falling rain (A sudden morning shower is falling); used as Occupies the open streetward portion of the frame and motivates 이미경's upward gaze; street near the kindergarten (Visible through the falling rain); used as Provides the exterior depth beyond the sheltered foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diffuse overcast morning daylight creates soft, restrained contrast beneath the bus-stop roof.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국의 버스 정류장 쉘터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 정류장 지붕 아래 서서 난감한 표정으로 떨어지는 빗줄기를 올려다보는 이미경의 상체\n\nLOCATION (lock): Under the roof of a roadside bus shelter near the kindergarten, open to the falling rain.\n\nTIME OF DAY (lock): morning, overcast with rain showers.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Under the bus-stop roof, the dolly-in settles at chest height beside and slightly ahead of 이미경, viewing her in profile across a medium upper-body frame. She holds her weight back beneath the shelter and tilts her face toward the falling rain outside, with the roof edge separating her protected position from the shower beyond.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bus-stop roof (In use as shelter from the shower) — Its outer edge crosses the upper frame, with the street visible beyond it; used as Creates an overhead boundary that frames 이미경 beneath shelter; falling rain (A sudden morning shower is falling); used as Occupies the open streetward portion of the frame and motivates 이미경's upward gaze; street near the kindergarten (Visible through the falling rain); used as Provides the exterior depth beyond the sheltered foreground.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diffuse overcast morning daylight creates soft, restrained contrast beneath the bus-stop roof.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대-2010년대 한국의 버스 정류장 쉘터: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S82sh2__bgfirst_bg.png",
  "asset_id": "24e13a05-a3c1-4d78-8208-b02f2b8b0696",
  "input_asset_ids": [
   "b710884d-75c5-478b-bf3e-f2a2e6e31bde",
   "37c616bc-e139-4551-8bfb-ab29811453bc"
  ],
  "era_research": {
   "subject": "2000년대-2010년대 한국의 버스 정류장 쉘터",
   "queries": [
    [
     "2010년 한국 버스정류장",
     "2000년대 한국 버스 쉘터 버스정류소 승강장"
    ]
   ],
   "picked_url": "https://mblogthumb-phinf.pstatic.net/MjAyNDAzMTJfMTQg/MDAxNzEwMjM2MjQ0NDUx.NcqHf14R2BZdf2Fum0oP0uV3bg1D4tAJ28PuOxK0A4sg.GYtgFSjAqy9Q0qjdJ6OVmTBcehMEwLXXuk_8Wn-1Y9Ag.JPEG/20240312_124131.jpg?type=w800",
   "sha256": "ecdd146658c64ad708618d21b1e60c7ec58ca08f4d428e985cb5ced027276e59",
   "file": "eraref_234b507d11b86947.png"
  }
 },
 "S82sh2": {
  "input_fingerprint": "8624d128adff80b1",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, overcast with rain showers.\n\nSHOT TEXT (authoritative, Korean): 정류장 지붕 아래 서서 난감한 표정으로 떨어지는 빗줄기를 올려다보는 이미경의 상체\n\nLOCATION (lock): Under the roof of a roadside bus shelter near the kindergarten, open to the falling rain. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Under the bus-stop roof, the dolly-in settles at chest height beside and slightly ahead of 이미경, viewing her in profile across a medium upper-body frame. She holds her weight back beneath the shelter and tilts her face toward the falling rain outside, with the roof edge separating her protected position from the shower beyond.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bus-stop roof (In use as shelter from the shower) — Its outer edge crosses the upper frame, with the street visible beyond it; used as Creates an overhead boundary that frames 이미경 beneath shelter; falling rain (A sudden morning shower is falling); used as Occupies the open streetward portion of the frame and motivates 이미경's upward gaze; street near the kindergarten (Visible through the falling rain); used as Provides the exterior depth beyond the sheltered foreground.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diffuse overcast morning daylight creates soft, restrained contrast beneath the bus-stop roof.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong has no umbrella and remains under the bus-stop roof while Taeksu approaches carrying one.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 버스정류장 표지판: \"유치원 앞\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, overcast with rain showers.\n\nSHOT TEXT (authoritative, Korean): 정류장 지붕 아래 서서 난감한 표정으로 떨어지는 빗줄기를 올려다보는 이미경의 상체\n\nLOCATION (lock): Under the roof of a roadside bus shelter near the kindergarten, open to the falling rain. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Under the bus-stop roof, the dolly-in settles at chest height beside and slightly ahead of 이미경, viewing her in profile across a medium upper-body frame. She holds her weight back beneath the shelter and tilts her face toward the falling rain outside, with the roof edge separating her protected position from the shower beyond.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bus-stop roof (In use as shelter from the shower) — Its outer edge crosses the upper frame, with the street visible beyond it; used as Creates an overhead boundary that frames 이미경 beneath shelter; falling rain (A sudden morning shower is falling); used as Occupies the open streetward portion of the frame and motivates 이미경's upward gaze; street near the kindergarten (Visible through the falling rain); used as Provides the exterior depth beyond the sheltered foreground.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diffuse overcast morning daylight creates soft, restrained contrast beneath the bus-stop roof.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong has no umbrella and remains under the bus-stop roof while Taeksu approaches carrying one.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 버스정류장 표지판: \"유치원 앞\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, overcast with rain showers.\n\nSHOT TEXT (authoritative, Korean): 정류장 지붕 아래 서서 난감한 표정으로 떨어지는 빗줄기를 올려다보는 이미경의 상체\n\nLOCATION (lock): Under the roof of a roadside bus shelter near the kindergarten, open to the falling rain. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Under the bus-stop roof, the dolly-in settles at chest height beside and slightly ahead of 이미경, viewing her in profile across a medium upper-body frame. She holds her weight back beneath the shelter and tilts her face toward the falling rain outside, with the roof edge separating her protected position from the shower beyond.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: bus-stop roof (In use as shelter from the shower) — Its outer edge crosses the upper frame, with the street visible beyond it; used as Creates an overhead boundary that frames 이미경 beneath shelter; falling rain (A sudden morning shower is falling); used as Occupies the open streetward portion of the frame and motivates 이미경's upward gaze; street near the kindergarten (Visible through the falling rain); used as Provides the exterior depth beyond the sheltered foreground.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Diffuse overcast morning daylight creates soft, restrained contrast beneath the bus-stop roof.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Mi-gyeong has no umbrella and remains under the bus-stop roof while Taeksu approaches carrying one.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 버스정류장 표지판: \"유치원 앞\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S82sh2__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S82sh2.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ],
   "B": [
    {
     "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_roadside_bus_stop_sel.png"
    },
    {
     "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:832451>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 샷 텍스트에 따라 인물을 단독으로 배치하였고, 참조된 위치의 구조와 표지판 디테일을 프레이밍에 맞게 정확히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "샷 텍스트에 명시되지 않은 인물이 배경에 등장하여 지침을 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "여자의 시선은 정류장 밖 떨어지는 빗줄기를 향해 위로 향해 있다.",
      "built_space": "버스 정류장 지붕과 노선도 유리가 보이나, 참조 사진에 없는 사각형 간판이 지붕에 매달려 있다.",
      "entities": "이미경의 복장(핑크 카디건, 리본)은 일치하나, 배경에 우산을 쓴 남자가 불필요하게 등장한다.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(우산 쓴 남자) 추가"
      ],
      "physics": "여자와 배경의 남자 모두 지면에 안정적으로 서 있거나 걷고 있다."
     },
     {
      "label": "B",
      "direction": "여자의 시선은 난감한 표정과 함께 정류장 밖의 빗줄기를 올려다보고 있다.",
      "built_space": "버스 정류장의 지붕이 화면 상단을 덮고 있으며, 참조 사진의 파란색 표지판 기둥과 배경의 유치원 건물이 올바르게 배치되었다.",
      "entities": "이미경 단독으로 등장하며 참조된 인상착의와 완전히 일치한다. 지정된 텍스트('유치원 앞')가 표지판에 정확히 적혀 있다.",
      "hard_violations": [],
      "physics": "여자는 지면에 안정적으로 서서 체중을 지탱하고 있다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "지정된 샷 텍스트에 따라 인물을 단독으로 배치하였고, 참조된 위치의 구조와 표지판 디테일을 프레이밍에 맞게 정확히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "샷 텍스트에 명시되지 않은 인물이 배경에 등장하여 지침을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "여자의 시선은 정류장 밖 떨어지는 빗줄기를 향해 위로 향해 있다.",
      "built_space": "버스 정류장 지붕과 노선도 유리가 보이나, 참조 사진에 없는 사각형 간판이 지붕에 매달려 있다.",
      "entities": "이미경의 복장(핑크 카디건, 리본)은 일치하나, 배경에 우산을 쓴 남자가 불필요하게 등장한다.",
      "hard_violations": [
       "샷 텍스트에 없는 인물(우산 쓴 남자) 추가"
      ],
      "physics": "여자와 배경의 남자 모두 지면에 안정적으로 서 있거나 걷고 있다."
     },
     {
      "label": "B",
      "direction": "여자의 시선은 난감한 표정과 함께 정류장 밖의 빗줄기를 올려다보고 있다.",
      "built_space": "버스 정류장의 지붕이 화면 상단을 덮고 있으며, 참조 사진의 파란색 표지판 기둥과 배경의 유치원 건물이 올바르게 배치되었다.",
      "entities": "이미경 단독으로 등장하며 참조된 인상착의와 완전히 일치한다. 지정된 텍스트('유치원 앞')가 표지판에 정확히 적혀 있다.",
      "hard_violations": [],
      "physics": "여자는 지면에 안정적으로 서서 체중을 지탱하고 있다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "상체 위주의 프레이밍과 단일 인물 조건을 잘 따랐으나, 배경 정류장 구조물과 표지판 형태가 참조 이미지와 다소 다름."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "정류장의 공간 구현은 우수하나 샷 텍스트에 없는 인물을 배경에 추가하여 지시사항을 크게 위반함."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경의 시선은 화면 위쪽 허공에 떨어지는 비를 향함.",
      "built_space": "화면 상단에 정류장 지붕 끝이 보이며, 파란색 원형 표지판이 기둥에 매달려 있음.",
      "entities": "이미경(분홍 가디건, 흰 블라우스), '유치원 앞' 텍스트가 적힌 표지판. 샷 텍스트에 명시된 인물만 존재함.",
      "hard_violations": [],
      "physics": "인물은 바닥에 발을 딛고 안정적으로 서 있으며, 빗방울은 공중에서 아래로 떨어짐."
     },
     {
      "label": "B",
      "direction": "이미경의 시선은 비를 향해 위로 향하며, 배경의 남자는 화면 앞쪽으로 걸어옴.",
      "built_space": "참조 이미지와 일치하는 정류장 내부(벤치, 노선도)가 보이며, 지붕 아래에 표지판이 매달려 있음.",
      "entities": "이미경(가디건, 민트색 치마), 우산을 쓴 남자. 샷 텍스트에 없는 추가 인물이 존재함.",
      "hard_violations": [
       "발명된 인물: 샷 텍스트에 명시되지 않은 우산을 쓴 남자가 배경에 추가됨"
      ],
      "physics": "두 인물 모두 바닥을 단단히 디디고 있으며, 남자는 손으로 우산을 쥐고 체중을 지탱함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "상체 위주의 프레이밍과 단일 인물 조건을 잘 따랐으나, 배경 정류장 구조물과 표지판 형태가 참조 이미지와 다소 다름."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "정류장의 공간 구현은 우수하나 샷 텍스트에 없는 인물을 배경에 추가하여 지시사항을 크게 위반함."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "이미경의 시선은 화면 위쪽 허공에 떨어지는 비를 향함.",
      "built_space": "화면 상단에 정류장 지붕 끝이 보이며, 파란색 원형 표지판이 기둥에 매달려 있음.",
      "entities": "이미경(분홍 가디건, 흰 블라우스), '유치원 앞' 텍스트가 적힌 표지판. 샷 텍스트에 명시된 인물만 존재함.",
      "hard_violations": [],
      "physics": "인물은 바닥에 발을 딛고 안정적으로 서 있으며, 빗방울은 공중에서 아래로 떨어짐."
     },
     {
      "label": "A",
      "direction": "이미경의 시선은 비를 향해 위로 향하며, 배경의 남자는 화면 앞쪽으로 걸어옴.",
      "built_space": "참조 이미지와 일치하는 정류장 내부(벤치, 노선도)가 보이며, 지붕 아래에 표지판이 매달려 있음.",
      "entities": "이미경(가디건, 민트색 치마), 우산을 쓴 남자. 샷 텍스트에 없는 추가 인물이 존재함.",
      "hard_violations": [
       "발명된 인물: 샷 텍스트에 명시되지 않은 우산을 쓴 남자가 배경에 추가됨"
      ],
      "physics": "두 인물 모두 바닥을 단단히 디디고 있으며, 남자는 손으로 우산을 쥐고 체중을 지탱함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 13
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "A",
    "direction": "여자의 시선은 정류장 밖 떨어지는 빗줄기를 향해 위로 향해 있다.",
    "built_space": "버스 정류장 지붕과 노선도 유리가 보이나, 참조 사진에 없는 사각형 간판이 지붕에 매달려 있다.",
    "entities": "이미경의 복장(핑크 카디건, 리본)은 일치하나, 배경에 우산을 쓴 남자가 불필요하게 등장한다.",
    "hard_violations": [
     "샷 텍스트에 없는 인물(우산 쓴 남자) 추가"
    ],
    "physics": "여자와 배경의 남자 모두 지면에 안정적으로 서 있거나 걷고 있다."
   },
   {
    "label": "B",
    "direction": "여자의 시선은 난감한 표정과 함께 정류장 밖의 빗줄기를 올려다보고 있다.",
    "built_space": "버스 정류장의 지붕이 화면 상단을 덮고 있으며, 참조 사진의 파란색 표지판 기둥과 배경의 유치원 건물이 올바르게 배치되었다.",
    "entities": "이미경 단독으로 등장하며 참조된 인상착의와 완전히 일치한다. 지정된 텍스트('유치원 앞')가 표지판에 정확히 적혀 있다.",
    "hard_violations": [],
    "physics": "여자는 지면에 안정적으로 서서 체중을 지탱하고 있다."
   }
  ],
  "totals": {
   "A": 6,
   "B": 13
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "지정된 샷 텍스트에 따라 인물을 단독으로 배치하였고, 참조된 위치의 구조와 표지판 디테일을 프레이밍에 맞게 정확히 구현했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "샷 텍스트에 명시되지 않은 인물이 배경에 등장하여 지침을 크게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_roadside_bus_stop_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 왼쪽의 버스정류장 표지판 구조물이 원형으로 렌더링되어, 세로로 긴 형태인 로케이션 레퍼런스의 고정된 구조와 일치하지 않습니다.",
     "fix_en": "Change the circular bus stop sign on the left to match the tall, pill-shaped blue sign structure from the location reference.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "이미경 왼쪽 가슴에 캐릭터 참조 의상의 명찰이 없다",
     "fix_en": "Add the small rectangular name tag to the left chest of the pink cardigan, matching the character reference.",
     "severity": "minor",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 왼쪽의 버스정류장 표지판 구조물이 원형으로 렌더링되어, 세로로 긴 형태인 로케이션 레퍼런스의 고정된 구조와 일치하지 않습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "버스정류장 표지판이 위치 참조의 고정 구조(버스 아이콘이 있는 세로형 타원 폴)가 아니라 원형 표지판으로 바뀌어 있다",
     "severity": "major"
    },
    {
     "issue_ko": "이미경 왼쪽 가슴에 캐릭터 참조 의상의 명찰이 없다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 왼쪽의 버스정류장 표지판 구조물이 원형으로 렌더링되어, 세로로 긴 형태인 로케이션 레퍼런스의 고정된 구조와 일치하지 않습니다.",
    "fix_en": "Change the circular bus stop sign on the left to match the tall, pill-shaped blue sign structure from the location reference.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "이미경 왼쪽 가슴에 캐릭터 참조 의상의 명찰이 없다",
    "fix_en": "Add the small rectangular name tag to the left chest of the pink cardigan, matching the character reference.",
    "severity": "minor",
    "observation_index": 2
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S82sh2__bgfirst_bg.png",
   "bg_asset_id": "24e13a05-a3c1-4d78-8208-b02f2b8b0696",
   "bg_record_key": "S82sh2::bgfirst_bg",
   "chain_winner": false,
   "authority": "seed_bg"
  },
  "ref_mode": "seed-bg+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_ready"
 },
 "S82sh2::cine": {
  "applied": true,
  "fingerprint": "3d584c9c88410b44fbac4fd809a0d9db2e4825f9b45672a0ceaba8b5382ce495",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S82sh2_sel.png",
  "source_sha256": "d6c89498d2b35d09ae014dba390b2725f833cbb41e3ba151d7bff6c78bfa0875",
  "file": "S82sh2_cine.png",
  "latency_ms": 19689
 },
 "S82sh4::signage": {
  "fp": "372017bbd89ac725",
  "inscriptions": [
   {
    "surface_native": "어린이 보호구역 표지판",
    "text_native": "어린이 보호구역",
    "reason_ko": "유치원과 버스 정류장 사이의 한국 거리라는 공간적 배경을 사실적으로 나타내기 위해 도로변의 어린이 보호구역 표지판 문구가 필요합니다."
   }
  ]
 },
 "S82sh4": {
  "input_fingerprint": "82d696d6c3a1d634",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, overcast with rain showers.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 거리에서 하나의 우산을 나란히 쓴 채 한쪽 발을 내딛고 있는 걷기 중간 자세의 전택수와 이미경의 뒷모습\n\nLOCATION (lock): Along the rain-soaked street between the bus stop and kindergarten, beneath a shared umbrella. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera tracks at waist height several meters behind 전택수 and 이미경, offset slightly from their shared path while keeping both full figures beneath one umbrella. They occupy the middle of the frame and move toward the upper center, caught in the same beat of a stride but at different phases—전택수's front foot taking weight as 이미경's rear heel lifts.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, moves toward street ahead; 이미경 in the middle-right of the frame, midground, moves toward street ahead.\n- KEY BACKGROUND ELEMENTS: shared umbrella (Held above 전택수 and 이미경 as they walk) — The canopy is seen from behind and slightly below, centered over both walkers; used as Visually joins the two figures while preserving their awkward physical separation below it; rainy street (Rain is falling along the walking route); used as Supplies the receding path and procedural exterior context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Overcast morning daylight remains diffuse and subdued across the rain-filled street.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu and Mi-gyeong continue sharing Taeksu's single umbrella as they walk through the rain.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수와 이미경 right now — 전택수와 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수와 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 어린이 보호구역 표지판: \"어린이 보호구역\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, overcast with rain showers.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 거리에서 하나의 우산을 나란히 쓴 채 한쪽 발을 내딛고 있는 걷기 중간 자세의 전택수와 이미경의 뒷모습\n\nLOCATION (lock): Along the rain-soaked street between the bus stop and kindergarten, beneath a shared umbrella. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera tracks at waist height several meters behind 전택수 and 이미경, offset slightly from their shared path while keeping both full figures beneath one umbrella. They occupy the middle of the frame and move toward the upper center, caught in the same beat of a stride but at different phases—전택수's front foot taking weight as 이미경's rear heel lifts.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, moves toward street ahead; 이미경 in the middle-right of the frame, midground, moves toward street ahead.\n- KEY BACKGROUND ELEMENTS: shared umbrella (Held above 전택수 and 이미경 as they walk) — The canopy is seen from behind and slightly below, centered over both walkers; used as Visually joins the two figures while preserving their awkward physical separation below it; rainy street (Rain is falling along the walking route); used as Supplies the receding path and procedural exterior context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Overcast morning daylight remains diffuse and subdued across the rain-filled street.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu and Mi-gyeong continue sharing Taeksu's single umbrella as they walk through the rain.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수와 이미경 right now — 전택수와 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수와 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 어린이 보호구역 표지판: \"어린이 보호구역\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, overcast with rain showers.\n\nSHOT TEXT (authoritative, Korean): 비 내리는 거리에서 하나의 우산을 나란히 쓴 채 한쪽 발을 내딛고 있는 걷기 중간 자세의 전택수와 이미경의 뒷모습\n\nLOCATION (lock): Along the rain-soaked street between the bus stop and kindergarten, beneath a shared umbrella. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The camera tracks at waist height several meters behind 전택수 and 이미경, offset slightly from their shared path while keeping both full figures beneath one umbrella. They occupy the middle of the frame and move toward the upper center, caught in the same beat of a stride but at different phases—전택수's front foot taking weight as 이미경's rear heel lifts.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, moves toward street ahead; 이미경 in the middle-right of the frame, midground, moves toward street ahead.\n- KEY BACKGROUND ELEMENTS: shared umbrella (Held above 전택수 and 이미경 as they walk) — The canopy is seen from behind and slightly below, centered over both walkers; used as Visually joins the two figures while preserving their awkward physical separation below it; rainy street (Rain is falling along the walking route); used as Supplies the receding path and procedural exterior context.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Overcast morning daylight remains diffuse and subdued across the rain-filled street.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu and Mi-gyeong continue sharing Taeksu's single umbrella as they walk through the rain.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수와 이미경 right now — 전택수와 이미경's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수와 이미경 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 어린이 보호구역 표지판: \"어린이 보호구역\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.667
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.667
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 1667
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "이전 샷 레퍼런스의 배경(버스 정류장, 유리문 건물, 타일 벽)을 매우 정확하게 유지하면서 요청된 인물들의 뒷모습 걷기 자세와 표지판을 충실히 구현했습니다."
   },
   {
    "label": "B",
    "score": 1667,
    "verdict_ko": "인물과 복장은 잘 묘사되었으나, 이전 샷 레퍼런스에서 고정된 장소의 배경(버스 정류장 디자인, 건물 외관 등)이 일치하지 않습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S82sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   },
   {
    "label": "PROP REFERENCE — 검은 우산: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1008354>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우산이 전택수만 가리고 있으며 이미경은 우산 밖에서 비를 맞고 있어, 두 사람이 하나의 우산을 나란히 쓰고 우산이 두 사람 중앙에 위치해야 한다는 프롬프트 지시를 위반함.",
     "fix_en": "Redraw the black umbrella canopy to span centered over both the man and the woman, replacing the previous canopy's far-left side with the background building and trees, while strictly preserving the characters, their clothing, the street, the yellow sign, and the rain.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "전택수의 오른팔 옆에 위치한 우산의 J자형 손잡이가 비정상적인 각도로 옆으로 꺾여 물리적으로 불가능한 형태로 묘사됨.",
     "fix_en": "Redraw the man's right hand and the umbrella handle to form a single straight vertical shaft with a downward J-curve, replacing the broken horizontal section with his dark police jacket, preserving the characters, clothing, umbrella canopy, and background exactly.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "화면 좌측의 버스 정류장에 이전 장면 레퍼런스 이미지에서 확인되는 파란색 원형 표지판('유치원 앞')과 파란색 기둥이 누락됨.",
     "fix_en": "Add the missing blue circular sign and blue pole to the bus stop.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "카메라가 허리 높이가 아니어서 우산 캐노피가 뒤·아래가 아니라 윗면으로 프레임 상단에 보임",
     "fix_en": "Lower the camera viewpoint to waist height.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경 건물과 버스정류장 형태가 이전 컷에 고정된 장소와 다름",
     "fix_en": "Rebuild the background architecture to match the reference.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이미경의 뒷발 뒤꿈치가 들리지 않고 두 발이 모두 땅에 닿아 지정된 보행 위상이 아님",
     "fix_en": "Redraw the woman's right foot to lift the heel.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "이미경의 상의가 이전 컷의 분홍 가디건이 아니라 허리 골이 있는 스웨터로 보임",
     "fix_en": "Redraw the woman's top as a pink cardigan with a white blouse collar visible.",
     "severity": "major",
     "observation_index": 6
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우산이 전택수만 가리고 있으며 이미경은 우산 밖에서 비를 맞고 있어, 두 사람이 하나의 우산을 나란히 쓰고 우산이 두 사람 중앙에 위치해야 한다는 프롬프트 지시를 위반함.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수의 오른팔 옆에 위치한 우산의 J자형 손잡이가 비정상적인 각도로 옆으로 꺾여 물리적으로 불가능한 형태로 묘사됨.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 좌측의 버스 정류장에 이전 장면 레퍼런스 이미지에서 확인되는 파란색 원형 표지판('유치원 앞')과 파란색 기둥이 누락됨.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 허리 높이가 아니어서 우산 캐노피가 뒤·아래가 아니라 윗면으로 프레임 상단에 보임",
     "severity": "major"
    },
    {
     "issue_ko": "배경 건물과 버스정류장 형태가 이전 컷에 고정된 장소와 다름",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 뒷발 뒤꿈치가 들리지 않고 두 발이 모두 땅에 닿아 지정된 보행 위상이 아님",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 상의가 이전 컷의 분홍 가디건이 아니라 허리 골이 있는 스웨터로 보임",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 좌측의 버스 정류장에 이전 장면 레퍼런스 이미지에서 확인되는 파란색 원형 표지판('유치원 앞')과 파란색 기둥이 누락됨.",
    "fix_en": "Add the missing blue circular sign and blue pole to the bus stop.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "카메라가 허리 높이가 아니어서 우산 캐노피가 뒤·아래가 아니라 윗면으로 프레임 상단에 보임",
    "fix_en": "Lower the camera viewpoint to waist height.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경 건물과 버스정류장 형태가 이전 컷에 고정된 장소와 다름",
    "fix_en": "Rebuild the background architecture to match the reference.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "이미경의 뒷발 뒤꿈치가 들리지 않고 두 발이 모두 땅에 닿아 지정된 보행 위상이 아님",
    "fix_en": "Redraw the woman's right foot to lift the heel.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "이미경의 상의가 이전 컷의 분홍 가디건이 아니라 허리 골이 있는 스웨터로 보임",
    "fix_en": "Redraw the woman's top as a pink cardigan with a white blouse collar visible.",
    "severity": "major",
    "observation_index": 6
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the black umbrella canopy to span centered over both the man and the woman, replacing the previous canopy's far-left side with the background building and trees, while strictly preserving the characters, their clothing, the street, the yellow sign, and the rain.\n- Redraw the man's right hand and the umbrella handle to form a single straight vertical shaft with a downward J-curve, replacing the broken horizontal section with his dark police jacket, preserving the characters, clothing, umbrella canopy, and background exactly.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트에서 요구한 뒷모습 카메라 앵글, 이전 숏과 일치하는 배경, 지정된 간판 텍스트, 그리고 걷고 있는 중간 자세를 완벽하게 구현한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "프롬프트가 명시한 '뒷모습'을 무시하고 정면을 묘사했으며, 지정된 장소(버스 정류장 앞)를 벗어나 완전히 다른 거리를 배경으로 삼은 데다, 걷는 자세가 아닌 서 있는 자세를 렌더링하여 지침을 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 인물은 카메라를 등지고 앞을 향해 걸어가고 있으며, 시선과 진행 방향이 프레임 안쪽을 향함",
      "built_space": "이전 숏 레퍼런스와 일치하는 버스 정류장 구조물과 유치원 건물이 배경에 정확히 배치되었으며, 우측에 지정된 문구의 표지판이 있음",
      "entities": "전택수와 이미경의 뒷모습이 레퍼런스의 복장 및 체형과 일치하며, 지시된 검은색 우산과 '어린이 보호구역' 표지판이 정확히 렌더링됨",
      "hard_violations": [],
      "physics": "두 인물 모두 한쪽 발이 땅에 닿고 다른 발의 뒤꿈치가 들려있는 걷기 중간 자세를 취하고 있으며, 전택수의 오른손이 우산대 손잡이를 자연스럽게 쥐고 있음"
     },
     {
      "label": "B",
      "direction": "두 인물이 카메라를 정면으로 응시하고 있음",
      "built_space": "이전 숏 레퍼런스와 전혀 일치하지 않는 일반적인 시가지 배경이 렌더링됨",
      "entities": "전택수와 이미경이 정면으로 묘사되었으나, 요구된 표지판 내용이 없으며 무관한 간판이 배치됨",
      "hard_violations": [
       "명시된 뒷모습(카메라 앵글 및 방향) 위반 - 정면 묘사",
       "명시된 장소(이전 숏의 배경) 위반 - 완전히 다른 거리 묘사",
       "걷기 중간 자세 위반 - 제자리에 가만히 서 있음"
      ],
      "physics": "두 인물 모두 양발을 땅에 디디고 정지해 있으며, 보행의 동력이 전혀 보이지 않음"
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "프롬프트에서 요구한 뒷모습 카메라 앵글, 이전 숏과 일치하는 배경, 지정된 간판 텍스트, 그리고 걷고 있는 중간 자세를 완벽하게 구현한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "프롬프트가 명시한 '뒷모습'을 무시하고 정면을 묘사했으며, 지정된 장소(버스 정류장 앞)를 벗어나 완전히 다른 거리를 배경으로 삼은 데다, 걷는 자세가 아닌 서 있는 자세를 렌더링하여 지침을 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "두 인물은 카메라를 등지고 앞을 향해 걸어가고 있으며, 시선과 진행 방향이 프레임 안쪽을 향함",
      "built_space": "이전 숏 레퍼런스와 일치하는 버스 정류장 구조물과 유치원 건물이 배경에 정확히 배치되었으며, 우측에 지정된 문구의 표지판이 있음",
      "entities": "전택수와 이미경의 뒷모습이 레퍼런스의 복장 및 체형과 일치하며, 지시된 검은색 우산과 '어린이 보호구역' 표지판이 정확히 렌더링됨",
      "hard_violations": [],
      "physics": "두 인물 모두 한쪽 발이 땅에 닿고 다른 발의 뒤꿈치가 들려있는 걷기 중간 자세를 취하고 있으며, 전택수의 오른손이 우산대 손잡이를 자연스럽게 쥐고 있음"
     },
     {
      "label": "B",
      "direction": "두 인물이 카메라를 정면으로 응시하고 있음",
      "built_space": "이전 숏 레퍼런스와 전혀 일치하지 않는 일반적인 시가지 배경이 렌더링됨",
      "entities": "전택수와 이미경이 정면으로 묘사되었으나, 요구된 표지판 내용이 없으며 무관한 간판이 배치됨",
      "hard_violations": [
       "명시된 뒷모습(카메라 앵글 및 방향) 위반 - 정면 묘사",
       "명시된 장소(이전 숏의 배경) 위반 - 완전히 다른 거리 묘사",
       "걷기 중간 자세 위반 - 제자리에 가만히 서 있음"
      ],
      "physics": "두 인물 모두 양발을 땅에 디디고 정지해 있으며, 보행의 동력이 전혀 보이지 않음"
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "프롬프트가 명확하게 '뒷모습'을 요구했음에도 불구하고 인물들이 정면을 응시하고 있어 샷의 기본 구성을 심각하게 위반했습니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "요구된 카메라 앵글(뒷모습), 걷는 동작, 지정된 텍스트('어린이 보호구역'), 그리고 이전 샷의 장소적 맥락을 충실히 구현한 우수한 결과물입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "두 인물이 카메라를 정면으로 바라보고 서 있음.",
      "built_space": "비 내리는 거리지만, 이전 샷에서 요구된 버스 정류장이나 유치원의 특정 구조물 배치가 보이지 않음.",
      "entities": "전택수와 이미경의 정면 얼굴과 복장이 레퍼런스와 일치함. 검은 우산이 쓰이고 있음. 간판 텍스트가 요구사항과 다름.",
      "hard_violations": [
       "프롬프트에서 지시한 '뒷모습(카메라가 뒤에서 따라가는 구도)'과 정반대로 정면을 보여줌 (심각한 연출 위반)",
       "걷기 중간 자세가 아닌 정지된 자세"
      ],
      "physics": "두 인물 모두 두 발을 땅에 딛고 정지해 있음. 남성이 우산을 들고 있음."
     },
     {
      "label": "B",
      "direction": "두 인물이 카메라를 등지고 앞을 향해 걷고 있음.",
      "built_space": "비 내리는 거리, 왼쪽에 버스 정류장 구조물이 있고 오른쪽에 '어린이 보호구역' 표지판이 배치됨.",
      "entities": "전택수와 이미경의 뒷모습. 의상이 레퍼런스와 일치함. 하나의 검은 우산을 함께 쓰고 있음. '어린이 보호구역' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 걷는 중(mid-stride). 전택수는 오른발에 체중을 싣고 있고, 이미경은 오른발 뒤꿈치가 들려 있음. 전택수의 손이 우산을 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "프롬프트가 명확하게 '뒷모습'을 요구했음에도 불구하고 인물들이 정면을 응시하고 있어 샷의 기본 구성을 심각하게 위반했습니다."
     },
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "요구된 카메라 앵글(뒷모습), 걷는 동작, 지정된 텍스트('어린이 보호구역'), 그리고 이전 샷의 장소적 맥락을 충실히 구현한 우수한 결과물입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "두 인물이 카메라를 정면으로 바라보고 서 있음.",
      "built_space": "비 내리는 거리지만, 이전 샷에서 요구된 버스 정류장이나 유치원의 특정 구조물 배치가 보이지 않음.",
      "entities": "전택수와 이미경의 정면 얼굴과 복장이 레퍼런스와 일치함. 검은 우산이 쓰이고 있음. 간판 텍스트가 요구사항과 다름.",
      "hard_violations": [
       "프롬프트에서 지시한 '뒷모습(카메라가 뒤에서 따라가는 구도)'과 정반대로 정면을 보여줌 (심각한 연출 위반)",
       "걷기 중간 자세가 아닌 정지된 자세"
      ],
      "physics": "두 인물 모두 두 발을 땅에 딛고 정지해 있음. 남성이 우산을 들고 있음."
     },
     {
      "label": "A",
      "direction": "두 인물이 카메라를 등지고 앞을 향해 걷고 있음.",
      "built_space": "비 내리는 거리, 왼쪽에 버스 정류장 구조물이 있고 오른쪽에 '어린이 보호구역' 표지판이 배치됨.",
      "entities": "전택수와 이미경의 뒷모습. 의상이 레퍼런스와 일치함. 하나의 검은 우산을 함께 쓰고 있음. '어린이 보호구역' 텍스트가 정확히 표기됨.",
      "hard_violations": [],
      "physics": "두 인물 모두 걷는 중(mid-stride). 전택수는 오른발에 체중을 싣고 있고, 이미경은 오른발 뒤꿈치가 들려 있음. 전택수의 손이 우산을 지지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S82sh2"
  }
 },
 "S82sh4::cine": {
  "applied": true,
  "fingerprint": "3979ab9246eab7c95b3ac2b143af93793d994e4f71b1d43e5d10717179290d5a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S82sh4_sel.png",
  "source_sha256": "b6604ff5ceee5a96a3c969896559c9d50ab614611d9d9b4cfa46713da494a8f6",
  "file": "S82sh4_cine.png",
  "latency_ms": 18598
 },
 "S83sh1::signage": {
  "fp": "c4226ea2e17eb457",
  "inscriptions": [
   {
    "surface_native": "유치원 간판",
    "text_native": "새싹유치원",
    "reason_ko": "유치원 현관 앞이라는 공간적 배경을 직관적으로 보여주기 위해 유치원 이름이 적힌 간판이 필요합니다."
   }
  ]
 },
 "S83sh1": {
  "input_fingerprint": "2a647a89322f4e98",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, after rain.\n\nSHOT TEXT (authoritative, Korean): 비가 그친 유치원 현관 앞, 전택수를 향해 가볍게 고개를 숙인 채 어색하게 묵례하는 이미경의 상체\n\nLOCATION (lock): Outside the kindergarten's covered front entrance after the rain has eased. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From beside the kindergarten entrance and slightly above shoulder height, the static camera holds a tightened three-quarter upper-body view of 이미경, with part of 전택수's near shoulder at the opposite edge. 이미경 bends her head in a restrained bow toward him, leaving an awkward interval of empty space between their bodies as they acknowledge one another.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: kindergarten entrance (They have arrived in front of it) — The entrance is seen obliquely behind 이미경 from its outer side; used as Anchors their arrival and frames 이미경's side of the parting; umbrella (Held by 전택수) — Its handle and part of its form angle into the frame beside 전택수; used as Appears partially with 전택수 at the frame edge as evidence of the shared walk.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft morning daylight after the rain has subsided keeps the contrast naturalistic and restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu still holds the umbrella they shared, though the rain has now subsided at the kindergarten entrance.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유치원 간판: \"새싹유치원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, after rain.\n\nSHOT TEXT (authoritative, Korean): 비가 그친 유치원 현관 앞, 전택수를 향해 가볍게 고개를 숙인 채 어색하게 묵례하는 이미경의 상체\n\nLOCATION (lock): Outside the kindergarten's covered front entrance after the rain has eased. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From beside the kindergarten entrance and slightly above shoulder height, the static camera holds a tightened three-quarter upper-body view of 이미경, with part of 전택수's near shoulder at the opposite edge. 이미경 bends her head in a restrained bow toward him, leaving an awkward interval of empty space between their bodies as they acknowledge one another.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: kindergarten entrance (They have arrived in front of it) — The entrance is seen obliquely behind 이미경 from its outer side; used as Anchors their arrival and frames 이미경's side of the parting; umbrella (Held by 전택수) — Its handle and part of its form angle into the frame beside 전택수; used as Appears partially with 전택수 at the frame edge as evidence of the shared walk.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft morning daylight after the rain has subsided keeps the contrast naturalistic and restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu still holds the umbrella they shared, though the rain has now subsided at the kindergarten entrance.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유치원 간판: \"새싹유치원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, after rain.\n\nSHOT TEXT (authoritative, Korean): 비가 그친 유치원 현관 앞, 전택수를 향해 가볍게 고개를 숙인 채 어색하게 묵례하는 이미경의 상체\n\nLOCATION (lock): Outside the kindergarten's covered front entrance after the rain has eased. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From beside the kindergarten entrance and slightly above shoulder height, the static camera holds a tightened three-quarter upper-body view of 이미경, with part of 전택수's near shoulder at the opposite edge. 이미경 bends her head in a restrained bow toward him, leaving an awkward interval of empty space between their bodies as they acknowledge one another.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: kindergarten entrance (They have arrived in front of it) — The entrance is seen obliquely behind 이미경 from its outer side; used as Anchors their arrival and frames 이미경's side of the parting; umbrella (Held by 전택수) — Its handle and part of its form angle into the frame beside 전택수; used as Appears partially with 전택수 at the frame edge as evidence of the shared walk.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft morning daylight after the rain has subsided keeps the contrast naturalistic and restrained.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu still holds the umbrella they shared, though the rain has now subsided at the kindergarten entrance.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유치원 간판: \"새싹유치원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "이미경의 시선과 고개가 전택수를 향해 아래로 자연스럽게 숙여져 있음.",
    "built_space": "배경에 유치원 입구와 계단이 위치하며 참조 이미지의 건물 구조와 정확히 일치함.",
    "entities": "이미경의 외모와 의상이 참조와 일치함. 우측에 전택수의 어깨와 나무 손잡이 우산이 프레임에 걸쳐 있음. 간판에 '새싹유치원' 텍스트 정확함.",
    "hard_violations": [],
    "physics": "우산이 전택수에 의해 안정적으로 지지된 상태로 보이며, 인물들의 서 있는 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "이미경이 전택수 쪽으로 시선을 내리며 가볍게 고개를 숙이고 있음.",
    "built_space": "배경의 유치원 입구 구조와 간판 위치가 참조와 일치함.",
    "entities": "이미경의 인상착의는 비슷하나 가디건에 참조와 다른 금색 단추가 두드러짐. 우측 가장자리에 전택수의 어깨와 우산이 보임. '새싹유치원' 간판 텍스트 정확함.",
    "hard_violations": [],
    "physics": "우산은 지지되고 있으나 손잡이와 천 부분이 다소 형태가 뭉개져 연결됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 6
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "어색하게 묵례하는 자세와 지정된 프레이밍을 정확히 구현했으며, 참조 이미지의 인물 및 배경 요소와 매우 높은 일치도를 보입니다."
   },
   {
    "label": "B",
    "score": 6,
    "verdict_ko": "전반적인 구도와 배경은 좋으나, 가디건에 참조에 없는 눈에 띄는 단추가 생겼고 우산의 형태가 다소 어색하게 렌더링되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 이미경 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S78sh3_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이미경이 프롬프트의 지시와 달리 고개를 숙여 묵례하지 않고 똑바로 서서 앞을 응시하고 있습니다.",
     "fix_en": "Redraw Lee Mi-kyung's head and neck to tilt downward in a restrained bow, lowering her gaze toward the man. Maintain her identity, hair, pink cardigan, the background, the man on the right, and the umbrella.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "화면 우측의 우산을 쥐고 있는 손이나 지지대가 묘사되지 않아 우산이 인물 앞 허공에 떠 있습니다.",
     "fix_en": "Redraw the area around the wooden umbrella handle to include the man's hand firmly gripping it. Maintain the umbrella's current position, the man's dark shirt, Lee Mi-kyung, and the background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "유치원 현관 캐노피와 간판 구조가 이전 샷의 고정 시설과 다르게 바뀌어 있다",
     "fix_en": "Redraw the background building to match the flat blue and yellow facade from the previous shot, removing the protruding white entrance structure. Maintain the foreground characters and umbrella.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "간판 글자가 '새싹유치원'이 아니라 띄어 쓴 '새 싹유치원'이다",
     "fix_en": "Adjust the text spacing on the background sign to read '새싹유치원' without the wide gap. Maintain all other foreground and background elements.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "프레임 가장자리에 전택수의 어깨만이 아니라 얼굴까지 들어와 있다",
     "fix_en": "Redraw the right edge of the frame to remove the man's face and jaw, leaving only his shoulder visible. Maintain Lee Mi-kyung, the umbrella, and the background.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "이미경이 프롬프트의 지시와 달리 고개를 숙여 묵례하지 않고 똑바로 서서 앞을 응시하고 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 우측의 우산을 쥐고 있는 손이나 지지대가 묘사되지 않아 우산이 인물 앞 허공에 떠 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "이미경이 고개를 숙여 묵례하지 않고 고개를 든 채 옆을 노려보고 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "유치원 현관 캐노피와 간판 구조가 이전 샷의 고정 시설과 다르게 바뀌어 있다",
     "severity": "major"
    },
    {
     "issue_ko": "간판 글자가 '새싹유치원'이 아니라 띄어 쓴 '새 싹유치원'이다",
     "severity": "major"
    },
    {
     "issue_ko": "프레임 가장자리에 전택수의 어깨만이 아니라 얼굴까지 들어와 있다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "유치원 현관 캐노피와 간판 구조가 이전 샷의 고정 시설과 다르게 바뀌어 있다",
    "fix_en": "Redraw the background building to match the flat blue and yellow facade from the previous shot, removing the protruding white entrance structure. Maintain the foreground characters and umbrella.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "간판 글자가 '새싹유치원'이 아니라 띄어 쓴 '새 싹유치원'이다",
    "fix_en": "Adjust the text spacing on the background sign to read '새싹유치원' without the wide gap. Maintain all other foreground and background elements.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "프레임 가장자리에 전택수의 어깨만이 아니라 얼굴까지 들어와 있다",
    "fix_en": "Redraw the right edge of the frame to remove the man's face and jaw, leaving only his shoulder visible. Maintain Lee Mi-kyung, the umbrella, and the background.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw Lee Mi-kyung's head and neck to tilt downward in a restrained bow, lowering her gaze toward the man. Maintain her identity, hair, pink cardigan, the background, the man on the right, and the umbrella.\n- Redraw the area around the wooden umbrella handle to include the man's hand firmly gripping it. Maintain the umbrella's current position, the man's dark shirt, Lee Mi-kyung, and the background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "고개를 숙인 어색한 묵례의 자세를 잘 표현했으며, 우산을 쥔 손이 물리적으로 올바르게 묘사되어 프롬프트에 부합합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "캐릭터와 배경은 잘 구현되었으나, 우산을 잡은 손이 묘사되지 않아 우산이 허공에 떠 있는 치명적인 물리적 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경은 전택수 쪽을 향해 약간 아래로 시선을 두고 있음.",
      "built_space": "배경에 유치원 입구 계단과 간판이 비스듬하게 올바른 비율로 위치함.",
      "entities": "이미경의 얼굴, 헤어, 의상이 레퍼런스와 일치함. 간판 텍스트 '새싹유치원'이 정확함. 우측에 전택수의 어깨와 우산이 보임.",
      "hard_violations": [
       "아무것도 지지하지 않는 허공에 떠 있는 우산 (우산을 잡은 손이 없음)"
      ],
      "physics": "전택수 측의 우산 손잡이 아래로 우산을 지지하는 손이나 물체가 없어 우산이 공중에 떠 있음."
     },
     {
      "label": "B",
      "direction": "이미경이 고개를 살짝 숙인 채 전택수 측 아래를 향해 시선을 둠.",
      "built_space": "유치원 입구와 현관, 간판이 배경에 적절하게 배치됨.",
      "entities": "이미경의 인상착의가 레퍼런스와 완벽히 일치하며 간판 텍스트 '새싹유치원'도 정확함. 우측에 전택수의 어깨와 손, 우산이 있음.",
      "hard_violations": [],
      "physics": "전택수의 손이 우산 기둥을 물리적으로 단단히 쥐고 있어 정상적으로 지지됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "고개를 숙인 어색한 묵례의 자세를 잘 표현했으며, 우산을 쥔 손이 물리적으로 올바르게 묘사되어 프롬프트에 부합합니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "캐릭터와 배경은 잘 구현되었으나, 우산을 잡은 손이 묘사되지 않아 우산이 허공에 떠 있는 치명적인 물리적 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "이미경은 전택수 쪽을 향해 약간 아래로 시선을 두고 있음.",
      "built_space": "배경에 유치원 입구 계단과 간판이 비스듬하게 올바른 비율로 위치함.",
      "entities": "이미경의 얼굴, 헤어, 의상이 레퍼런스와 일치함. 간판 텍스트 '새싹유치원'이 정확함. 우측에 전택수의 어깨와 우산이 보임.",
      "hard_violations": [
       "아무것도 지지하지 않는 허공에 떠 있는 우산 (우산을 잡은 손이 없음)"
      ],
      "physics": "전택수 측의 우산 손잡이 아래로 우산을 지지하는 손이나 물체가 없어 우산이 공중에 떠 있음."
     },
     {
      "label": "B",
      "direction": "이미경이 고개를 살짝 숙인 채 전택수 측 아래를 향해 시선을 둠.",
      "built_space": "유치원 입구와 현관, 간판이 배경에 적절하게 배치됨.",
      "entities": "이미경의 인상착의가 레퍼런스와 완벽히 일치하며 간판 텍스트 '새싹유치원'도 정확함. 우측에 전택수의 어깨와 손, 우산이 있음.",
      "hard_violations": [],
      "physics": "전택수의 손이 우산 기둥을 물리적으로 단단히 쥐고 있어 정상적으로 지지됨."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "이미경의 포즈와 배경은 프롬프트에 부합하나, 우산이 손에 들려 있지 않고 공중에 떠 있는 구조적 오류(지원되지 않는 물체)로 인해 실격됨.  ★위반: [gemini-pro] 공중에 떠 있는 물체 (아무것도 우산을 지탱하거나 쥐고 있지 않음)"
     },
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "우산을 쥐고 있는 손을 추가하여 수정을 시도했으나, 팔이 남자의 몸통 한가운데서 돋아난 형태의 해부학적으로 불가능한 오류를 발생시켜 실격됨.  ★위반: [gemini-pro] 해부학적으로 불가능한 신체 구조 (우산을 쥔 손과 팔이 남자의 가슴/복부에서 돋아나 있으며 소매 색상도 다름)"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.75,
      "B": 1.75
     },
     "adjusted": {
      "A": 1.5,
      "B": 1.5
     },
     "violations": {
      "A": [
       "[gemini-pro] 해부학적으로 불가능한 신체 구조 (우산을 쥔 손과 팔이 남자의 가슴/복부에서 돋아나 있으며 소매 색상도 다름)"
      ],
      "B": [
       "[gemini-pro] 공중에 떠 있는 물체 (아무것도 우산을 지탱하거나 쥐고 있지 않음)"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.25,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1500,
      "verdict_ko": "이미경의 포즈와 배경은 프롬프트에 부합하나, 우산이 손에 들려 있지 않고 공중에 떠 있는 구조적 오류(지원되지 않는 물체)로 인해 실격됨.  ★위반: [gemini-pro] 공중에 떠 있는 물체 (아무것도 우산을 지탱하거나 쥐고 있지 않음)"
     },
     {
      "label": "B",
      "score": 1500,
      "verdict_ko": "우산을 쥐고 있는 손을 추가하여 수정을 시도했으나, 팔이 남자의 몸통 한가운데서 돋아난 형태의 해부학적으로 불가능한 오류를 발생시켜 실격됨.  ★위반: [gemini-pro] 해부학적으로 불가능한 신체 구조 (우산을 쥔 손과 팔이 남자의 가슴/복부에서 돋아나 있으며 소매 색상도 다름)"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1503,
     "B": 1508
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S78sh3"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S83sh1::cine": {
  "applied": true,
  "fingerprint": "eb727d7d0c61a1391d2218c9a1c47ed9a82d11b3a049410e3f48eb1b574de2b3",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S83sh1_sel.png",
  "source_sha256": "da343fddf16a641cbbd1eec0514ea5bc57d696608cbc8936dbd024227f5a4e20",
  "file": "S83sh1_cine.png",
  "latency_ms": 11571
 },
 "S83sh6::signage": {
  "fp": "a381f076afd6e7a2",
  "inscriptions": [
   {
    "surface_native": "유치원 입구 간판",
    "text_native": "샛별유치원",
    "reason_ko": "인물들이 서 있는 장소가 유치원 입구 근처임을 나타내기 위해 배경에 유치원 이름 표지판이 필요합니다."
   }
  ]
 },
 "S83sh6": {
  "input_fingerprint": "3035057f41352965",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, after rain.\n\nSHOT TEXT (authoritative, Korean): 이미경을 향해 자신이 든 우산을 가볍게 치켜든 채 미소 짓는 전택수의 상체\n\nLOCATION (lock): On the pavement just beyond the kindergarten entrance, holding the umbrella after the rain. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The dolly-in settles just behind and beside 이미경's shoulder at chest height, aiming diagonally toward 전택수 in a close medium three-quarter view. 전택수 lifts the umbrella with a bent elbow and gives her an embarrassed smile, while 이미경 remains a soft shoulder profile at the near edge, watching the small gesture.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: umbrella (Lightly raised toward 이미경) — The umbrella is angled upward from 전택수's hand, with its side presented toward the camera; used as Serves as the clearly readable gesture object beside 전택수's upper body; kindergarten entrance (Visible at their parting point) — The entrance is viewed obliquely beyond 이미경's near shoulder; used as Places 이미경 at the threshold while 전택수 prepares to leave.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft morning daylight after the shower has subsided supports a restrained, naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same kindergarten entrance, damp pavement, softened morning light, and the continuing umbrella. Exclude the woman's bowing pose and show the man lifting the umbrella with a small smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu retains the same umbrella and lifts it toward Mi-gyeong before walking away.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유치원 입구 간판: \"샛별유치원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, after rain.\n\nSHOT TEXT (authoritative, Korean): 이미경을 향해 자신이 든 우산을 가볍게 치켜든 채 미소 짓는 전택수의 상체\n\nLOCATION (lock): On the pavement just beyond the kindergarten entrance, holding the umbrella after the rain. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The dolly-in settles just behind and beside 이미경's shoulder at chest height, aiming diagonally toward 전택수 in a close medium three-quarter view. 전택수 lifts the umbrella with a bent elbow and gives her an embarrassed smile, while 이미경 remains a soft shoulder profile at the near edge, watching the small gesture.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: umbrella (Lightly raised toward 이미경) — The umbrella is angled upward from 전택수's hand, with its side presented toward the camera; used as Serves as the clearly readable gesture object beside 전택수's upper body; kindergarten entrance (Visible at their parting point) — The entrance is viewed obliquely beyond 이미경's near shoulder; used as Places 이미경 at the threshold while 전택수 prepares to leave.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft morning daylight after the shower has subsided supports a restrained, naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same kindergarten entrance, damp pavement, softened morning light, and the continuing umbrella. Exclude the woman's bowing pose and show the man lifting the umbrella with a small smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu retains the same umbrella and lifts it toward Mi-gyeong before walking away.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유치원 입구 간판: \"샛별유치원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): morning, after rain.\n\nSHOT TEXT (authoritative, Korean): 이미경을 향해 자신이 든 우산을 가볍게 치켜든 채 미소 짓는 전택수의 상체\n\nLOCATION (lock): On the pavement just beyond the kindergarten entrance, holding the umbrella after the rain. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: The dolly-in settles just behind and beside 이미경's shoulder at chest height, aiming diagonally toward 전택수 in a close medium three-quarter view. 전택수 lifts the umbrella with a bent elbow and gives her an embarrassed smile, while 이미경 remains a soft shoulder profile at the near edge, watching the small gesture.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: umbrella (Lightly raised toward 이미경) — The umbrella is angled upward from 전택수's hand, with its side presented toward the camera; used as Serves as the clearly readable gesture object beside 전택수's upper body; kindergarten entrance (Visible at their parting point) — The entrance is viewed obliquely beyond 이미경's near shoulder; used as Places 이미경 at the threshold while 전택수 prepares to leave.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Soft morning daylight after the shower has subsided supports a restrained, naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same kindergarten entrance, damp pavement, softened morning light, and the continuing umbrella. Exclude the woman's bowing pose and show the man lifting the umbrella with a small smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu retains the same umbrella and lifts it toward Mi-gyeong before walking away.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유치원 입구 간판: \"샛별유치원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.167,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.833,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.833,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1833,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1833,
    "verdict_ko": "캐릭터 모자의 금장 디테일과 검은색 우산 레퍼런스를 충실히 반영했으나, 간판 텍스트('샛별유치원') 생성에 실패함."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "모자 챙 디테일이 누락되었고 지정된 검은 우산 대신 이전 샷의 나무 손잡이를 재현했으며, 간판 텍스트 역시 프롬프트를 따르지 않음."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S83sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "PROP REFERENCE — 검은 우산: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1008354>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "남성이 쥐고 있는 우산의 J자형 손잡이가 손 바로 위쪽 우산대 중간에서 솟아나 있어, 우산의 물리적 구조가 불가능하게 왜곡되었습니다.",
     "fix_en": "Redraw the lower section of the black umbrella shaft to be a continuous, straight rod extending downward from the man's grip, completely removing the anomalous J-shaped hook appended just below his hand. Maintain the man's hand, his pose, his clothing, the woman's profile, the upper umbrella canopy, the background building, and the lighting exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "배경의 유치원 입구 간판 텍스트가 프롬프트에서 지정한 '샛별유치원'이 아닌 '새뿔유치원'으로 잘못 표기되었습니다.",
     "fix_en": "Correct the dark blue Korean text on the white sign above the building entrance in the background to read exactly '샛별유치원'. Maintain the sign's shape, the building structure, both characters, their poses and clothing, the umbrella, and the overall lighting exactly as they are.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "전택수 점퍼 왼쪽 가슴에 레퍼런스와 다른 깨진 영문 'SALIKZ'가 있다",
     "fix_en": "Remove the 'SALIKZ' text from the right chest of the man's jacket.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "유치원 입구가 이미경 어깨 너머가 아니라 전택수 등 뒤 오른편에 보인다",
     "fix_en": "Relocate the kindergarten building to be visible over the woman's near shoulder on the left.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "남성이 쥐고 있는 우산의 J자형 손잡이가 손 바로 위쪽 우산대 중간에서 솟아나 있어, 우산의 물리적 구조가 불가능하게 왜곡되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "배경의 유치원 입구 간판 텍스트가 프롬프트에서 지정한 '샛별유치원'이 아닌 '새뿔유치원'으로 잘못 표기되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "유치원 간판이 지정된 '샛별유치원'이 아니라 '새빌유치원'으로 잘못 적혀 있다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수 점퍼 왼쪽 가슴에 레퍼런스와 다른 깨진 영문 'SALIKZ'가 있다",
     "severity": "major"
    },
    {
     "issue_ko": "유치원 입구가 이미경 어깨 너머가 아니라 전택수 등 뒤 오른편에 보인다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "전택수 점퍼 왼쪽 가슴에 레퍼런스와 다른 깨진 영문 'SALIKZ'가 있다",
    "fix_en": "Remove the 'SALIKZ' text from the right chest of the man's jacket.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "유치원 입구가 이미경 어깨 너머가 아니라 전택수 등 뒤 오른편에 보인다",
    "fix_en": "Relocate the kindergarten building to be visible over the woman's near shoulder on the left.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the lower section of the black umbrella shaft to be a continuous, straight rod extending downward from the man's grip, completely removing the anomalous J-shaped hook appended just below his hand. Maintain the man's hand, his pose, his clothing, the woman's profile, the upper umbrella canopy, the background building, and the lighting exactly as they are.\n- Correct the dark blue Korean text on the white sign above the building entrance in the background to read exactly '샛별유치원'. Maintain the sign's shape, the building structure, both characters, their poses and clothing, the umbrella, and the overall lighting exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지시된 오버더숄더 앵글과 전택수의 상체 프레이밍, 미소 짓는 표정 등을 완벽하게 구현했으나 유치원 간판에 약간의 오타가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "지시된 카메라 앵글을 완전히 무시하고 이전 컷의 구도를 그대로 가져왔으며, 근경에 우산을 든 미상의 인물이 남은 상태에서 배경에 전택수를 추가해 인물 중복 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선과 우산의 방향이 프레임 왼쪽 근경에 걸쳐진 이미경을 향하고 있습니다.",
      "built_space": "배경에 유치원 건물의 입구와 계단이 올바르게 위치해 있으나, 간판 텍스트에 오타('새삘유치원')가 보입니다.",
      "entities": "전택수는 레퍼런스와 일치하는 얼굴과 복장으로 상체만 등장하며, 이미경은 지시대로 왼쪽 가장자리에 어깨와 뒷모습만 보입니다. 검은색 우산의 형태와 재질도 잘 표현되었습니다.",
      "hard_violations": [],
      "physics": "전택수의 손이 우산 손잡이를 자연스럽게 쥐고 있으며, 우산의 각도와 무게 지탱이 물리적으로 타당합니다."
     },
     {
      "label": "B",
      "direction": "근경의 이미경은 아래를 향해 시선을 두고 있으며, 원경의 전택수는 정면 카메라 방향을 응시하고 있습니다.",
      "built_space": "배경에 유치원 건물과 입구가 보이며, 간판 글씨('샛별유치원')가 정확하게 렌더링되었습니다.",
      "entities": "지시사항과 달리 이미경이 정면 얼굴로 등장하고, 전택수가 원경에 전신으로 서 있습니다. 또한 프레임 우측 근경에 지시되지 않은 미상의 인물(우산을 든 팔과 어깨)이 등장합니다.",
      "hard_violations": [
       "지시된 카메라 프레이밍(이미경의 어깨 너머로 전택수의 상체를 잡는 클로즈 미디엄 샷)을 완전히 위반함",
       "프레임 우측 근경에 텍스트가 지시하지 않은 신원 미상의 인물 파트(우산을 든 팔과 어깨)가 존재함 (extra bodies)"
      ],
      "physics": "근경의 미상 인물의 손이 우산을 지탱하고 있으며, 원경의 전택수는 땅에 제대로 서 있으나 샷의 맥락상 물리적 위치 관계가 지시사항과 완전히 틀립니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "지시된 오버더숄더 앵글과 전택수의 상체 프레이밍, 미소 짓는 표정 등을 완벽하게 구현했으나 유치원 간판에 약간의 오타가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "지시된 카메라 앵글을 완전히 무시하고 이전 컷의 구도를 그대로 가져왔으며, 근경에 우산을 든 미상의 인물이 남은 상태에서 배경에 전택수를 추가해 인물 중복 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "전택수의 시선과 우산의 방향이 프레임 왼쪽 근경에 걸쳐진 이미경을 향하고 있습니다.",
      "built_space": "배경에 유치원 건물의 입구와 계단이 올바르게 위치해 있으나, 간판 텍스트에 오타('새삘유치원')가 보입니다.",
      "entities": "전택수는 레퍼런스와 일치하는 얼굴과 복장으로 상체만 등장하며, 이미경은 지시대로 왼쪽 가장자리에 어깨와 뒷모습만 보입니다. 검은색 우산의 형태와 재질도 잘 표현되었습니다.",
      "hard_violations": [],
      "physics": "전택수의 손이 우산 손잡이를 자연스럽게 쥐고 있으며, 우산의 각도와 무게 지탱이 물리적으로 타당합니다."
     },
     {
      "label": "B",
      "direction": "근경의 이미경은 아래를 향해 시선을 두고 있으며, 원경의 전택수는 정면 카메라 방향을 응시하고 있습니다.",
      "built_space": "배경에 유치원 건물과 입구가 보이며, 간판 글씨('샛별유치원')가 정확하게 렌더링되었습니다.",
      "entities": "지시사항과 달리 이미경이 정면 얼굴로 등장하고, 전택수가 원경에 전신으로 서 있습니다. 또한 프레임 우측 근경에 지시되지 않은 미상의 인물(우산을 든 팔과 어깨)이 등장합니다.",
      "hard_violations": [
       "지시된 카메라 프레이밍(이미경의 어깨 너머로 전택수의 상체를 잡는 클로즈 미디엄 샷)을 완전히 위반함",
       "프레임 우측 근경에 텍스트가 지시하지 않은 신원 미상의 인물 파트(우산을 든 팔과 어깨)가 존재함 (extra bodies)"
      ],
      "physics": "근경의 미상 인물의 손이 우산을 지탱하고 있으며, 원경의 전택수는 땅에 제대로 서 있으나 샷의 맥락상 물리적 위치 관계가 지시사항과 완전히 틀립니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "요구된 카메라 앵글과 숄더샷 구도를 완전히 무시하고 이전 샷의 정면 구도를 그대로 사용했으며, 우산을 든 팔 외에 배경에 지시받지 않은 방식으로 전택수를 추가 배치하여 프레임 내 인물이 중복되는 치명적인 오류를 범했습니다."
     },
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "이미경의 어깨를 근경에 걸고 우산을 들어 올리며 미소 짓는 전택수를 미디엄 샷으로 담아내라는 카메라 구도와 행동 지시, 인물의 외양 및 조명 분위기를 모두 완벽하게 구현했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "정면을 향해 시선을 아래로 둔 여성, 우측에서 들어온 우산을 든 알 수 없는 팔, 배경에서 정면을 응시하는 남성이 각기 다른 방향을 향함.",
      "built_space": "뒷배경에 유치원 입구와 계단, 간판이 보이며 이전 샷의 공간과 동일함.",
      "entities": "이전 샷의 여성(정면), 우측의 우산을 든 팔, 배경의 전택수(레퍼런스 복장 일치) 등 세 명의 존재가 나타남.",
      "hard_violations": [
       "지시된 카메라 앵글 및 구도(오버 더 숄더 샷) 완전 무시",
       "우산을 든 제3의 팔과 배경의 인물이 동시에 존재하는 신체/인물 중복 (extra bodies)"
      ],
      "physics": "우산을 든 우측의 팔이 우산을 지탱하고 있으나, 배경의 남성은 가만히 서 있음."
     },
     {
      "label": "B",
      "direction": "전택수의 시선과 우산의 방향이 화면 좌측 근경에 있는 이미경을 자연스럽게 향하고 있음.",
      "built_space": "뒷배경에 샛별유치원 입구와 계단이 위치하며 심도와 공간감이 적절함.",
      "entities": "좌측 가장자리의 이미경(어깨 프로필), 우산을 들고 미소 짓는 전택수(레퍼런스 얼굴, 복장 일치), 그리고 레퍼런스와 일치하는 검은 우산이 정확하게 묘사됨.",
      "hard_violations": [],
      "physics": "전택수의 오른손이 우산 손잡이를 단단히 쥐고 올바른 자세로 우산을 지탱하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "요구된 카메라 앵글과 숄더샷 구도를 완전히 무시하고 이전 샷의 정면 구도를 그대로 사용했으며, 우산을 든 팔 외에 배경에 지시받지 않은 방식으로 전택수를 추가 배치하여 프레임 내 인물이 중복되는 치명적인 오류를 범했습니다."
     },
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "이미경의 어깨를 근경에 걸고 우산을 들어 올리며 미소 짓는 전택수를 미디엄 샷으로 담아내라는 카메라 구도와 행동 지시, 인물의 외양 및 조명 분위기를 모두 완벽하게 구현했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "정면을 향해 시선을 아래로 둔 여성, 우측에서 들어온 우산을 든 알 수 없는 팔, 배경에서 정면을 응시하는 남성이 각기 다른 방향을 향함.",
      "built_space": "뒷배경에 유치원 입구와 계단, 간판이 보이며 이전 샷의 공간과 동일함.",
      "entities": "이전 샷의 여성(정면), 우측의 우산을 든 팔, 배경의 전택수(레퍼런스 복장 일치) 등 세 명의 존재가 나타남.",
      "hard_violations": [
       "지시된 카메라 앵글 및 구도(오버 더 숄더 샷) 완전 무시",
       "우산을 든 제3의 팔과 배경의 인물이 동시에 존재하는 신체/인물 중복 (extra bodies)"
      ],
      "physics": "우산을 든 우측의 팔이 우산을 지탱하고 있으나, 배경의 남성은 가만히 서 있음."
     },
     {
      "label": "A",
      "direction": "전택수의 시선과 우산의 방향이 화면 좌측 근경에 있는 이미경을 자연스럽게 향하고 있음.",
      "built_space": "뒷배경에 샛별유치원 입구와 계단이 위치하며 심도와 공간감이 적절함.",
      "entities": "좌측 가장자리의 이미경(어깨 프로필), 우산을 들고 미소 짓는 전택수(레퍼런스 얼굴, 복장 일치), 그리고 레퍼런스와 일치하는 검은 우산이 정확하게 묘사됨.",
      "hard_violations": [],
      "physics": "전택수의 오른손이 우산 손잡이를 단단히 쥐고 올바른 자세로 우산을 지탱하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 19,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S83sh1"
  },
  "lane_policy": "ab_select_bypass:prev"
 },
 "S83sh6::cine": {
  "applied": true,
  "fingerprint": "ccbd4325ef3aef78ea49ec82eac37fbd0d440a06f57c5c50a9d91cd5110e7748",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S83sh6_sel.png",
  "source_sha256": "214850bac42ea4e9a28fdc977841a030a88006ed99ececc4d08883cdf265a89c",
  "file": "S83sh6_cine.png",
  "latency_ms": 12663
 },
 "S84sh1::signage": {
  "fp": "83b5c8cfe861e533",
  "inscriptions": [
   {
    "surface_native": "법원 청사 외벽",
    "text_native": "광주지방법원",
    "reason_ko": "2000년대 한국의 지방법원 건물 정면을 보여주는 장면이므로 사실적인 묘사를 위해 법원의 공식 명칭이 표시되어야 합니다."
   }
  ]
 },
 "S84sh1": {
  "input_fingerprint": "4f1927bbf0a6724f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, sunny.\n\nSHOT TEXT (authoritative, Korean): 맑고 화창한 햇빛이 내리쬐는 웅장한 광주지방법원 건물 정면의 전경\n\nLOCATION (lock): In the open courthouse plaza, facing the grand front facade under bright sunshine. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: On the courthouse's frontal axis, the slow push settles at a moderately elevated long distance with the complete front elevation centered and fully legible. The building occupies the central and upper frame while sufficient foreground remains below it to preserve its scale and institutional weight.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 광주지방법원 building (Fully visible in the exterior establishing view) — Its complete front elevation faces the camera along the building's frontal axis; used as Primary architectural subject and establishing image for the courtroom sequence.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Clear, bright daytime sunlight renders the courthouse in a naturalistic neutral palette with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 청사 외벽: \"광주지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, sunny.\n\nSHOT TEXT (authoritative, Korean): 맑고 화창한 햇빛이 내리쬐는 웅장한 광주지방법원 건물 정면의 전경\n\nLOCATION (lock): In the open courthouse plaza, facing the grand front facade under bright sunshine. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: On the courthouse's frontal axis, the slow push settles at a moderately elevated long distance with the complete front elevation centered and fully legible. The building occupies the central and upper frame while sufficient foreground remains below it to preserve its scale and institutional weight.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 광주지방법원 building (Fully visible in the exterior establishing view) — Its complete front elevation faces the camera along the building's frontal axis; used as Primary architectural subject and establishing image for the courtroom sequence.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Clear, bright daytime sunlight renders the courthouse in a naturalistic neutral palette with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 청사 외벽: \"광주지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, sunny.\n\nSHOT TEXT (authoritative, Korean): 맑고 화창한 햇빛이 내리쬐는 웅장한 광주지방법원 건물 정면의 전경\n\nLOCATION (lock): In the open courthouse plaza, facing the grand front facade under bright sunshine. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: On the courthouse's frontal axis, the slow push settles at a moderately elevated long distance with the complete front elevation centered and fully legible. The building occupies the central and upper frame while sufficient foreground remains below it to preserve its scale and institutional weight.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 광주지방법원 building (Fully visible in the exterior establishing view) — Its complete front elevation faces the camera along the building's frontal axis; used as Primary architectural subject and establishing image for the courtroom sequence.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Clear, bright daytime sunlight renders the courthouse in a naturalistic neutral palette with moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 청사 외벽: \"광주지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "카메라가 법원 건물 정면을 똑바로 향함.",
    "built_space": "넓은 포장 광장, 계단, 대칭형 기둥이 있는 법원 건물 정면 구조가 정확히 배치됨.",
    "entities": "광주지방법원 건물, 정확히 쓰인 '광주지방법원' 외벽 간판, 레퍼런스와 일치하는 파란색 지붕의 '자유평등정의' 텍스트.",
    "hard_violations": [],
    "physics": "모든 건축물과 광장이 지면에 안정적으로 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "카메라가 법원 건물 정면을 똑바로 향함.",
    "built_space": "넓은 포장 광장, 계단, 대칭형 기둥이 있는 법원 건물 정면 구조가 배치됨.",
    "entities": "광주지방법원 건물, 정확히 쓰인 '광주지방법원' 외벽 간판, 텍스트가 비어있는 파란색 지붕.",
    "hard_violations": [],
    "physics": "모든 건축물과 광장이 지면에 안정적으로 고정되어 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "레퍼런스 이미지에 확립된 파란색 구조물의 '자유평등정의' 텍스트까지 정확히 유지하여 지시사항을 훌륭하게 따름."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "지정된 간판과 건물 형태는 잘 구현했으나, 레퍼런스에 존재하던 파란색 구조물의 텍스트가 누락됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S72sh1_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "카메라 구도를 복사하지 말라는 참조 지침과 달리, 참조 이미지와 완전히 동일한 구도로 렌더링되었습니다.",
     "fix_en": "Crop the image slightly to alter the framing so it does not identically match the reference photo, preserving the central courthouse building, lighting, and overall scene.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "카메라 구도를 복사하지 말라는 참조 지침과 달리, 참조 이미지와 완전히 동일한 구도로 렌더링되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 건물 정면 축이 아니라 왼쪽 사선에 있어 정면 입면이 카메라를 마주하지 않는다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "카메라 구도를 복사하지 말라는 참조 지침과 달리, 참조 이미지와 완전히 동일한 구도로 렌더링되었습니다.",
    "fix_en": "Crop the image slightly to alter the framing so it does not identically match the reference photo, preserving the central courthouse building, lighting, and overall scene.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S72sh1"
  },
  "lane_policy": "ab_select_bypass:bg_only:share_plan_prev_bgonly"
 },
 "S84sh1::cine": {
  "applied": true,
  "fingerprint": "81f8ed6434da1ddbcf3a59408deeaa21cfaadedc1374da0f6cb6f3fd618d73d7",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S84sh1_sel.png",
  "source_sha256": "3d9025d8c206c0ed13ab2fd17ca606fc74c1b5a19114409909344301728f6dec",
  "file": "S84sh1_cine.png",
  "latency_ms": 9693
 },
 "S85sh17::signage": {
  "fp": "ce947ee0959420a5",
  "inscriptions": []
 },
 "S85sh17": {
  "input_fingerprint": "d768633529a7bffb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이미경 쪽을 향해 한쪽 입꼬리를 올린 채 옅은 미소를 짓고 있는 지국현의 얄미운 표정 클로즈업.\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the witness side at seated eye height, the dolly-in ends on a tight three-quarter close-up of 지국현 (현재) across the courtroom. His face occupies most of the frame as his eyes stay fixed on 이미경 and one corner of his mouth lifts into a faint, needling smile; the defendant's area remains only as restrained contextual blur.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: defendant's seating area (Occupied by 지국현 (현재)) — The seating area is seen obliquely from the witness side; used as Provides minimal spatial context behind the isolated expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination keeps the palette subdued and the facial contrast moderate.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom layout, screen and witness area, formal lighting, and defendant's clothing. Exclude the attorney and projected photograph from the close frame, isolating the defendant's slight smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이미경 쪽을 향해 한쪽 입꼬리를 올린 채 옅은 미소를 짓고 있는 지국현의 얄미운 표정 클로즈업.\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the witness side at seated eye height, the dolly-in ends on a tight three-quarter close-up of 지국현 (현재) across the courtroom. His face occupies most of the frame as his eyes stay fixed on 이미경 and one corner of his mouth lifts into a faint, needling smile; the defendant's area remains only as restrained contextual blur.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: defendant's seating area (Occupied by 지국현 (현재)) — The seating area is seen obliquely from the witness side; used as Provides minimal spatial context behind the isolated expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination keeps the palette subdued and the facial contrast moderate.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom layout, screen and witness area, formal lighting, and defendant's clothing. Exclude the attorney and projected photograph from the close frame, isolating the defendant's slight smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 이미경 쪽을 향해 한쪽 입꼬리를 올린 채 옅은 미소를 짓고 있는 지국현의 얄미운 표정 클로즈업.\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the witness stand. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the witness side at seated eye height, the dolly-in ends on a tight three-quarter close-up of 지국현 (현재) across the courtroom. His face occupies most of the frame as his eyes stay fixed on 이미경 and one corner of his mouth lifts into a faint, needling smile; the defendant's area remains only as restrained contextual blur.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: defendant's seating area (Occupied by 지국현 (현재)) — The seating area is seen obliquely from the witness side; used as Provides minimal spatial context behind the isolated expression.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination keeps the palette subdued and the facial contrast moderate.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom layout, screen and witness area, formal lighting, and defendant's clothing. Exclude the attorney and projected photograph from the close frame, isolating the defendant's slight smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "화면 왼쪽 밖(증인석 방향)을 응시하며, 한쪽 입꼬리를 올린 옅은 미소를 지음.",
    "built_space": "흐리게 처리된 법정의 나무 벽면과 창문이 보이며, 모순되는 명패나 요소가 없음.",
    "entities": "지국현의 얼굴 및 체형이 레퍼런스와 일치하며, '교도소 1234'가 적힌 수의를 입음.",
    "hard_violations": [],
    "physics": "보이지 않는 의자에 자연스럽게 앉아 체중을 지탱함."
   },
   {
    "label": "B",
    "direction": "화면 왼쪽 밖을 응시하며 옅은 미소를 지음.",
    "built_space": "법정 내부 배경에 창문과 나무 벽면이 있고, 앞쪽에 '검사석'이라고 적힌 명패와 책상이 배치됨.",
    "entities": "지국현의 인상착의 및 수의 디테일이 레퍼런스와 일치함.",
    "hard_violations": [
     "프롬프트는 피고인석(defendant's table)을 지시했으나, 수의를 입은 피고인이 '검사석' 명패가 놓인 자리에 앉아 있어 인물 배치 및 공간 설정 위반."
    ],
    "physics": "의자에 앉아 있는 자세로 물리적 지지 상태는 정상적임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 3
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "요구된 타이트한 클로즈업 프레이밍을 잘 따랐으며, 피고인석이라는 설정에 맞게 배경을 적절히 블러 처리하여 지정된 미소를 충실히 구현함."
   },
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "수의를 입은 피고인을 '검사석' 명패가 있는 자리에 배치하는 심각한 위치 설정 오류를 범함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S80sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "생성된 인물의 얼굴이 캐릭터 레퍼런스(지국현)가 아닌 이전 샷(PREVIOUS SHOT STILL)에 등장하는 검사의 얼굴로 렌더링되었습니다.",
     "fix_en": "Replace the man's face with the face of the young man from the character reference (지국현), ensuring he has a faint, needling smile. Preserve the blue prison uniform, the blurred wooden courtroom background, the lighting, and the current camera framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "타이트한 3/4 클로즈업이 아니며 얼굴이 프레임 대부분을 차지하지 않고 상반신이 크게 보인다.",
     "fix_en": "Crop and redraw the image to a tight close-up where the face occupies most of the frame.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "생성된 인물의 얼굴이 캐릭터 레퍼런스(지국현)가 아닌 이전 샷(PREVIOUS SHOT STILL)에 등장하는 검사의 얼굴로 렌더링되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "지국현의 얼굴이 캐릭터 레퍼런스와 일치하지 않는다.",
     "severity": "major"
    },
    {
     "issue_ko": "타이트한 3/4 클로즈업이 아니며 얼굴이 프레임 대부분을 차지하지 않고 상반신이 크게 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "타이트한 3/4 클로즈업이 아니며 얼굴이 프레임 대부분을 차지하지 않고 상반신이 크게 보인다.",
    "fix_en": "Crop and redraw the image to a tight close-up where the face occupies most of the frame.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the man's face with the face of the young man from the character reference (지국현), ensuring he has a faint, needling smile. Preserve the blue prison uniform, the blurred wooden courtroom background, the lighting, and the current camera framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "두 후보 모두 이전 샷 인물의 얼굴을 그대로 가져오는 치명적 오류를 범했으나, A가 요구된 얄미운 미소와 명찰 디테일을 그나마 더 온전하게 표현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "이전 샷 인물의 얼굴을 차용하여 신원 지시를 완전히 위반했으며, 표정이 다소 밋밋하고 명찰 하단이 잘려 보입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "화면 우측 화면 밖 증인석 방향으로 시선이 고정됨.",
      "built_space": "블러 처리된 법정 내부의 나무 패널 벽면이 배경으로 배치됨.",
      "entities": "파란색 수의(교도소 1234)를 입은 남성. 한쪽 입꼬리가 올라간 미소. 얼굴 생김새가 캐릭터 레퍼런스가 아닌 이전 샷의 변호사와 동일함.",
      "hard_violations": [
       "이전 샷 인물의 얼굴을 그대로 차용함 (캐릭터 레퍼런스 완전 불일치 및 배제 지시 위반)"
      ],
      "physics": "피고인석 의자에 앉아 상체가 안정적으로 지지된 상태."
     },
     {
      "label": "B",
      "direction": "화면 우측 화면 밖 증인석 방향으로 시선이 고정됨.",
      "built_space": "블러 처리된 법정 내부의 나무 패널 벽면이 배경으로 배치됨.",
      "entities": "파란색 수의(명찰 하단 숫자 잘림)를 입은 남성. 옅은 미소. 얼굴 생김새가 캐릭터 레퍼런스가 아닌 이전 샷의 변호사와 동일함.",
      "hard_violations": [
       "이전 샷 인물의 얼굴을 그대로 차용함 (캐릭터 레퍼런스 완전 불일치 및 배제 지시 위반)"
      ],
      "physics": "피고인석 의자에 앉아 상체가 안정적으로 지지된 상태."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "두 후보 모두 이전 샷 인물의 얼굴을 그대로 가져오는 치명적 오류를 범했으나, A가 요구된 얄미운 미소와 명찰 디테일을 그나마 더 온전하게 표현했습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "이전 샷 인물의 얼굴을 차용하여 신원 지시를 완전히 위반했으며, 표정이 다소 밋밋하고 명찰 하단이 잘려 보입니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "화면 우측 화면 밖 증인석 방향으로 시선이 고정됨.",
      "built_space": "블러 처리된 법정 내부의 나무 패널 벽면이 배경으로 배치됨.",
      "entities": "파란색 수의(교도소 1234)를 입은 남성. 한쪽 입꼬리가 올라간 미소. 얼굴 생김새가 캐릭터 레퍼런스가 아닌 이전 샷의 변호사와 동일함.",
      "hard_violations": [
       "이전 샷 인물의 얼굴을 그대로 차용함 (캐릭터 레퍼런스 완전 불일치 및 배제 지시 위반)"
      ],
      "physics": "피고인석 의자에 앉아 상체가 안정적으로 지지된 상태."
     },
     {
      "label": "B",
      "direction": "화면 우측 화면 밖 증인석 방향으로 시선이 고정됨.",
      "built_space": "블러 처리된 법정 내부의 나무 패널 벽면이 배경으로 배치됨.",
      "entities": "파란색 수의(명찰 하단 숫자 잘림)를 입은 남성. 옅은 미소. 얼굴 생김새가 캐릭터 레퍼런스가 아닌 이전 샷의 변호사와 동일함.",
      "hard_violations": [
       "이전 샷 인물의 얼굴을 그대로 차용함 (캐릭터 레퍼런스 완전 불일치 및 배제 지시 위반)"
      ],
      "physics": "피고인석 의자에 앉아 상체가 안정적으로 지지된 상태."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시문이 요구한 '한쪽 입꼬리를 올린 채' 짓는 얄미운 미소를 비대칭적인 입 모양으로 정확하게 구현하여 의도를 잘 살렸습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "캐릭터와 배경은 잘 구현되었으나, 미소가 비교적 대칭적이어서 '한쪽 입꼬리를 올린' 디테일이 다소 부족합니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "시선은 프레임 왼쪽(증인석 방향)을 향하고 있습니다.",
      "built_space": "배경은 흐리게 처리된 법정 내부로, 나무 패널 벽과 창문이 보입니다.",
      "entities": "지국현(현재)이 파란색 죄수복('교도소 1234' 명찰)을 입고 등장합니다.",
      "hard_violations": [],
      "physics": "자연스럽게 앉아있는 자세로, 등받이 일부가 보여 의자에 지지되어 있음이 확인됩니다."
     },
     {
      "label": "A",
      "direction": "시선은 프레임 왼쪽(증인석 방향)을 향하고 있습니다.",
      "built_space": "배경은 흐리게 처리된 법정 내부로, 나무 패널 벽과 창문이 보입니다.",
      "entities": "지국현(현재)이 파란색 죄수복('교도소 1234' 명찰)을 입고 등장합니다.",
      "hard_violations": [],
      "physics": "자연스럽게 앉아있는 자세로, 등받이 일부가 보여 의자에 지지되어 있음이 확인됩니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시문이 요구한 '한쪽 입꼬리를 올린 채' 짓는 얄미운 미소를 비대칭적인 입 모양으로 정확하게 구현하여 의도를 잘 살렸습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "캐릭터와 배경은 잘 구현되었으나, 미소가 비교적 대칭적이어서 '한쪽 입꼬리를 올린' 디테일이 다소 부족합니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "시선은 프레임 왼쪽(증인석 방향)을 향하고 있습니다.",
      "built_space": "배경은 흐리게 처리된 법정 내부로, 나무 패널 벽과 창문이 보입니다.",
      "entities": "지국현(현재)이 파란색 죄수복('교도소 1234' 명찰)을 입고 등장합니다.",
      "hard_violations": [],
      "physics": "자연스럽게 앉아있는 자세로, 등받이 일부가 보여 의자에 지지되어 있음이 확인됩니다."
     },
     {
      "label": "B",
      "direction": "시선은 프레임 왼쪽(증인석 방향)을 향하고 있습니다.",
      "built_space": "배경은 흐리게 처리된 법정 내부로, 나무 패널 벽과 창문이 보입니다.",
      "entities": "지국현(현재)이 파란색 죄수복('교도소 1234' 명찰)을 입고 등장합니다.",
      "hard_violations": [],
      "physics": "자연스럽게 앉아있는 자세로, 등받이 일부가 보여 의자에 지지되어 있음이 확인됩니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 11,
     "B": 10
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S80sh4"
  }
 },
 "S85sh17::cine": {
  "applied": true,
  "fingerprint": "c32683a1628c2e4dd1e5614d49ec7a9b5a9ae93ffa826f4eb51a977d2a86f388",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S85sh17_sel.png",
  "source_sha256": "4fe49a1eb086a48b47892ae5aa5916fce2d23b6bf6596bde442775d8c7b653cf",
  "file": "S85sh17_cine.png",
  "latency_ms": 12007
 },
 "S85sh19::signage": {
  "fp": "2e4c5874fd4da446",
  "inscriptions": [
   {
    "surface_native": "증인석 명패",
    "text_native": "증인석",
    "reason_ko": "법정 내 증언대 전면에 설치된 명패를 통해 이미경이 앉아 있는 자리가 증인석임을 시각적으로 명확히 전달하기 위함."
   }
  ]
 },
 "S85sh19": {
  "input_fingerprint": "df19b8137d0be6d8",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 증언석 마이크를 향해 목을 곧게 세운 채 입술을 살짝 뗀 이미경의 결연한 상체.\n\nLOCATION (lock): At the courtroom witness stand, directly before its microphone. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the witness position and slightly below 이미경's seated eye line, the inward move settles into a firm upper-body three-quarter frame angled across the microphone. 이미경 straightens her neck and parts her lips toward the microphone while fixing her gaze across the courtroom on 지국현 (현재), her body held steady against the surrounding silence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: witness microphone (Positioned before 이미경 as she begins her testimony) — Its speaking end angles upward from the foreground toward 이미경; used as Crosses the lower foreground and directs the composition toward 이미경's mouth without obscuring her face; witness stand (Occupied by 이미경) — Its inner side surrounds 이미경 and is viewed from an oblique side angle; used as Frames her seated posture and identifies her formal role in the courtroom.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination holds a subdued neutral palette and clear but unsensational facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same witness stand, courtroom materials, microphone, and daylight. Exclude the defendant and attorneys from the close composition, showing only the witness speaking resolutely.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 증언석 마이크를 향해 목을 곧게 세운 채 입술을 살짝 뗀 이미경의 결연한 상체.\n\nLOCATION (lock): At the courtroom witness stand, directly before its microphone. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the witness position and slightly below 이미경's seated eye line, the inward move settles into a firm upper-body three-quarter frame angled across the microphone. 이미경 straightens her neck and parts her lips toward the microphone while fixing her gaze across the courtroom on 지국현 (현재), her body held steady against the surrounding silence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: witness microphone (Positioned before 이미경 as she begins her testimony) — Its speaking end angles upward from the foreground toward 이미경; used as Crosses the lower foreground and directs the composition toward 이미경's mouth without obscuring her face; witness stand (Occupied by 이미경) — Its inner side surrounds 이미경 and is viewed from an oblique side angle; used as Frames her seated posture and identifies her formal role in the courtroom.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination holds a subdued neutral palette and clear but unsensational facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same witness stand, courtroom materials, microphone, and daylight. Exclude the defendant and attorneys from the close composition, showing only the witness speaking resolutely.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 증언석 마이크를 향해 목을 곧게 세운 채 입술을 살짝 뗀 이미경의 결연한 상체.\n\nLOCATION (lock): At the courtroom witness stand, directly before its microphone. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside the witness position and slightly below 이미경's seated eye line, the inward move settles into a firm upper-body three-quarter frame angled across the microphone. 이미경 straightens her neck and parts her lips toward the microphone while fixing her gaze across the courtroom on 지국현 (현재), her body held steady against the surrounding silence.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: witness microphone (Positioned before 이미경 as she begins her testimony) — Its speaking end angles upward from the foreground toward 이미경; used as Crosses the lower foreground and directs the composition toward 이미경's mouth without obscuring her face; witness stand (Occupied by 이미경) — Its inner side surrounds 이미경 and is viewed from an oblique side angle; used as Frames her seated posture and identifies her formal role in the courtroom.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination holds a subdued neutral palette and clear but unsensational facial detail.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same witness stand, courtroom materials, microphone, and daylight. Exclude the defendant and attorneys from the close composition, showing only the witness speaking resolutely.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "시선은 화면 밖 대상을 향함. 입술은 굳게 닫혀 있음.",
    "built_space": "법정 증인석 단상, 스탠드 마이크, 카메라를 향한 명패 위치함.",
    "entities": "이미경 캐릭터(외모, 의상) 일치. '증인석' 텍스트 정확함.",
    "hard_violations": [],
    "physics": "단상에 팔을 얹고 안정적으로 착석함. 마이크와 명패는 지지면에 놓임."
   },
   {
    "label": "B",
    "direction": "시선은 화면 밖을 향하며, 입술이 살짝 벌어져 마이크를 향함.",
    "built_space": "법정 증인석 단상, 구스넥 마이크, 카메라를 향한 명패 위치함.",
    "entities": "이미경 캐릭터(외모, 의상) 일치. '증인석' 텍스트 정확함.",
    "hard_violations": [],
    "physics": "단상에 팔을 얹고 안정적으로 착석함. 물체들은 올바르게 지지됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "입술을 살짝 뗀 표정과 시선 처리 등 프롬프트의 세부 지시사항을 정확히 반영함."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "전반적인 구성과 명패는 정확하나 '입술을 살짝 뗀' 묘사가 누락됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S85sh17_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "이미경의 얼굴·이목구비가 캐릭터 레퍼런스와 다른 사람으로 보임",
     "fix_en": "Modify the woman's facial features to exactly match the character reference image, preserving her current expression, pose, clothing, the microphone, the wooden desk, and the courtroom background.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "이미경 분홍 카디건 왼쪽 가슴에 레퍼런스의 명찰이 없음",
     "fix_en": "Add the small name tag and round badge from the character reference to the left chest of the pink cardigan, preserving the woman's face, pose, all other clothing, the microphone, the desk, and the courtroom background.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "증인석 명패 글자가 '증인석'이 아니라 글자 사이가 띄어져 있음",
     "fix_en": "Remove the spaces between the characters on the nameplate so it reads exactly '증인석', preserving the nameplate's overall design, the wooden desk, the woman, the microphone, and the courtroom background.",
     "severity": "minor",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "레퍼런스 이미지에서 인물의 왼쪽 가슴에 부착된 배지와 이름표가 생성된 이미지의 의상에서는 누락되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "이미경의 얼굴·이목구비가 캐릭터 레퍼런스와 다른 사람으로 보임",
     "severity": "major"
    },
    {
     "issue_ko": "이미경 분홍 카디건 왼쪽 가슴에 레퍼런스의 명찰이 없음",
     "severity": "major"
    },
    {
     "issue_ko": "증인석 명패 글자가 '증인석'이 아니라 글자 사이가 띄어져 있음",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "이미경의 얼굴·이목구비가 캐릭터 레퍼런스와 다른 사람으로 보임",
    "fix_en": "Modify the woman's facial features to exactly match the character reference image, preserving her current expression, pose, clothing, the microphone, the wooden desk, and the courtroom background.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "이미경 분홍 카디건 왼쪽 가슴에 레퍼런스의 명찰이 없음",
    "fix_en": "Add the small name tag and round badge from the character reference to the left chest of the pink cardigan, preserving the woman's face, pose, all other clothing, the microphone, the desk, and the courtroom background.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "증인석 명패 글자가 '증인석'이 아니라 글자 사이가 띄어져 있음",
    "fix_en": "Remove the spaces between the characters on the nameplate so it reads exactly '증인석', preserving the nameplate's overall design, the wooden desk, the woman, the microphone, and the courtroom background.",
    "severity": "minor",
    "observation_index": 3
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S85sh17"
  }
 },
 "S85sh19::cine": {
  "applied": true,
  "fingerprint": "a6fb4a4a8dfa908bdc5978cf3a591af51314a60285df9d597ed64578c7588699",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S85sh19_sel.png",
  "source_sha256": "fbde3fd0c6779e6365dff10e1b1b74e84ed0ae56f9fac5ed5121aeaee7498d9c",
  "file": "S85sh19_cine.png",
  "latency_ms": 10972
 },
 "S85sh21::signage": {
  "fp": "dfeae7511b4123e0",
  "inscriptions": [
   {
    "surface_native": "피고인석 명패",
    "text_native": "피고인",
    "reason_ko": "대한민국 법정의 피고인 테이블 앞에 놓인 신분 표시용 명패로, 현재 지국현이 피고인석에 앉아 있음을 보여주는 장치입니다."
   }
  ]
 },
 "S85sh21": {
  "input_fingerprint": "deacf35eaf719a94",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 당황한 기색으로 동공이 미세하게 확장된 지국현의 얼굴 클로즈업.\n\nLOCATION (lock): At the defendant's table inside the courtroom, under the witness's direct gaze. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the witness side and slightly above 지국현 (현재)'s seated eye line, the static camera holds the tightest non-extreme three-quarter facial close-up of the reaction. His composure breaks in a minute backward recoil, his pupils subtly widening as he continues to look toward 이미경 before he can turn away.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: defendant's seating area (Occupied by 지국현 (현재)) — The seating area is viewed from a slightly elevated witness-side diagonal; used as Leaves a narrow contextual trace around the close reaction without competing with the face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination preserves the subtle eye reaction with moderate contrast and a restrained neutral tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom lighting, defendant's seat, formal wood finishes, and prison clothing. Exclude the witness from the close frame and isolate the defendant's alarmed reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피고인석 명패: \"피고인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 당황한 기색으로 동공이 미세하게 확장된 지국현의 얼굴 클로즈업.\n\nLOCATION (lock): At the defendant's table inside the courtroom, under the witness's direct gaze. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the witness side and slightly above 지국현 (현재)'s seated eye line, the static camera holds the tightest non-extreme three-quarter facial close-up of the reaction. His composure breaks in a minute backward recoil, his pupils subtly widening as he continues to look toward 이미경 before he can turn away.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: defendant's seating area (Occupied by 지국현 (현재)) — The seating area is viewed from a slightly elevated witness-side diagonal; used as Leaves a narrow contextual trace around the close reaction without competing with the face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination preserves the subtle eye reaction with moderate contrast and a restrained neutral tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom lighting, defendant's seat, formal wood finishes, and prison clothing. Exclude the witness from the close frame and isolate the defendant's alarmed reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피고인석 명패: \"피고인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 당황한 기색으로 동공이 미세하게 확장된 지국현의 얼굴 클로즈업.\n\nLOCATION (lock): At the defendant's table inside the courtroom, under the witness's direct gaze. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From the witness side and slightly above 지국현 (현재)'s seated eye line, the static camera holds the tightest non-extreme three-quarter facial close-up of the reaction. His composure breaks in a minute backward recoil, his pupils subtly widening as he continues to look toward 이미경 before he can turn away.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: defendant's seating area (Occupied by 지국현 (현재)) — The seating area is viewed from a slightly elevated witness-side diagonal; used as Leaves a narrow contextual trace around the close reaction without competing with the face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Naturalistic daytime courtroom illumination preserves the subtle eye reaction with moderate contrast and a restrained neutral tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom lighting, defendant's seat, formal wood finishes, and prison clothing. Exclude the witness from the close frame and isolate the defendant's alarmed reaction.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피고인석 명패: \"피고인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.333,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.667,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.667,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1667,
   "B": 1714
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1667,
    "verdict_ko": "참조 이미지의 수의 명찰 텍스트('교도소 1234')를 정확히 재현했으며, 미세하게 당황하며 뒤로 물러나는 자세를 잘 표현함."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "지시된 위치와 표정은 무난하게 연출되었으나, 수의 명찰에 '교도스'라는 오타가 발생하여 의상 재현도가 떨어짐."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S85sh19_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프레이밍이 프롬프트에서 요구한 타이트한 얼굴 클로즈업(close-up)이 아니라, 가슴 부위까지 널찍하게 잡힌 미디엄 클로즈업/미디엄 샷으로 연출되었습니다.",
     "fix_en": "Crop the image to a tight facial close-up of the man, removing the desk and his lower torso from the frame. Maintain the man's identity, his facial expression, his prison uniform, the courtroom background, and the lighting.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프레이밍이 프롬프트에서 요구한 타이트한 얼굴 클로즈업(close-up)이 아니라, 가슴 부위까지 널찍하게 잡힌 미디엄 클로즈업/미디엄 샷으로 연출되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 클로즈업이 아니라 상반신과 책상·피고인 명패가 크게 보이는 미디엄 클로즈업이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "프레이밍이 프롬프트에서 요구한 타이트한 얼굴 클로즈업(close-up)이 아니라, 가슴 부위까지 널찍하게 잡힌 미디엄 클로즈업/미디엄 샷으로 연출되었습니다.",
    "fix_en": "Crop the image to a tight facial close-up of the man, removing the desk and his lower torso from the frame. Maintain the man's identity, his facial expression, his prison uniform, the courtroom background, and the lighting.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S85sh19"
  }
 },
 "S85sh21::cine": {
  "applied": true,
  "fingerprint": "0f62982e09737cb057e21dd77aba4e2373ddf20042200d8c62f54c1b0e33970a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S85sh21_sel.png",
  "source_sha256": "084cf40d04df4e0d8ee9fd5451ac265bd101c3de0c1c7d2b9a17a8dd74e5f03f",
  "file": "S85sh21_cine.png",
  "latency_ms": 11633
 },
 "S86sh4::signage": {
  "fp": "9a6dbaa187614c78",
  "inscriptions": [
   {
    "surface_native": "법원 광장의 표지석",
    "text_native": "법원",
    "reason_ko": "법원 광장이라는 공간적 배경을 사실적으로 나타내기 위해 배경에 위치한 표지석에 '법원'이라는 글자가 표시되어야 합니다."
   }
  ]
 },
 "S86sh4": {
  "input_fingerprint": "ff9b4ff92a4ee2eb",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 쪽을 향해 고개를 돌린 채 물끄러미 바라보는 이미경의 측면.\n\nLOCATION (lock): On the open courthouse plaza, some distance from the building entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From shoulder height beside 이미경, hold a static close profile with her occupying the left half of frame and open looking room extending toward the off-screen 전택수. Her body remains aligned with her departing path while her head has turned fully toward him, isolating the moment of recognition without breaking the lateral scene axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 광주지방법원 건물 전면 (Visible in the daytime exterior) — The building frontage is seen obliquely from the side of 이미경's path; used as Soft spatial context behind 이미경 and along her departing path.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with a restrained neutral-to-cool palette and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courthouse facade, open plaza, bright sunlight, and exterior paving. Exclude arriving vehicles and show the woman turning back across the plaza.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 광장의 표지석: \"법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 쪽을 향해 고개를 돌린 채 물끄러미 바라보는 이미경의 측면.\n\nLOCATION (lock): On the open courthouse plaza, some distance from the building entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From shoulder height beside 이미경, hold a static close profile with her occupying the left half of frame and open looking room extending toward the off-screen 전택수. Her body remains aligned with her departing path while her head has turned fully toward him, isolating the moment of recognition without breaking the lateral scene axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 광주지방법원 건물 전면 (Visible in the daytime exterior) — The building frontage is seen obliquely from the side of 이미경's path; used as Soft spatial context behind 이미경 and along her departing path.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with a restrained neutral-to-cool palette and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courthouse facade, open plaza, bright sunlight, and exterior paving. Exclude arriving vehicles and show the woman turning back across the plaza.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 광장의 표지석: \"법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수 쪽을 향해 고개를 돌린 채 물끄러미 바라보는 이미경의 측면.\n\nLOCATION (lock): On the open courthouse plaza, some distance from the building entrance. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From shoulder height beside 이미경, hold a static close profile with her occupying the left half of frame and open looking room extending toward the off-screen 전택수. Her body remains aligned with her departing path while her head has turned fully toward him, isolating the moment of recognition without breaking the lateral scene axis.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 광주지방법원 건물 전면 (Visible in the daytime exterior) — The building frontage is seen obliquely from the side of 이미경's path; used as Soft spatial context behind 이미경 and along her departing path.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light with a restrained neutral-to-cool palette and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courthouse facade, open plaza, bright sunlight, and exterior paving. Exclude arriving vehicles and show the woman turning back across the plaza.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 이미경 (Korean 여성, 30대 초반, 타원형 얼굴, 검은색 중간 길이 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 광장의 표지석: \"법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 1714,
   "A": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "지정된 클로즈업 샷보다 프레임이 다소 넓게 잡혔으나, 몸은 진행 방향에 두고 고개만 돌려 대상을 바라보는 프롬프트의 핵심 행동 지시를 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "프레임 크기는 클로즈업에 조금 더 가깝지만, 인물의 몸통과 시선이 모두 같은 방향을 향하고 있어 고개만 돌린 채 바라보는 연출을 실패했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S84sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 이미경: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832451>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 인물의 몸이 걷던 경로를 유지한 채 고개만 완전히 돌린 자세를 지시했으나, 이미지에서는 어깨와 몸통 전체가 시선과 같은 방향(화면 오른쪽)을 향하고 있음.",
     "fix_en": "Redraw the woman's torso to face the departing path while keeping her head turned right. Preserve her face, hair, clothing, the background, and the lighting.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "캐릭터 레퍼런스 이미지의 왼쪽 가슴에 부착된 명찰과 배지가 생성된 이미지에서는 누락됨.",
     "fix_en": "Render the small badge and name tag on the left chest of the pink cardigan. Preserve the woman's face, pose, clothing, the background, and the lighting.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "이미경이 좌측 절반을 차지하는 정면 클로즈 프로파일이 아니라 중간 거리에서 더 많이 보이는 3/4 상반신으로 잡혀 있다.",
     "fix_en": "Crop the image to simulate a closer profile framing of the woman's head and shoulders. Preserve all character details, clothing, and the background.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이미경의 얼굴이 캐릭터 레퍼런스의 타원형 얼굴·이목구비와 일치하지 않는다.",
     "fix_en": "Redraw the woman's face to exactly match the oval face shape and specific facial features of the character reference. Preserve her head angle, hair, pink cardigan, white blouse, the background courthouse plaza, and the lighting.",
     "severity": "critical",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 인물의 몸이 걷던 경로를 유지한 채 고개만 완전히 돌린 자세를 지시했으나, 이미지에서는 어깨와 몸통 전체가 시선과 같은 방향(화면 오른쪽)을 향하고 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "캐릭터 레퍼런스 이미지의 왼쪽 가슴에 부착된 명찰과 배지가 생성된 이미지에서는 누락됨.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경이 좌측 절반을 차지하는 정면 클로즈 프로파일이 아니라 중간 거리에서 더 많이 보이는 3/4 상반신으로 잡혀 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "몸이 떠나는 경로와 정렬된 채 머리만 돌린 것이 아니라 몸통과 얼굴이 함께 오른쪽을 향해 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경의 얼굴이 캐릭터 레퍼런스의 타원형 얼굴·이목구비와 일치하지 않는다.",
     "severity": "critical"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 인물의 몸이 걷던 경로를 유지한 채 고개만 완전히 돌린 자세를 지시했으나, 이미지에서는 어깨와 몸통 전체가 시선과 같은 방향(화면 오른쪽)을 향하고 있음.",
    "fix_en": "Redraw the woman's torso to face the departing path while keeping her head turned right. Preserve her face, hair, clothing, the background, and the lighting.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "캐릭터 레퍼런스 이미지의 왼쪽 가슴에 부착된 명찰과 배지가 생성된 이미지에서는 누락됨.",
    "fix_en": "Render the small badge and name tag on the left chest of the pink cardigan. Preserve the woman's face, pose, clothing, the background, and the lighting.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "이미경이 좌측 절반을 차지하는 정면 클로즈 프로파일이 아니라 중간 거리에서 더 많이 보이는 3/4 상반신으로 잡혀 있다.",
    "fix_en": "Crop the image to simulate a closer profile framing of the woman's head and shoulders. Preserve all character details, clothing, and the background.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the woman's face to exactly match the oval face shape and specific facial features of the character reference. Preserve her head angle, hair, pink cardigan, white blouse, the background courthouse plaza, and the lighting.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시문이 요구한 앵글, 배경 구조물(광주지방법원, 표지석)과 텍스트를 완벽하게 구현했으며, 레퍼런스 인물의 부드러운 얼굴선과 정체성을 가장 정확하게 유지했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "구도와 배경 디테일은 훌륭하게 유지되었으나, 측면 얼굴(코와 턱선)이 날카롭게 변형되고 뒷머리가 어색하게 갈라지는 등 캐릭터의 정체성 보존 측면에서 손실이 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "이미경이 프레임 오른쪽(화면 밖 전택수 방향)을 향해 고개를 돌리고 시선을 고정하고 있음.",
      "built_space": "광주지방법원 전면이 배경에 올바르게 위치하며, 현관의 '자유평등정의'와 상단의 '광주지방법원', 중경의 '법원' 표지석 텍스트가 정확한 위치에 렌더링됨.",
      "entities": "이미경(30대 초반 여성, 검은 머리, 분홍색 가디건, 리본 블라우스)의 외형과 의상이 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "광장 바닥에 안정적으로 서 있으며 물리적으로 어색한 부분 없이 자연스러움."
     },
     {
      "label": "B",
      "direction": "이미경이 프레임 오른쪽(화면 밖 전택수 방향)을 응시하고 있음.",
      "built_space": "A와 동일하게 광주지방법원 건물, 현관, 표지석 및 텍스트들이 정확한 위치에 묘사됨.",
      "entities": "복장과 기본 외형은 레퍼런스와 일치하나, 코와 턱선이 날카로워져 레퍼런스의 부드러운 인상과 다소 차이가 생겼음.",
      "hard_violations": [],
      "physics": "광장 바닥에 안정적으로 지탱해 서 있으나, 등 뒤로 떨어지는 머리카락이 부자연스럽게 갈라져 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시문이 요구한 앵글, 배경 구조물(광주지방법원, 표지석)과 텍스트를 완벽하게 구현했으며, 레퍼런스 인물의 부드러운 얼굴선과 정체성을 가장 정확하게 유지했습니다."
     },
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "구도와 배경 디테일은 훌륭하게 유지되었으나, 측면 얼굴(코와 턱선)이 날카롭게 변형되고 뒷머리가 어색하게 갈라지는 등 캐릭터의 정체성 보존 측면에서 손실이 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "이미경이 프레임 오른쪽(화면 밖 전택수 방향)을 향해 고개를 돌리고 시선을 고정하고 있음.",
      "built_space": "광주지방법원 전면이 배경에 올바르게 위치하며, 현관의 '자유평등정의'와 상단의 '광주지방법원', 중경의 '법원' 표지석 텍스트가 정확한 위치에 렌더링됨.",
      "entities": "이미경(30대 초반 여성, 검은 머리, 분홍색 가디건, 리본 블라우스)의 외형과 의상이 레퍼런스와 정확히 일치함.",
      "hard_violations": [],
      "physics": "광장 바닥에 안정적으로 서 있으며 물리적으로 어색한 부분 없이 자연스러움."
     },
     {
      "label": "B",
      "direction": "이미경이 프레임 오른쪽(화면 밖 전택수 방향)을 응시하고 있음.",
      "built_space": "A와 동일하게 광주지방법원 건물, 현관, 표지석 및 텍스트들이 정확한 위치에 묘사됨.",
      "entities": "복장과 기본 외형은 레퍼런스와 일치하나, 코와 턱선이 날카로워져 레퍼런스의 부드러운 인상과 다소 차이가 생겼음.",
      "hard_violations": [],
      "physics": "광장 바닥에 안정적으로 지탱해 서 있으나, 등 뒤로 떨어지는 머리카락이 부자연스럽게 갈라져 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "요구된 화면 구도와 인물의 시선, 지정된 의상과 표지석의 '법원' 텍스트를 모두 완벽하게 구현했으며, 배경의 간판 텍스트가 약간 더 선명합니다."
     },
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "프롬프트의 모든 조건을 훌륭하게 충족했으나, 텍스트의 선명도에서 후보 B에 비해 아주 미세하게 떨어집니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "이미경의 시선과 얼굴 방향이 프레임 우측 오프스크린(전택수)을 올바르게 향하고 있음.",
      "built_space": "이전 숏과 동일한 광주지방법원 건물 전경, 파란색 캐노피, 광장 중앙에 3단 받침을 가진 표지석이 위치함.",
      "entities": "이미경의 얼굴, 헤어스타일, 핑크 가디건, 흰색 리본 블라우스, 민트색 치마가 레퍼런스와 정확히 일치하며, 표지석에 '법원' 텍스트가 명확히 새겨짐.",
      "hard_violations": [],
      "physics": "인물이 광장 바닥에 안정적으로 서 있으며, 광원과 그림자의 방향이 자연스럽고 지지점에 문제가 없음."
     },
     {
      "label": "A",
      "direction": "이미경의 시선과 얼굴 방향이 프레임 우측 오프스크린(전택수)을 향하고 있음.",
      "built_space": "광주지방법원 건물 전경과 입구 캐노피, 광장 중앙의 표지석 배치가 정확함.",
      "entities": "이미경의 외모와 지정된 의상이 모두 일치하며, 표지석의 '법원' 텍스트 및 배경 건물의 텍스트가 존재함.",
      "hard_violations": [],
      "physics": "인물이 광장에 안정적으로 서 있으며 물리적으로 어색한 묘사나 지지 불량 문제가 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "요구된 화면 구도와 인물의 시선, 지정된 의상과 표지석의 '법원' 텍스트를 모두 완벽하게 구현했으며, 배경의 간판 텍스트가 약간 더 선명합니다."
     },
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "프롬프트의 모든 조건을 훌륭하게 충족했으나, 텍스트의 선명도에서 후보 B에 비해 아주 미세하게 떨어집니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "이미경의 시선과 얼굴 방향이 프레임 우측 오프스크린(전택수)을 올바르게 향하고 있음.",
      "built_space": "이전 숏과 동일한 광주지방법원 건물 전경, 파란색 캐노피, 광장 중앙에 3단 받침을 가진 표지석이 위치함.",
      "entities": "이미경의 얼굴, 헤어스타일, 핑크 가디건, 흰색 리본 블라우스, 민트색 치마가 레퍼런스와 정확히 일치하며, 표지석에 '법원' 텍스트가 명확히 새겨짐.",
      "hard_violations": [],
      "physics": "인물이 광장 바닥에 안정적으로 서 있으며, 광원과 그림자의 방향이 자연스럽고 지지점에 문제가 없음."
     },
     {
      "label": "B",
      "direction": "이미경의 시선과 얼굴 방향이 프레임 우측 오프스크린(전택수)을 향하고 있음.",
      "built_space": "광주지방법원 건물 전경과 입구 캐노피, 광장 중앙의 표지석 배치가 정확함.",
      "entities": "이미경의 외모와 지정된 의상이 모두 일치하며, 표지석의 '법원' 텍스트 및 배경 건물의 텍스트가 존재함.",
      "hard_violations": [],
      "physics": "인물이 광장에 안정적으로 서 있으며 물리적으로 어색한 묘사나 지지 불량 문제가 없음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 20,
     "B": 17
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S84sh1"
  }
 },
 "S86sh4::cine": {
  "applied": true,
  "fingerprint": "b4788b07491cb29d7f418ddfd090d9fae56519c144c6ce65f56f9fd5687ab7bf",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S86sh4_sel.png",
  "source_sha256": "ad0419cfa6dff47286e42120d2f73f32a9b07fbfb5873c816feffb43cff12bc3",
  "file": "S86sh4_cine.png",
  "latency_ms": 12068
 },
 "S86sh6::signage": {
  "fp": "3e90fa783b3b6c7b",
  "inscriptions": [
   {
    "surface_native": "법원 입구 표석",
    "text_native": "지방법원",
    "reason_ko": "법원 광장이라는 공간적 배경의 설득력을 높이기 위해 입구 표석에 실제 한국 법원 스타일의 명칭을 표시합니다."
   }
  ]
 },
 "S86sh6": {
  "input_fingerprint": "debac297021cb80d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거리를 둔 채 이미경을 향해 허리를 깊숙이 굽혀 깍듯이 인사하는 전택수의 굳건한 자세.\n\nLOCATION (lock): Across the courthouse plaza, with a respectful distance between the two figures. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At lower-torso height and long full-body distance, complete the dolly out from one side of the line between 전택수 and 이미경, placing 전택수 in the left midground and the distant 이미경 across the right background. Hold at the deepest point of his bow so his full-body form and the unclosed distance between them remain equally legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 이미경 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방법원 건물 전면 (Visible in the daytime exterior) — The frontage is viewed obliquely behind the separated figures; used as Provides the exterior setting while the lateral composition preserves the distance between the two figures; 법원 앞 외부 공간 (The two figures remain separated across it); used as Forms the uninterrupted depth interval between 전택수 and 이미경.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light keeps the farewell restrained in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courthouse plaza, bright daylight, paving, and the two characters' clothing. Exclude the woman's earlier turn as the focal action and show the man bowing deeply at a distance.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 입구 표석: \"지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거리를 둔 채 이미경을 향해 허리를 깊숙이 굽혀 깍듯이 인사하는 전택수의 굳건한 자세.\n\nLOCATION (lock): Across the courthouse plaza, with a respectful distance between the two figures. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At lower-torso height and long full-body distance, complete the dolly out from one side of the line between 전택수 and 이미경, placing 전택수 in the left midground and the distant 이미경 across the right background. Hold at the deepest point of his bow so his full-body form and the unclosed distance between them remain equally legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 이미경 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방법원 건물 전면 (Visible in the daytime exterior) — The frontage is viewed obliquely behind the separated figures; used as Provides the exterior setting while the lateral composition preserves the distance between the two figures; 법원 앞 외부 공간 (The two figures remain separated across it); used as Forms the uninterrupted depth interval between 전택수 and 이미경.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light keeps the farewell restrained in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courthouse plaza, bright daylight, paving, and the two characters' clothing. Exclude the woman's earlier turn as the focal action and show the man bowing deeply at a distance.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 입구 표석: \"지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 거리를 둔 채 이미경을 향해 허리를 깊숙이 굽혀 깍듯이 인사하는 전택수의 굳건한 자세.\n\nLOCATION (lock): Across the courthouse plaza, with a respectful distance between the two figures. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At lower-torso height and long full-body distance, complete the dolly out from one side of the line between 전택수 and 이미경, placing 전택수 in the left midground and the distant 이미경 across the right background. Hold at the deepest point of his bow so his full-body form and the unclosed distance between them remain equally legible.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground; 이미경 in the middle-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 광주지방법원 건물 전면 (Visible in the daytime exterior) — The frontage is viewed obliquely behind the separated figures; used as Provides the exterior setting while the lateral composition preserves the distance between the two figures; 법원 앞 외부 공간 (The two figures remain separated across it); used as Forms the uninterrupted depth interval between 전택수 and 이미경.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural daytime ambient light keeps the farewell restrained in subdued neutral-to-cool tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courthouse plaza, bright daylight, paving, and the two characters' clothing. Exclude the woman's earlier turn as the focal action and show the man bowing deeply at a distance.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법원 입구 표석: \"지방법원\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "전택수는 우측의 이미경을 향해 인사하고, 이미경은 정면을 응시함.",
    "built_space": "광주지방법원 건물이 배경에 있고 중간에 '지방법원' 표석이 배치됨.",
    "entities": "전택수(수사 복장)와 이미경(가디건 차림)의 외형이 지정된 대로 표현됨.",
    "hard_violations": [],
    "physics": "두 사람 모두 바닥을 단단히 디디고 서 있으며 인사 자세가 자연스러움."
   },
   {
    "label": "B",
    "direction": "전택수는 우측을 향해 인사하고, 이미경은 전택수를 바라봄.",
    "built_space": "법원 건물이 배경에 있고 중간에 '법원' 표석이 위치함.",
    "entities": "전택수와 이미경의 의상이 참조 이미지 및 지시와 일치함.",
    "hard_violations": [],
    "physics": "두 인물 모두 지면에 안정적으로 서 있으나 프레임 하단이 잘려 발목 아래가 보이지 않음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "표석에 요구된 '지방법원' 텍스트를 정확히 렌더링하고 전신의 모습이 온전히 담겨 프레이밍 지시를 더 잘 따름."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "표석 텍스트 지시를 누락하여 '법원'으로 표기했으며, 발 부분이 잘려 전신 풀샷 요구를 온전히 충족하지 못함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S86sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 이미경을 우측 원경(background)에 멀리 배치할 것을 지시했으나, 전택수와 비슷한 깊이의 전경/중경에 크게 배치되어 거리감과 구도 지시를 위반함.",
     "fix_en": "Scale down the woman on the right and adjust her placement to make her appear distant in the background. Preserve the bowing man on the left, both characters' clothing, the plaza paving, and the courthouse exterior.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경의 법원 건물 입구 파란색 간판 글씨(레퍼런스의 '자유평등정의') 중 세 번째 글자가 왜곡되어 '동'처럼 잘못 렌더링됨.",
     "fix_en": "Correct the distorted third character on the blue awning to clearly read '정', restoring the full '자유평등정의' text. Preserve both characters, their poses and clothing, the plaza ground, and the courthouse building.",
     "severity": "minor",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 이미경을 우측 원경(background)에 멀리 배치할 것을 지시했으나, 전택수와 비슷한 깊이의 전경/중경에 크게 배치되어 거리감과 구도 지시를 위반함.",
     "severity": "major"
    },
    {
     "issue_ko": "배경의 법원 건물 입구 파란색 간판 글씨(레퍼런스의 '자유평등정의') 중 세 번째 글자가 왜곡되어 '동'처럼 잘못 렌더링됨.",
     "severity": "minor"
    },
    {
     "issue_ko": "전택수가 왼쪽 중경이 아니라 화면 왼쪽 전경에 크게 가까이 서 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "이미경이 오른쪽 원경이 아니라 중경 오른쪽에 비교적 크게 서 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "건물 입구 파란 현판 글자가 이전 장면의 ‘자유평등정의’와 달리 ‘자유평등’만 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 이미경을 우측 원경(background)에 멀리 배치할 것을 지시했으나, 전택수와 비슷한 깊이의 전경/중경에 크게 배치되어 거리감과 구도 지시를 위반함.",
    "fix_en": "Scale down the woman on the right and adjust her placement to make her appear distant in the background. Preserve the bowing man on the left, both characters' clothing, the plaza paving, and the courthouse exterior.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경의 법원 건물 입구 파란색 간판 글씨(레퍼런스의 '자유평등정의') 중 세 번째 글자가 왜곡되어 '동'처럼 잘못 렌더링됨.",
    "fix_en": "Correct the distorted third character on the blue awning to clearly read '정', restoring the full '자유평등정의' text. Preserve both characters, their poses and clothing, the plaza ground, and the courthouse building.",
    "severity": "minor",
    "observation_index": 1
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S86sh4"
  }
 },
 "S86sh6::cine": {
  "applied": true,
  "fingerprint": "3af087252f05cfc44b25a6d16b6eabc9493af271561495e812f946045ce08140",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S86sh6_sel.png",
  "source_sha256": "1dc1fc94697a08ac883dcc3088f53e460e4a79d8806359f77ec3fa147f2f2657",
  "file": "S86sh6_cine.png",
  "latency_ms": 10676
 },
 "S87sh1::signage": {
  "fp": "3a1a399f4ef531c0",
  "inscriptions": [
   {
    "surface_native": "법정 피고석 테이블 위의 명패",
    "text_native": "변호인",
    "reason_ko": "대한민국 법정 내부의 사실감을 높이기 위해 장원섭이 서 있는 테이블 위에 한국어로 적힌 '변호인' 명패가 필요합니다."
   }
  ]
 },
 "S87sh1": {
  "input_fingerprint": "764b9d5719725312",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 법대 위의 부장판사를 향해 낡은 다이어리를 두 손으로 뻗어 내민 장원섭의 상체.\n\nLOCATION (lock): Before the raised judges' bench inside the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside the central courtroom axis, dolly in on 장원섭 from his three-quarter side as his upper body occupies the near center and both hands extend the worn diary diagonally toward the elevated bench. Keep 부장판사 smaller in the upper background, receiving the line of the gesture without flattening the difference in height.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-center of the frame, midground, reaches for elevated court bench; 부장판사 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 낡은 다이어리 (Worn and being submitted) — The opened written side is angled toward 부장판사, with the camera seeing it obliquely between 장원섭's hands; used as Primary evidence object forming a diagonal from 장원섭's hands toward the bench; 법대 (Occupied by 부장판사 during proceedings) — Its front and upper working side are seen diagonally from the central courtroom axis; used as Establishes the judge's elevated position at the end of the evidence diagonal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination with subdued cool tonality and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, judge's bench, formal woodwork, daylight, and prosecutor's attire. Exclude the alibi photograph and defense attorney, showing the diary being submitted.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Wonseop holds out Seonyeong's worn diary to the presiding judge as evidence.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 피고석 테이블 위의 명패: \"변호인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 법대 위의 부장판사를 향해 낡은 다이어리를 두 손으로 뻗어 내민 장원섭의 상체.\n\nLOCATION (lock): Before the raised judges' bench inside the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside the central courtroom axis, dolly in on 장원섭 from his three-quarter side as his upper body occupies the near center and both hands extend the worn diary diagonally toward the elevated bench. Keep 부장판사 smaller in the upper background, receiving the line of the gesture without flattening the difference in height.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-center of the frame, midground, reaches for elevated court bench; 부장판사 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 낡은 다이어리 (Worn and being submitted) — The opened written side is angled toward 부장판사, with the camera seeing it obliquely between 장원섭's hands; used as Primary evidence object forming a diagonal from 장원섭's hands toward the bench; 법대 (Occupied by 부장판사 during proceedings) — Its front and upper working side are seen diagonally from the central courtroom axis; used as Establishes the judge's elevated position at the end of the evidence diagonal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination with subdued cool tonality and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, judge's bench, formal woodwork, daylight, and prosecutor's attire. Exclude the alibi photograph and defense attorney, showing the diary being submitted.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Wonseop holds out Seonyeong's worn diary to the presiding judge as evidence.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 피고석 테이블 위의 명패: \"변호인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 법대 위의 부장판사를 향해 낡은 다이어리를 두 손으로 뻗어 내민 장원섭의 상체.\n\nLOCATION (lock): Before the raised judges' bench inside the courtroom. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At chest height beside the central courtroom axis, dolly in on 장원섭 from his three-quarter side as his upper body occupies the near center and both hands extend the worn diary diagonally toward the elevated bench. Keep 부장판사 smaller in the upper background, receiving the line of the gesture without flattening the difference in height.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 장원섭 in the middle-center of the frame, midground, reaches for elevated court bench; 부장판사 in the upper-right of the frame, background.\n- KEY BACKGROUND ELEMENTS: 낡은 다이어리 (Worn and being submitted) — The opened written side is angled toward 부장판사, with the camera seeing it obliquely between 장원섭's hands; used as Primary evidence object forming a diagonal from 장원섭's hands toward the bench; 법대 (Occupied by 부장판사 during proceedings) — Its front and upper working side are seen diagonally from the central courtroom axis; used as Establishes the judge's elevated position at the end of the evidence diagonal.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination with subdued cool tonality and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, judge's bench, formal woodwork, daylight, and prosecutor's attire. Exclude the alibi photograph and defense attorney, showing the diary being submitted.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Wonseop holds out Seonyeong's worn diary to the presiding judge as evidence.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 장원섭 right now — 장원섭's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 장원섭 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법정 피고석 테이블 위의 명패: \"변호인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.2,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.8,
    "B": 1.833
   },
   "adjusted": {
    "A": 1.8,
    "B": 1.833
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1800,
   "B": 1833
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1800,
    "verdict_ko": "인물과 소품의 외형 및 '변호인' 명패는 훌륭하게 재현했으나, 다이어리의 펼친 면이 판사가 아닌 카메라를 향해 있어 지시된 방향성이 어긋남."
   },
   {
    "label": "B",
    "score": 1833,
    "verdict_ko": "명패와 소품 텍스트는 정확하나 다이어리가 카메라를 향해 있고, 손가락 묘사가 어색하며 죄수복의 명찰이 누락됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S85sh21_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:796050>"
   },
   {
    "label": "PROP REFERENCE — 김선영의 낡은 노트와 다이어리: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1240939>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "다이어리의 펼쳐진 필기면이 배경의 부장판사를 향해야 한다는 지시와 달리 카메라 정면을 향해 열려 있음.",
     "fix_en": "Redraw the worn diary and the hands holding it so that the open written pages angle away from the camera and point toward the background judge, showing the text obliquely. Preserve the people present and their positions, their clothing, the courtroom set, the light, and the framing.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "화면 한가운데 전경 인물이 장원섭 레퍼런스가 아니라 이전 스틸 피고인의 얼굴과 푸른 셔츠이다.",
     "fix_en": "Replace the foreground man's face and blue shirt with the oval face, neat black hair, and black robe with white cravat from the 장원섭 character reference. Preserve the people's positions, the background judge, the courtroom set, the worn diary, the light, and the framing.",
     "severity": "critical",
     "observation_index": 4
    },
    {
     "issue_ko": "장원섭 레퍼런스 얼굴의 인물이 오른쪽 위 배경 법대에 부장판사로 앉아 있다.",
     "fix_en": "Replace the background judge's face with a generic older Korean man to remove the 장원섭 character reference likeness from him. Preserve the people's positions, their clothing, the courtroom set, the worn diary, the light, and the framing.",
     "severity": "critical",
     "observation_index": 5
    },
    {
     "issue_ko": "전경 인물이 법대를 향해 다이어리를 뻗지 않고 몸 앞에서 들고만 있다.",
     "fix_en": "Redraw the foreground person's arms and hands to extend the diary diagonally forward and upward toward the judge's bench rather than holding it close to the chest. Preserve the person's identity, clothing, the diary, the courtroom set, the light, and the framing.",
     "severity": "major",
     "observation_index": 7
    },
    {
     "issue_ko": "전경 인물 시선이 부장판사가 아니라 카메라 왼쪽을 향한다.",
     "fix_en": "Redirect the foreground person's eyes so his gaze looks up and back toward the presiding judge on the bench instead of camera left. Preserve the face identity, clothing, pose, the courtroom set, the light, and the framing.",
     "severity": "major",
     "observation_index": 8
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "전경의 장원섭은 명시적으로 금지된 이전 컷 인물의 얼굴을 하고 있으며 배경의 판사는 장원섭의 캐릭터 레퍼런스 얼굴을 하고 있어 인물 식별이 완전히 뒤바뀜.",
     "severity": "critical"
    },
    {
     "issue_ko": "전경 인물이 이전 컷 이미지에서 가져오지 말라고 지시된 파란색 의상을 그대로 착용하고 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "다이어리의 펼쳐진 필기면이 배경의 부장판사를 향해야 한다는 지시와 달리 카메라 정면을 향해 열려 있음.",
     "severity": "critical"
    },
    {
     "issue_ko": "다이어리 우측 하단을 잡고 있는 전경 인물의 왼손 손가락이 비정상적으로 길고 구조가 뭉개져 있음.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 한가운데 전경 인물이 장원섭 레퍼런스가 아니라 이전 스틸 피고인의 얼굴과 푸른 셔츠이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "장원섭 레퍼런스 얼굴의 인물이 오른쪽 위 배경 법대에 부장판사로 앉아 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "펼친 다이어리 글자면이 법대 판사가 아니라 카메라 정면을 향한다.",
     "severity": "major"
    },
    {
     "issue_ko": "전경 인물이 법대를 향해 다이어리를 뻗지 않고 몸 앞에서 들고만 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "전경 인물 시선이 부장판사가 아니라 카메라 왼쪽을 향한다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "전경 인물이 법대를 향해 다이어리를 뻗지 않고 몸 앞에서 들고만 있다.",
    "fix_en": "Redraw the foreground person's arms and hands to extend the diary diagonally forward and upward toward the judge's bench rather than holding it close to the chest. Preserve the person's identity, clothing, the diary, the courtroom set, the light, and the framing.",
    "severity": "major",
    "observation_index": 7
   },
   {
    "issue_ko": "전경 인물 시선이 부장판사가 아니라 카메라 왼쪽을 향한다.",
    "fix_en": "Redirect the foreground person's eyes so his gaze looks up and back toward the presiding judge on the bench instead of camera left. Preserve the face identity, clothing, pose, the courtroom set, the light, and the framing.",
    "severity": "major",
    "observation_index": 8
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the worn diary and the hands holding it so that the open written pages angle away from the camera and point toward the background judge, showing the text obliquely. Preserve the people present and their positions, their clothing, the courtroom set, the light, and the framing.\n- Replace the foreground man's face and blue shirt with the oval face, neat black hair, and black robe with white cravat from the 장원섭 character reference. Preserve the people's positions, the background judge, the courtroom set, the worn diary, the light, and the framing.\n- Replace the background judge's face with a generic older Korean man to remove the 장원섭 character reference likeness from him. Preserve the people's positions, their clothing, the courtroom set, the worn diary, the light, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "장원섭이 판사를 향해 다이어리를 내미는 구도, '변호인' 명패, 캐릭터의 외모와 의상 등 텍스트의 요구사항을 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "지정된 인물이 아닌 엉뚱한 노인이 등장하고, 소품 레퍼런스의 회색 가이드 손이 그대로 화면에 합성되는 치명적인 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선과 다이어리를 쥔 두 손이 배경 우측 상단의 부장판사를 정확히 향하고 있습니다.",
      "built_space": "카메라 구도에 맞게 근경의 테이블, 중경의 장원섭, 원경의 판사석 및 판사가 올바른 깊이감으로 배치되었습니다.",
      "entities": "장원섭의 얼굴과 수의, 낡은 다이어리, 배경의 판사, '변호인' 명패가 프롬프트 및 레퍼런스와 정확히 일치합니다.",
      "hard_violations": [],
      "physics": "장원섭의 두 손이 다이어리의 양옆을 물리적으로 올바르게 쥐고 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "프레임 정중앙의 인물이 특정 대상 없이 카메라 정면 부근을 응시하고 있습니다.",
      "built_space": "지정된 대각선 구도와 깊이감이 무시되었으며, 판사석을 향하는 동선이 나타나지 않습니다.",
      "entities": "장원섭이 아닌 다른 노인이 등장했으며, 판사복에 죄수 번호표가 붙어 있고 명패도 '피고인'으로 잘못 유지되었습니다.",
      "hard_violations": [
       "invented people (요구되지 않은 노인 캐릭터 등장 및 장원섭 누락)",
       "leaked markers (레퍼런스 이미지의 회색 2D 손 실루엣이 그대로 출력됨)",
       "physically impossible staging (2D 그래픽 형태의 손이 입체적인 다이어리를 들고 있음)"
      ],
      "physics": "물리적인 실체가 없는 회색 2D 그래픽 손이 다이어리를 허공에서 받치고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "장원섭이 판사를 향해 다이어리를 내미는 구도, '변호인' 명패, 캐릭터의 외모와 의상 등 텍스트의 요구사항을 완벽하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "지정된 인물이 아닌 엉뚱한 노인이 등장하고, 소품 레퍼런스의 회색 가이드 손이 그대로 화면에 합성되는 치명적인 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "장원섭의 시선과 다이어리를 쥔 두 손이 배경 우측 상단의 부장판사를 정확히 향하고 있습니다.",
      "built_space": "카메라 구도에 맞게 근경의 테이블, 중경의 장원섭, 원경의 판사석 및 판사가 올바른 깊이감으로 배치되었습니다.",
      "entities": "장원섭의 얼굴과 수의, 낡은 다이어리, 배경의 판사, '변호인' 명패가 프롬프트 및 레퍼런스와 정확히 일치합니다.",
      "hard_violations": [],
      "physics": "장원섭의 두 손이 다이어리의 양옆을 물리적으로 올바르게 쥐고 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "프레임 정중앙의 인물이 특정 대상 없이 카메라 정면 부근을 응시하고 있습니다.",
      "built_space": "지정된 대각선 구도와 깊이감이 무시되었으며, 판사석을 향하는 동선이 나타나지 않습니다.",
      "entities": "장원섭이 아닌 다른 노인이 등장했으며, 판사복에 죄수 번호표가 붙어 있고 명패도 '피고인'으로 잘못 유지되었습니다.",
      "hard_violations": [
       "invented people (요구되지 않은 노인 캐릭터 등장 및 장원섭 누락)",
       "leaked markers (레퍼런스 이미지의 회색 2D 손 실루엣이 그대로 출력됨)",
       "physically impossible staging (2D 그래픽 형태의 손이 입체적인 다이어리를 들고 있음)"
      ],
      "physics": "물리적인 실체가 없는 회색 2D 그래픽 손이 다이어리를 허공에서 받치고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "다이어리 텍스트와 프레이밍, '변호인' 명패는 완벽히 구현했으나, 장원섭의 위치에 금지된 이전 샷 인물을 배치하고 참조 인물은 배경에 두는 치명적인 캐릭터 및 신원 배치 오류가 있습니다."
     },
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "레퍼런스의 마스킹 실루엣이 그래픽 오버레이로 그대로 노출되는 치명적 오류가 있으며, 지정되지 않은 노인 캐릭터 등장 및 명패 텍스트('피고인') 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "전경의 남성이 배경 상단의 판사석을 향해 다이어리를 대각선으로 뻗어 내밀고 있으며, 판사는 아래를 내려다봄.",
      "built_space": "법정 내부. 전경에 '변호인' 명패가 있는 책상이 있고, 상단 배경에 판사석이 배치되어 올바른 공간감과 높이 차이를 형성함.",
      "entities": "내용이 정확히 렌더링된 낡은 다이어리, '변호인' 명패, 파란 죄수복의 남성(이전 샷 인물), 검은 법복을 입은 배경의 판사(장원섭 참조 인물).",
      "hard_violations": [
       "지시문에서 명시적으로 배제한 이전 샷 인물의 외모와 복장을 전경의 장원섭 위치에 잘못 적용함",
       "장원섭 캐릭터 참조 인물을 배경의 판사로 잘못 배치함"
      ],
      "physics": "남성의 양손이 다이어리 양쪽을 물리적으로 단단히 잡고 허공으로 뻗어 안정적으로 지탱함."
     },
     {
      "label": "A",
      "direction": "화면 하단의 회색 실루엣 손이 다이어리를 펼쳐 정면의 노인에게 보여주고 있음.",
      "built_space": "법정 내부. 인물 앞쪽에 명패가 놓인 책상이 있음.",
      "entities": "낡은 다이어리, 텍스트('피고인')가 적힌 명패, 회색 그래픽 실루엣 손, 판사복에 죄수 번호표를 단 지정되지 않은 노인.",
      "hard_violations": [
       "참조 이미지의 회색 마스킹 실루엣이 물리적 질감 없이 그래픽 오버레이로 렌더링됨 (leaked markers)",
       "참조 이미지와 전혀 다른 새롭게 발명된 노인 캐릭터 등장",
       "판사복에 죄수 번호표가 붙은 모순되고 불가능한 복장"
      ],
      "physics": "물리적 실체가 없는 그래픽 마스킹 실루엣이 허공에서 다이어리를 지탱하고 있어 물리적 지지 위반."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "다이어리 텍스트와 프레이밍, '변호인' 명패는 완벽히 구현했으나, 장원섭의 위치에 금지된 이전 샷 인물을 배치하고 참조 인물은 배경에 두는 치명적인 캐릭터 및 신원 배치 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "레퍼런스의 마스킹 실루엣이 그래픽 오버레이로 그대로 노출되는 치명적 오류가 있으며, 지정되지 않은 노인 캐릭터 등장 및 명패 텍스트('피고인') 오류가 있습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "전경의 남성이 배경 상단의 판사석을 향해 다이어리를 대각선으로 뻗어 내밀고 있으며, 판사는 아래를 내려다봄.",
      "built_space": "법정 내부. 전경에 '변호인' 명패가 있는 책상이 있고, 상단 배경에 판사석이 배치되어 올바른 공간감과 높이 차이를 형성함.",
      "entities": "내용이 정확히 렌더링된 낡은 다이어리, '변호인' 명패, 파란 죄수복의 남성(이전 샷 인물), 검은 법복을 입은 배경의 판사(장원섭 참조 인물).",
      "hard_violations": [
       "지시문에서 명시적으로 배제한 이전 샷 인물의 외모와 복장을 전경의 장원섭 위치에 잘못 적용함",
       "장원섭 캐릭터 참조 인물을 배경의 판사로 잘못 배치함"
      ],
      "physics": "남성의 양손이 다이어리 양쪽을 물리적으로 단단히 잡고 허공으로 뻗어 안정적으로 지탱함."
     },
     {
      "label": "B",
      "direction": "화면 하단의 회색 실루엣 손이 다이어리를 펼쳐 정면의 노인에게 보여주고 있음.",
      "built_space": "법정 내부. 인물 앞쪽에 명패가 놓인 책상이 있음.",
      "entities": "낡은 다이어리, 텍스트('피고인')가 적힌 명패, 회색 그래픽 실루엣 손, 판사복에 죄수 번호표를 단 지정되지 않은 노인.",
      "hard_violations": [
       "참조 이미지의 회색 마스킹 실루엣이 물리적 질감 없이 그래픽 오버레이로 렌더링됨 (leaked markers)",
       "참조 이미지와 전혀 다른 새롭게 발명된 노인 캐릭터 등장",
       "판사복에 죄수 번호표가 붙은 모순되고 불가능한 복장"
      ],
      "physics": "물리적 실체가 없는 그래픽 마스킹 실루엣이 허공에서 다이어리를 지탱하고 있어 물리적 지지 위반."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 0
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S85sh21"
  }
 },
 "S87sh1::cine": {
  "applied": true,
  "fingerprint": "514589d6deee604081ea2828f71bc2d83a650c579444e2017a115c2f303d7495",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S87sh1_sel.png",
  "source_sha256": "b14eaab055c9cf108ee627fc6c082cbeca02808d1dee3bb68fa6d3231aa49090",
  "file": "S87sh1_cine.png",
  "latency_ms": 15238
 },
 "S87sh5::signage": {
  "fp": "af16e2a454ba5c5e",
  "inscriptions": [
   {
    "surface_native": "명패",
    "text_native": "검사",
    "reason_ko": "법정 내 검사석임을 명시하여 캐릭터의 역할과 대립 구도의 설득력을 높이기 위한 명패 표기."
   }
  ]
 },
 "S87sh5": {
  "input_fingerprint": "740eca3224fdf749",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현의 변호인 쪽을 향해 목에 핏대를 세운 채 분노로 입을 크게 벌린 장원섭의 단호한 얼굴.\n\nLOCATION (lock): At the prosecution side of the courtroom, turned toward the defense table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just beyond the defense side of the axis and slightly below 장원섭's eye line, finish the dolly in on a tight opposing three-quarter close-up, leaving a narrow strip of looking room toward the off-screen 지국현의 변호인. 장원섭's drawn brow, open mouth, and tightened neck carry the rebuttal while the courtroom recedes without introducing the lawyer into frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 장원섭 in the middle-center of the frame, foreground, looks toward off-screen defense counsel.\n- KEY BACKGROUND ELEMENTS: 법정 내부 (Court proceedings are in progress); used as Soft, limited context behind the isolated face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination preserves facial strain with moderate contrast and restrained cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, judge's bench, formal lighting, and prosecutor's clothing. Exclude the diary handover and isolate the prosecutor arguing toward the defense.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's diary remains before the court, open for the judge to inspect the dated “Magic” entries.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현의 변호인 쪽을 향해 목에 핏대를 세운 채 분노로 입을 크게 벌린 장원섭의 단호한 얼굴.\n\nLOCATION (lock): At the prosecution side of the courtroom, turned toward the defense table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just beyond the defense side of the axis and slightly below 장원섭's eye line, finish the dolly in on a tight opposing three-quarter close-up, leaving a narrow strip of looking room toward the off-screen 지국현의 변호인. 장원섭's drawn brow, open mouth, and tightened neck carry the rebuttal while the courtroom recedes without introducing the lawyer into frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 장원섭 in the middle-center of the frame, foreground, looks toward off-screen defense counsel.\n- KEY BACKGROUND ELEMENTS: 법정 내부 (Court proceedings are in progress); used as Soft, limited context behind the isolated face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination preserves facial strain with moderate contrast and restrained cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, judge's bench, formal lighting, and prosecutor's clothing. Exclude the diary handover and isolate the prosecutor arguing toward the defense.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's diary remains before the court, open for the judge to inspect the dated “Magic” entries.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 지국현의 변호인 쪽을 향해 목에 핏대를 세운 채 분노로 입을 크게 벌린 장원섭의 단호한 얼굴.\n\nLOCATION (lock): At the prosecution side of the courtroom, turned toward the defense table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Just beyond the defense side of the axis and slightly below 장원섭's eye line, finish the dolly in on a tight opposing three-quarter close-up, leaving a narrow strip of looking room toward the off-screen 지국현의 변호인. 장원섭's drawn brow, open mouth, and tightened neck carry the rebuttal while the courtroom recedes without introducing the lawyer into frame.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 장원섭 in the middle-center of the frame, foreground, looks toward off-screen defense counsel.\n- KEY BACKGROUND ELEMENTS: 법정 내부 (Court proceedings are in progress); used as Soft, limited context behind the isolated face.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination preserves facial strain with moderate contrast and restrained cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, judge's bench, formal lighting, and prosecutor's clothing. Exclude the diary handover and isolate the prosecutor arguing toward the defense.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's diary remains before the court, open for the judge to inspect the dated “Magic” entries.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 명패: \"검사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "장원섭의 시선과 고개가 화면 왼쪽 바깥(변호인석 방향)을 향해 있음.",
    "built_space": "'검사' 명패가 놓인 법정 책상 뒤편이며, 배경에 판사석이 위치함.",
    "entities": "장원섭(검은 법복, 분노한 표정), 'Magic'이 적힌 다이어리, 배경의 판사, 화면 왼쪽의 파란 소매를 입은 손.",
    "hard_violations": [
     "이전 스틸컷에 있던 배경의 판사를 그대로 포함해 '다른 인물 배제' 지시 위반.",
     "화면 왼쪽에서 다이어리를 쥐고 있는 파란 소매의 손이 등장하여 '다른 인물 및 신체 일부 배제' 지시 위반."
    ],
    "physics": "다이어리는 화면 밖 인물의 손에 의해 공중에 들려 지탱되고 있음."
   },
   {
    "label": "B",
    "direction": "장원섭의 시선과 고개가 화면 왼쪽 바깥(변호인석 방향)을 강하게 향해 있음.",
    "built_space": "'검사' 명패가 놓인 법정 책상 뒤편이며, 배경에는 빈 의자가 보임.",
    "entities": "장원섭(검은 법복, 핏대 선 목, 크게 벌린 입), 날짜가 적힌 펼쳐진 다이어리.",
    "hard_violations": [],
    "physics": "다이어리는 양쪽 표지 하단이 책상 바닥에 닿은 채 V자로 펼쳐져 스스로 서 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "이전 컷의 다른 인물들을 성공적으로 배제하고 장원섭의 단독 샷을 요구사항에 맞춰 잘 구현했으나, 법복 깃의 형태와 다이어리의 'Magic' 글자가 누락된 점이 아쉽습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "장원섭의 표정은 좋으나, 이전 스틸컷의 판사와 타인의 손(파란 소매)을 화면에 포함시켜 인물 배제 지시를 명백히 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 장원섭 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S87sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:796050>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 좌측 하단에 있는 펼쳐진 일기장이 손이나 다른 물리적 지지대 없이 불가능한 각도로 허공에 떠 있습니다.",
     "fix_en": "Redraw the floating diary so it lies flat and fully supported on the wooden desk surface, preserving the man, his pose, his clothing, and the courtroom background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "일기장 안의 내용이 레퍼런스 이미지의 빼곡한 영문 필기체와 다르게 날짜 두 줄만 남은 형태로 임의로 변형되었습니다.",
     "fix_en": "Fill the open diary pages with dense cursive writing to match the previous shot, preserving the man and the surrounding desk.",
     "severity": "major",
     "observation_index": 1
    },
    {
     "issue_ko": "타이트 클로즈업이 아니라 일기와 책상이 크게 보이는 미디엄 클로즈업이다.",
     "fix_en": "Crop the image tightly around the man's head and shoulders to simulate a close-up, preserving his facial expression and lighting.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "장원섭이 프레임 정중앙이 아니라 오른쪽에 치우쳐 있다.",
     "fix_en": "Crop the left side of the frame to center the man, preserving his pose, expression, and the background.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "레퍼런스 법복의 흰색 탭이 칼라 앞에 없다.",
     "fix_en": "Draw the two long white legal tabs hanging down over the black robe from the collar, preserving the man's face and the rest of his clothing.",
     "severity": "major",
     "observation_index": 5
    },
    {
     "issue_ko": "목에 핏대가 서지 않았다.",
     "fix_en": "Add prominent bulging veins to the man's neck to emphasize his shouting, preserving his facial expression and clothing.",
     "severity": "minor",
     "observation_index": 7
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 좌측 하단에 있는 펼쳐진 일기장이 손이나 다른 물리적 지지대 없이 불가능한 각도로 허공에 떠 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "일기장 안의 내용이 레퍼런스 이미지의 빼곡한 영문 필기체와 다르게 날짜 두 줄만 남은 형태로 임의로 변형되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 속 남성이 장원섭 캐릭터 레퍼런스가 아니라 이전 스틸의 다른 인물 얼굴이다.",
     "severity": "major"
    },
    {
     "issue_ko": "타이트 클로즈업이 아니라 일기와 책상이 크게 보이는 미디엄 클로즈업이다.",
     "severity": "major"
    },
    {
     "issue_ko": "장원섭이 프레임 정중앙이 아니라 오른쪽에 치우쳐 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "레퍼런스 법복의 흰색 탭이 칼라 앞에 없다.",
     "severity": "major"
    },
    {
     "issue_ko": "펼친 일기에 본문 없이 날짜만 있어 지정된 Magic 항목이 없다.",
     "severity": "major"
    },
    {
     "issue_ko": "목에 핏대가 서지 않았다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 6
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "일기장 안의 내용이 레퍼런스 이미지의 빼곡한 영문 필기체와 다르게 날짜 두 줄만 남은 형태로 임의로 변형되었습니다.",
    "fix_en": "Fill the open diary pages with dense cursive writing to match the previous shot, preserving the man and the surrounding desk.",
    "severity": "major",
    "observation_index": 1
   },
   {
    "issue_ko": "타이트 클로즈업이 아니라 일기와 책상이 크게 보이는 미디엄 클로즈업이다.",
    "fix_en": "Crop the image tightly around the man's head and shoulders to simulate a close-up, preserving his facial expression and lighting.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "장원섭이 프레임 정중앙이 아니라 오른쪽에 치우쳐 있다.",
    "fix_en": "Crop the left side of the frame to center the man, preserving his pose, expression, and the background.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "레퍼런스 법복의 흰색 탭이 칼라 앞에 없다.",
    "fix_en": "Draw the two long white legal tabs hanging down over the black robe from the collar, preserving the man's face and the rest of his clothing.",
    "severity": "major",
    "observation_index": 5
   },
   {
    "issue_ko": "목에 핏대가 서지 않았다.",
    "fix_en": "Add prominent bulging veins to the man's neck to emphasize his shouting, preserving his facial expression and clothing.",
    "severity": "minor",
    "observation_index": 7
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the floating diary so it lies flat and fully supported on the wooden desk surface, preserving the man, his pose, his clothing, and the courtroom background.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "다이어리를 책상 위에 눕혀 A의 물리적 오류를 완벽하게 수정하였으며, 프롬프트가 요구한 인물의 감정 표현과 상황을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "다이어리가 지탱하는 사람이나 장치 없이 허공에 떠 있는 형태(스티커처럼 붙여짐)로 심각한 물리적 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "화면 왼쪽 피고인석 방향을 향해 강하게 소리치고 있음.",
      "built_space": "법정 내부 검사석으로, 앞쪽에 책상과 명패가 바르게 위치함.",
      "entities": "장원섭(얼굴과 법복 모두 참조와 일치), '검사' 명패, 날짜가 적힌 다이어리 모두 일치.",
      "hard_violations": [],
      "physics": "다이어리가 책상 위에 평평하고 안정적으로 놓여 있음."
     },
     {
      "label": "A",
      "direction": "화면 왼쪽 피고인석 방향을 향해 강하게 소리치고 있음.",
      "built_space": "법정 내부 검사석으로, 앞쪽에 책상과 명패가 바르게 위치함.",
      "entities": "장원섭(얼굴과 법복 모두 참조와 일치), '검사' 명패, 날짜가 적힌 다이어리 모두 일치.",
      "hard_violations": [
       "물리적으로 불가능하게 공중에 떠 있는 다이어리(아무것도 지탱하지 않음)"
      ],
      "physics": "손으로 잡지 않은 다이어리가 책상 모서리에 세워져 있고, 왼쪽 끝부분은 허공에 떠 있어 지탱하는 것이 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "다이어리를 책상 위에 눕혀 A의 물리적 오류를 완벽하게 수정하였으며, 프롬프트가 요구한 인물의 감정 표현과 상황을 충실히 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "다이어리가 지탱하는 사람이나 장치 없이 허공에 떠 있는 형태(스티커처럼 붙여짐)로 심각한 물리적 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "화면 왼쪽 피고인석 방향을 향해 강하게 소리치고 있음.",
      "built_space": "법정 내부 검사석으로, 앞쪽에 책상과 명패가 바르게 위치함.",
      "entities": "장원섭(얼굴과 법복 모두 참조와 일치), '검사' 명패, 날짜가 적힌 다이어리 모두 일치.",
      "hard_violations": [],
      "physics": "다이어리가 책상 위에 평평하고 안정적으로 놓여 있음."
     },
     {
      "label": "A",
      "direction": "화면 왼쪽 피고인석 방향을 향해 강하게 소리치고 있음.",
      "built_space": "법정 내부 검사석으로, 앞쪽에 책상과 명패가 바르게 위치함.",
      "entities": "장원섭(얼굴과 법복 모두 참조와 일치), '검사' 명패, 날짜가 적힌 다이어리 모두 일치.",
      "hard_violations": [
       "물리적으로 불가능하게 공중에 떠 있는 다이어리(아무것도 지탱하지 않음)"
      ],
      "physics": "손으로 잡지 않은 다이어리가 책상 모서리에 세워져 있고, 왼쪽 끝부분은 허공에 떠 있어 지탱하는 것이 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 앵글과 인물의 격앙된 표정을 훌륭하게 구현했으며, 다이어리가 데스크에 자연스럽게 놓여 물리적 오류가 없습니다."
     },
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "표정과 앵글은 지시사항을 잘 따랐으나, 손이나 받침대 없이 다이어리가 허공에 뜬 것처럼 세워져 있는 심각한 물리적 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "장원섭이 화면 왼쪽 밖 변호인석을 향해 시선을 고정하고 강하게 소리치고 있음.",
      "built_space": "법정 내 목재 데스크와 '검사' 명패가 피사체 앞의 올바른 위치에 있음.",
      "entities": "장원섭의 얼굴, 머리스타일, 검사복이 레퍼런스와 일치하며, 책상 위에 텍스트가 적힌 다이어리가 놓여 있음.",
      "hard_violations": [],
      "physics": "인물은 데스크 뒤에 서 있고, 다이어리는 목재 데스크 표면 위에 안정적으로 평평하게 놓여 있어 물리적 지지가 명확함."
     },
     {
      "label": "B",
      "direction": "장원섭이 화면 왼쪽 밖 변호인석을 향해 시선을 고정하고 강하게 소리치고 있음.",
      "built_space": "법정 내 목재 데스크와 '검사' 명패가 피사체 앞의 올바른 위치에 있음.",
      "entities": "장원섭의 얼굴, 머리스타일, 검사복이 레퍼런스와 일치하며, 텍스트가 적힌 다이어리가 존재함.",
      "hard_violations": [
       "지지대 없이 부자연스럽게 서 있는 다이어리"
      ],
      "physics": "인물은 안정적으로 서 있으나, 다이어리가 손이나 다른 지지물 없이 V자 형태로 데스크 모서리에 비스듬히 서 있어 중력 및 물리 법칙에 위배됨."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "지시된 클로즈업 앵글과 인물의 격앙된 표정을 훌륭하게 구현했으며, 다이어리가 데스크에 자연스럽게 놓여 물리적 오류가 없습니다."
     },
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "표정과 앵글은 지시사항을 잘 따랐으나, 손이나 받침대 없이 다이어리가 허공에 뜬 것처럼 세워져 있는 심각한 물리적 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "장원섭이 화면 왼쪽 밖 변호인석을 향해 시선을 고정하고 강하게 소리치고 있음.",
      "built_space": "법정 내 목재 데스크와 '검사' 명패가 피사체 앞의 올바른 위치에 있음.",
      "entities": "장원섭의 얼굴, 머리스타일, 검사복이 레퍼런스와 일치하며, 책상 위에 텍스트가 적힌 다이어리가 놓여 있음.",
      "hard_violations": [],
      "physics": "인물은 데스크 뒤에 서 있고, 다이어리는 목재 데스크 표면 위에 안정적으로 평평하게 놓여 있어 물리적 지지가 명확함."
     },
     {
      "label": "A",
      "direction": "장원섭이 화면 왼쪽 밖 변호인석을 향해 시선을 고정하고 강하게 소리치고 있음.",
      "built_space": "법정 내 목재 데스크와 '검사' 명패가 피사체 앞의 올바른 위치에 있음.",
      "entities": "장원섭의 얼굴, 머리스타일, 검사복이 레퍼런스와 일치하며, 텍스트가 적힌 다이어리가 존재함.",
      "hard_violations": [
       "지지대 없이 부자연스럽게 서 있는 다이어리"
      ],
      "physics": "인물은 안정적으로 서 있으나, 다이어리가 손이나 다른 지지물 없이 V자 형태로 데스크 모서리에 비스듬히 서 있어 중력 및 물리 법칙에 위배됨."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S87sh1"
  }
 },
 "S87sh5::cine": {
  "applied": true,
  "fingerprint": "b214f4f0006668d0d33272e391a5f596e1e64db746126c4b887b8fd282f9af35",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S87sh5_sel.png",
  "source_sha256": "06c512aadcf6e16f14791b765d100dcb08b92c626c4ee8b002a9c9941349fbba",
  "file": "S87sh5_cine.png",
  "latency_ms": 17225
 },
 "S88sh1::signage": {
  "fp": "2f8950c07781bcfe",
  "inscriptions": [
   {
    "surface_native": "증인석 명패",
    "text_native": "증인석",
    "reason_ko": "한국 법정 증인석 데스크 전면에 부착되는 표지판으로, 인물이 선 자리의 법적 지위와 공간의 현실감을 명확하게 보여줍니다."
   }
  ]
 },
 "S88sh1": {
  "input_fingerprint": "2abb802347297cf0",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 증인석에 꼿꼿하게 선 채 앞을 주시하는 조일남의 전경\n\nLOCATION (lock): At the courtroom witness stand, facing the prosecutor and judges. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At standing chest height and medium-long distance on the counsel side of the axis, hold 조일남's full figure at the witness stand in an off-frontal three-quarter composition before the dolly begins. He remains prominent near center with his attention fixed diagonally on 장원섭, while enough surrounding courtroom space establishes the formal testimony setting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 증인석 (Occupied by 조일남 during testimony) — Its counsel-facing side and front edge are visible from the oblique camera position; used as Anchors 조일남's full figure and defines his formal position in court; 법정 내부 (Court proceedings are in progress); used as Provides restrained spatial depth around the witness position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination with subdued cool tones and a lightly documentary tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, large screen, and daylight. Exclude the previous prosecutor-defense confrontation and place the forensic expert upright at the stand.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 증인석에 꼿꼿하게 선 채 앞을 주시하는 조일남의 전경\n\nLOCATION (lock): At the courtroom witness stand, facing the prosecutor and judges. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At standing chest height and medium-long distance on the counsel side of the axis, hold 조일남's full figure at the witness stand in an off-frontal three-quarter composition before the dolly begins. He remains prominent near center with his attention fixed diagonally on 장원섭, while enough surrounding courtroom space establishes the formal testimony setting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 증인석 (Occupied by 조일남 during testimony) — Its counsel-facing side and front edge are visible from the oblique camera position; used as Anchors 조일남's full figure and defines his formal position in court; 법정 내부 (Court proceedings are in progress); used as Provides restrained spatial depth around the witness position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination with subdued cool tones and a lightly documentary tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, large screen, and daylight. Exclude the previous prosecutor-defense confrontation and place the forensic expert upright at the stand.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 증인석에 꼿꼿하게 선 채 앞을 주시하는 조일남의 전경\n\nLOCATION (lock): At the courtroom witness stand, facing the prosecutor and judges. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At standing chest height and medium-long distance on the counsel side of the axis, hold 조일남's full figure at the witness stand in an off-frontal three-quarter composition before the dolly begins. He remains prominent near center with his attention fixed diagonally on 장원섭, while enough surrounding courtroom space establishes the formal testimony setting.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: 증인석 (Occupied by 조일남 during testimony) — Its counsel-facing side and front edge are visible from the oblique camera position; used as Anchors 조일남's full figure and defines his formal position in court; 법정 내부 (Court proceedings are in progress); used as Provides restrained spatial depth around the witness position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination with subdued cool tones and a lightly documentary tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom front, witness stand, large screen, and daylight. Exclude the previous prosecutor-defense confrontation and place the forensic expert upright at the stand.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증인석 명패: \"증인석\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "인물의 시선이 화면 밖 대각선 앞쪽을 향하고 있음.",
    "built_space": "이전 샷의 전경(검사석 명패와 펼쳐진 책)을 정확히 유지한 채 중경에 증인석을 배치했고, 좌측 창문과 패널 등 법정 구조가 일치함.",
    "entities": "조일남의 외모와 복장(의사가운, 파란 셔츠, 신분증)이 레퍼런스와 일치하며, '증인석'과 '검사' 명패의 텍스트가 명확함.",
    "hard_violations": [],
    "physics": "두 손을 증인석 단상 모서리에 자연스럽게 얹고 체중을 지탱하며 서 있음."
   },
   {
    "label": "B",
    "direction": "인물의 시선이 화면 밖 대각선 앞쪽을 향하고 있음.",
    "built_space": "증인석이 중앙에 배치되었으나, 배경의 가로형 나무 패널과 우측 창문 등 공간 구조가 이전 샷과 완전히 다름.",
    "entities": "조일남의 외모와 복장이 레퍼런스와 일치하며, '증인석' 명패 텍스트가 정확함.",
    "hard_violations": [],
    "physics": "두 손을 증인석 단상 위에 얹고 안정적인 자세로 서 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "이전 샷의 전경(검사석과 책)과 좌측 창문 등 법정 공간의 연속성을 완벽하게 유지하면서, 지시된 앵글과 인물의 모습을 매우 충실하게 구현함."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "인물의 묘사와 명패 텍스트는 훌륭하나, 이전 샷에서 지정된 법정의 구조(벽면 패널, 창문 위치 등)를 무시하여 공간의 연속성이 훼손됨."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S87sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 조일남: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878922>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 인물의 전신(full figure)이 포함된 와이드 샷(wide shot)을 명확히 요구했으나, 생성된 이미지는 허리 부근에서 피사체가 잘린 미디엄 샷으로 연출되었습니다.",
     "fix_en": "Ensure the visible lower half of the character accurately interacts with the witness stand within the current crop. Preserve the person, his position, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "증인석 위에 올려진 인물의 오른손(화면 좌측) 손가락 디테일이 다소 뭉개지고 불분명하게 렌더링되었습니다.",
     "fix_en": "Redraw the right hand with anatomically clear fingers. Preserve the person, his position, clothing, set, lighting, and framing.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "조일남이 화면 중앙 근처가 아니라 오른쪽 편에 치우쳐 있다.",
     "fix_en": "Maintain the current framing and ensure the right-side placement blends naturally with the courtroom elements. Preserve the person, his position, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전경 왼쪽 펼친 책에 장면이 요구하지 않은 영어 날짜가 선명히 읽힌다.",
     "fix_en": "Obscure the text on the foreground book using shallow focus to make it illegible. Preserve the person, his position, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 4
    },
    {
     "issue_ko": "왼쪽 창호가 이전 샷에 고정된 작은 창과 다른 큰 격자창이다.",
     "fix_en": "Change the left window to match the smaller frosted window shown in the previous shot. Preserve the person, his position, clothing, set, lighting, and framing.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 인물의 전신(full figure)이 포함된 와이드 샷(wide shot)을 명확히 요구했으나, 생성된 이미지는 허리 부근에서 피사체가 잘린 미디엄 샷으로 연출되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "증인석 위에 올려진 인물의 오른손(화면 좌측) 손가락 디테일이 다소 뭉개지고 불분명하게 렌더링되었습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "와이드·전신이어야 할 프레이밍이 허리 위까지인 미디엄에 가깝고 조일남 하반신이 잘렸다.",
     "severity": "major"
    },
    {
     "issue_ko": "조일남이 화면 중앙 근처가 아니라 오른쪽 편에 치우쳐 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "전경 왼쪽 펼친 책에 장면이 요구하지 않은 영어 날짜가 선명히 읽힌다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 창호가 이전 샷에 고정된 작은 창과 다른 큰 격자창이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 인물의 전신(full figure)이 포함된 와이드 샷(wide shot)을 명확히 요구했으나, 생성된 이미지는 허리 부근에서 피사체가 잘린 미디엄 샷으로 연출되었습니다.",
    "fix_en": "Ensure the visible lower half of the character accurately interacts with the witness stand within the current crop. Preserve the person, his position, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 0,
    "needs_regeneration": true
   },
   {
    "issue_ko": "증인석 위에 올려진 인물의 오른손(화면 좌측) 손가락 디테일이 다소 뭉개지고 불분명하게 렌더링되었습니다.",
    "fix_en": "Redraw the right hand with anatomically clear fingers. Preserve the person, his position, clothing, set, lighting, and framing.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "조일남이 화면 중앙 근처가 아니라 오른쪽 편에 치우쳐 있다.",
    "fix_en": "Maintain the current framing and ensure the right-side placement blends naturally with the courtroom elements. Preserve the person, his position, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "전경 왼쪽 펼친 책에 장면이 요구하지 않은 영어 날짜가 선명히 읽힌다.",
    "fix_en": "Obscure the text on the foreground book using shallow focus to make it illegible. Preserve the person, his position, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 4
   },
   {
    "issue_ko": "왼쪽 창호가 이전 샷에 고정된 작은 창과 다른 큰 격자창이다.",
    "fix_en": "Change the left window to match the smaller frosted window shown in the previous shot. Preserve the person, his position, clothing, set, lighting, and framing.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S87sh5"
  }
 },
 "S88sh1::cine": {
  "applied": true,
  "fingerprint": "082df1fbd1153bdd51834cacfa15cfd5b9886fe04d21c1984cbda182484c8455",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S88sh1_sel.png",
  "source_sha256": "1eb48a92e64f7f8ea7a96ee27384ef227403cbcdf9d0523ad8878d4dfe41ecdd",
  "file": "S88sh1_cine.png",
  "latency_ms": 28823
 },
 "S88sh2::signage": {
  "fp": "5f71eb84633dc9c0",
  "inscriptions": [
   {
    "surface_native": "증거물 표찰",
    "text_native": "증 제1호증",
    "reason_ko": "법정에서 증거물로 제시된 스노우볼에 부착된 한국 법원의 공식 증거 표찰을 재현하기 위함."
   }
  ]
 },
 "S88sh2": {
  "input_fingerprint": "ba5d0484a7d6369d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성 모형이 든 스노우볼을 허공을 향해 치켜든 조일남의 손 클로즈업\n\nLOCATION (lock): At the witness stand inside the courtroom, above the rail where the demonstration object is displayed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly below hand height beside 조일남's torso, end the upward tilt on an extreme close detail of his fully raised hand and the snow globe at its highest point. The globe and hand share the central area without the object exceeding two-fifths of frame, while a sliver of his sleeve and the witness-stand context remain as scale references; his face stays outside the image.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: raised snow globe in the upper-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 성 모형이 든 스노우볼 (Held aloft in 조일남's hand) — The globe is seen slightly from below, with the castle model and internal particles visible through its surface; used as Demonstration detail held at the endpoint of the upward tilt; 증인석 가장자리 (Occupied during testimony) — Only its upper edge is visible below the demonstration; used as Provides a narrow lower-frame spatial reference beneath the raised hand.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination renders the demonstration clearly without heightened or sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조일남 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same witness-stand background, courtroom lighting, expert's sleeve, and the glass snow globe. Exclude the expert's full body and focus on his raised hand holding the object.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Jo Ilnam holds up the castle snow globe as the visual model for explaining how suspended particles settle without motion.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 조일남 right now — 조일남's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 조일남 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증거물 표찰: \"증 제1호증\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성 모형이 든 스노우볼을 허공을 향해 치켜든 조일남의 손 클로즈업\n\nLOCATION (lock): At the witness stand inside the courtroom, above the rail where the demonstration object is displayed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly below hand height beside 조일남's torso, end the upward tilt on an extreme close detail of his fully raised hand and the snow globe at its highest point. The globe and hand share the central area without the object exceeding two-fifths of frame, while a sliver of his sleeve and the witness-stand context remain as scale references; his face stays outside the image.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: raised snow globe in the upper-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 성 모형이 든 스노우볼 (Held aloft in 조일남's hand) — The globe is seen slightly from below, with the castle model and internal particles visible through its surface; used as Demonstration detail held at the endpoint of the upward tilt; 증인석 가장자리 (Occupied during testimony) — Only its upper edge is visible below the demonstration; used as Provides a narrow lower-frame spatial reference beneath the raised hand.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination renders the demonstration clearly without heightened or sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조일남 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same witness-stand background, courtroom lighting, expert's sleeve, and the glass snow globe. Exclude the expert's full body and focus on his raised hand holding the object.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Jo Ilnam holds up the castle snow globe as the visual model for explaining how suspended particles settle without motion.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 조일남 right now — 조일남's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 조일남 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증거물 표찰: \"증 제1호증\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 성 모형이 든 스노우볼을 허공을 향해 치켜든 조일남의 손 클로즈업\n\nLOCATION (lock): At the witness stand inside the courtroom, above the rail where the demonstration object is displayed. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly below hand height beside 조일남's torso, end the upward tilt on an extreme close detail of his fully raised hand and the snow globe at its highest point. The globe and hand share the central area without the object exceeding two-fifths of frame, while a sliver of his sleeve and the witness-stand context remain as scale references; his face stays outside the image.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: raised snow globe in the upper-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: 성 모형이 든 스노우볼 (Held aloft in 조일남's hand) — The globe is seen slightly from below, with the castle model and internal particles visible through its surface; used as Demonstration detail held at the endpoint of the upward tilt; 증인석 가장자리 (Occupied during testimony) — Only its upper edge is visible below the demonstration; used as Provides a narrow lower-frame spatial reference beneath the raised hand.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination renders the demonstration clearly without heightened or sensational emphasis.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 조일남 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same witness-stand background, courtroom lighting, expert's sleeve, and the glass snow globe. Exclude the expert's full body and focus on his raised hand holding the object.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Jo Ilnam holds up the castle snow globe as the visual model for explaining how suspended particles settle without motion.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 조일남 right now — 조일남's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 조일남 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 조일남 (Korean 남성, 50대 중반, 세월감 있는 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 증거물 표찰: \"증 제1호증\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.286,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.714,
    "B": 1.571
   },
   "adjusted": {
    "A": 1.714,
    "B": 1.321
   },
   "violations": {
    "B": [
     "[gemini-pro] 지정된 텍스트('증 제1호증')가 스노우볼 금속 장식 위와 단상에 중복으로 생성 및 유출됨(leaked text/duplicated text)."
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1714,
   "B": 1321
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1714,
    "verdict_ko": "요구된 클로즈업 앵글과 손의 피사체 비율을 완벽히 구현했으며, 소매 의상과 소품 디테일이 참조와 정확히 일치함."
   },
   {
    "label": "B",
    "score": 1321,
    "verdict_ko": "앵글과 소품의 형태는 양호하나, 지정된 표찰 텍스트를 소품 받침대와 배경 두 곳에 중복으로 렌더링하는 치명적 오류를 범함.  ★위반: [gemini-pro] 지정된 텍스트('증 제1호증')가 스노우볼 금속 장식 위와 단상에 중복으로 생성 및 유출됨(leaked text/duplicated text)."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 조일남 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S88sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 조일남: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:878922>"
   },
   {
    "label": "PROP REFERENCE — 성 모형이 든 스노우볼: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1309901>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "스노우볼을 쥐고 있는 오른손의 우측면에 5개의 손가락이 묘사되어 있어, 보이지 않는 엄지까지 포함하면 총 6개의 손가락이 있는 해부학적 오류가 발생했습니다.",
     "fix_en": "Redraw the hand holding the snow globe to correctly show an anatomically accurate left hand, featuring exactly four fingers curling around the right side of the base and one thumb supporting the left side. Preserve the snow globe, the white and blue sleeve, the witness stand, the background, the lighting, and the overall framing exactly as they are.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "프롬프트는 손 높이 약간 아래에서 완전히 치켜든 손을 올려다보는 '상향 틸트(upward tilt)' 앵글과 '익스트림 클로즈업'을 요구했으나, 사진은 눈높이에서 수평으로 뻗은 팔을 넓게 잡은 정면 앵글로 연출되었습니다.",
     "fix_en": "Crop the image tightly around the hand and the snow globe to simulate an extreme close-up. Preserve the snow globe, the hand, the sleeve, the witness stand, the background, the lighting, and the existing geometry exactly as they are.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "증인석은 하단 가장자리만 좁게 보여야 하는데 화면 오른쪽 하단을 크게 차지하며 내부와 정면까지 드러난다.",
     "fix_en": "Darken and blur the lower right area of the witness stand to reduce its prominence in the frame. Preserve the hand, the snow globe, the sleeve, the background, the lighting, and the framing exactly as they are.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "스노우볼을 쥐고 있는 오른손의 우측면에 5개의 손가락이 묘사되어 있어, 보이지 않는 엄지까지 포함하면 총 6개의 손가락이 있는 해부학적 오류가 발생했습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "프롬프트는 손 높이 약간 아래에서 완전히 치켜든 손을 올려다보는 '상향 틸트(upward tilt)' 앵글과 '익스트림 클로즈업'을 요구했으나, 사진은 눈높이에서 수평으로 뻗은 팔을 넓게 잡은 정면 앵글로 연출되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "카메라가 조일남 몸통 옆 아래에서 치켜든 손을 올려다보지 않고 법정 쪽에서 증인석 정면을 바라본다.",
     "severity": "major"
    },
    {
     "issue_ko": "증인석은 하단 가장자리만 좁게 보여야 하는데 화면 오른쪽 하단을 크게 차지하며 내부와 정면까지 드러난다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 손 높이 약간 아래에서 완전히 치켜든 손을 올려다보는 '상향 틸트(upward tilt)' 앵글과 '익스트림 클로즈업'을 요구했으나, 사진은 눈높이에서 수평으로 뻗은 팔을 넓게 잡은 정면 앵글로 연출되었습니다.",
    "fix_en": "Crop the image tightly around the hand and the snow globe to simulate an extreme close-up. Preserve the snow globe, the hand, the sleeve, the witness stand, the background, the lighting, and the existing geometry exactly as they are.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "증인석은 하단 가장자리만 좁게 보여야 하는데 화면 오른쪽 하단을 크게 차지하며 내부와 정면까지 드러난다.",
    "fix_en": "Darken and blur the lower right area of the witness stand to reduce its prominence in the frame. Preserve the hand, the snow globe, the sleeve, the background, the lighting, and the framing exactly as they are.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 4,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the hand holding the snow globe to correctly show an anatomically accurate left hand, featuring exactly four fingers curling around the right side of the base and one thumb supporting the left side. Preserve the snow globe, the white and blue sleeve, the witness stand, the background, the lighting, and the overall framing exactly as they are.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "인물의 전신을 배제하고 스노우볼을 치켜든 손만을 클로즈업하라는 카메라 프레이밍 지시를 완벽하게 구현했으며, 요구된 표찰의 텍스트와 소품 디테일도 훌륭하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "손 클로즈업 지시를 무시하고 인물의 상반신과 얼굴을 그대로 노출하는 심각한 프레이밍 오류를 범했으며, 스노우볼이 허공에 떠 있는 치명적인 물리적 결함이 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "손이 화면 중앙에서 스노우볼을 위로 치켜들고 있으며, 스노우볼 안의 성 모형이 정면을 향함.",
      "built_space": "법정 내 증인석 위. 배경에 창문과 목재 벽면, 그리고 증인석 단상이 배치되어 있음.",
      "entities": "조일남의 손과 하얀 의사 가운, 파란 셔츠 소매. 레퍼런스와 일치하는 성 모형 스노우볼. '증 제1호증' 텍스트가 정확히 적힌 표찰.",
      "hard_violations": [],
      "physics": "손바닥과 손가락이 스노우볼의 금속 받침대를 아래에서 단단히 받치며 안정적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "인물이 정면 밖을 응시하며 오른손을 뻗어 스노우볼을 가리키거나 들고 있음.",
      "built_space": "법정 내 증인석 주변. 법대, 증인석, 검사석 표찰 등 법정 구조물들이 보임.",
      "entities": "조일남의 상반신 전체(얼굴, 머리, 의상 모두 레퍼런스 일치). 성 모형 스노우볼.",
      "hard_violations": [
       "스노우볼을 쥐고 있어야 할 손가락이 물체와 제대로 닿지 않아 스노우볼이 손 앞 허공에 물리적으로 불가능하게 떠 있음"
      ],
      "physics": "스노우볼이 손바닥이나 손가락에 제대로 쥐어지지 않은 채 오른손 앞 허공에 지지대 없이 어색하게 떠 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "인물의 전신을 배제하고 스노우볼을 치켜든 손만을 클로즈업하라는 카메라 프레이밍 지시를 완벽하게 구현했으며, 요구된 표찰의 텍스트와 소품 디테일도 훌륭하게 재현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "손 클로즈업 지시를 무시하고 인물의 상반신과 얼굴을 그대로 노출하는 심각한 프레이밍 오류를 범했으며, 스노우볼이 허공에 떠 있는 치명적인 물리적 결함이 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "손이 화면 중앙에서 스노우볼을 위로 치켜들고 있으며, 스노우볼 안의 성 모형이 정면을 향함.",
      "built_space": "법정 내 증인석 위. 배경에 창문과 목재 벽면, 그리고 증인석 단상이 배치되어 있음.",
      "entities": "조일남의 손과 하얀 의사 가운, 파란 셔츠 소매. 레퍼런스와 일치하는 성 모형 스노우볼. '증 제1호증' 텍스트가 정확히 적힌 표찰.",
      "hard_violations": [],
      "physics": "손바닥과 손가락이 스노우볼의 금속 받침대를 아래에서 단단히 받치며 안정적으로 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "인물이 정면 밖을 응시하며 오른손을 뻗어 스노우볼을 가리키거나 들고 있음.",
      "built_space": "법정 내 증인석 주변. 법대, 증인석, 검사석 표찰 등 법정 구조물들이 보임.",
      "entities": "조일남의 상반신 전체(얼굴, 머리, 의상 모두 레퍼런스 일치). 성 모형 스노우볼.",
      "hard_violations": [
       "스노우볼을 쥐고 있어야 할 손가락이 물체와 제대로 닿지 않아 스노우볼이 손 앞 허공에 물리적으로 불가능하게 떠 있음"
      ],
      "physics": "스노우볼이 손바닥이나 손가락에 제대로 쥐어지지 않은 채 오른손 앞 허공에 지지대 없이 어색하게 떠 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 10,
      "verdict_ko": "지시된 인서트 클로즈업 앵글을 완벽하게 구현하여 인물의 얼굴을 배제하고 손과 스노우볼에 집중했으며, 의상의 소매 디테일과 '증 제1호증' 표찰까지 정확하게 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "인서트 클로즈업 및 얼굴 배제라는 프레이밍 지시를 완전히 무시하였으며, 스노우볼이 손에 쥐어지지 않고 허공에 떠 있는 치명적인 물리적 오류가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "스노우볼을 들고 있는 손이 화면 상단을 향함",
      "built_space": "법정 증인석의 나무 패널과 배경",
      "entities": "조일남의 손과 소매(흰 가운, 파란 셔츠), 성 모형 스노우볼, '증 제1호증'이 적힌 표찰",
      "hard_violations": [],
      "physics": "손이 스노우볼의 하단 받침대를 물리적으로 자연스럽게 지탱하며 쥐고 있음"
     },
     {
      "label": "A",
      "direction": "인물의 시선은 정면을 향함",
      "built_space": "법정 내부, 증인석과 검사석",
      "entities": "조일남(얼굴 및 상반신 포함), 성 모형 스노우볼, 재판 기록물 표지",
      "hard_violations": [
       "스노우볼이 손에 닿지 않은 채 허공에 떠 있음 (물리적 불가능)",
       "얼굴을 화면에서 배제하라는 프레이밍 지시 위반"
      ],
      "physics": "스노우볼을 지탱하는 것이 아무것도 없으며, 인물의 오른손 옆에 마법처럼 떠 있음"
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 10,
      "verdict_ko": "지시된 인서트 클로즈업 앵글을 완벽하게 구현하여 인물의 얼굴을 배제하고 손과 스노우볼에 집중했으며, 의상의 소매 디테일과 '증 제1호증' 표찰까지 정확하게 반영한 훌륭한 결과물입니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "인서트 클로즈업 및 얼굴 배제라는 프레이밍 지시를 완전히 무시하였으며, 스노우볼이 손에 쥐어지지 않고 허공에 떠 있는 치명적인 물리적 오류가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "스노우볼을 들고 있는 손이 화면 상단을 향함",
      "built_space": "법정 증인석의 나무 패널과 배경",
      "entities": "조일남의 손과 소매(흰 가운, 파란 셔츠), 성 모형 스노우볼, '증 제1호증'이 적힌 표찰",
      "hard_violations": [],
      "physics": "손이 스노우볼의 하단 받침대를 물리적으로 자연스럽게 지탱하며 쥐고 있음"
     },
     {
      "label": "B",
      "direction": "인물의 시선은 정면을 향함",
      "built_space": "법정 내부, 증인석과 검사석",
      "entities": "조일남(얼굴 및 상반신 포함), 성 모형 스노우볼, 재판 기록물 표지",
      "hard_violations": [
       "스노우볼이 손에 닿지 않은 채 허공에 떠 있음 (물리적 불가능)",
       "얼굴을 화면에서 배제하라는 프레이밍 지시 위반"
      ],
      "physics": "스노우볼을 지탱하는 것이 아무것도 없으며, 인물의 오른손 옆에 마법처럼 떠 있음"
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 20,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S88sh1"
  }
 },
 "S88sh6::signage": {
  "fp": "f5f010ca22fd30ef",
  "inscriptions": [
   {
    "surface_native": "피고인석 명패",
    "text_native": "피고인",
    "reason_ko": "지국현이 법정의 피고인석에 앉아 있음을 나타내기 위해 테이블 위의 명패에 한글 표기가 필요합니다."
   }
  ]
 },
 "S88sh6": {
  "input_fingerprint": "45630d405bceff4a",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 조일남 쪽을 향해 상기된 얼굴로 미간을 잔뜩 구긴 채 살벌한 눈빛을 쏘아보는 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the expert witness. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 지국현's seated eye height and close range, cut to an off-axis three-quarter facial close-up with his face weighted to one side and empty space extending along his diagonal glare toward the off-screen 조일남. His rigid shoulders remain at the lower edge while the tightened brow and fixed eyes become the sole emotional change carried by the cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 법정 내부 (Testimony is in progress); used as Minimal soft context behind 지국현 so his directed glare remains dominant.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination holds the flushed facial tension within restrained cool tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom layout, witness stand, formal light, and defendant's clothing. Exclude the expert and snow globe from the close frame, isolating the defendant's furious stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피고인석 명패: \"피고인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 조일남 쪽을 향해 상기된 얼굴로 미간을 잔뜩 구긴 채 살벌한 눈빛을 쏘아보는 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the expert witness. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 지국현's seated eye height and close range, cut to an off-axis three-quarter facial close-up with his face weighted to one side and empty space extending along his diagonal glare toward the off-screen 조일남. His rigid shoulders remain at the lower edge while the tightened brow and fixed eyes become the sole emotional change carried by the cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 법정 내부 (Testimony is in progress); used as Minimal soft context behind 지국현 so his directed glare remains dominant.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination holds the flushed facial tension within restrained cool tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom layout, witness stand, formal light, and defendant's clothing. Exclude the expert and snow globe from the close frame, isolating the defendant's furious stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피고인석 명패: \"피고인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 조일남 쪽을 향해 상기된 얼굴로 미간을 잔뜩 구긴 채 살벌한 눈빛을 쏘아보는 지국현의 얼굴 클로즈업\n\nLOCATION (lock): At the defendant's table inside the courtroom, facing the expert witness. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At 지국현's seated eye height and close range, cut to an off-axis three-quarter facial close-up with his face weighted to one side and empty space extending along his diagonal glare toward the off-screen 조일남. His rigid shoulders remain at the lower edge while the tightened brow and fixed eyes become the sole emotional change carried by the cut.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: 법정 내부 (Testimony is in progress); used as Minimal soft context behind 지국현 so his directed glare remains dominant.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination holds the flushed facial tension within restrained cool tones and moderate contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom layout, witness stand, formal light, and defendant's clothing. Exclude the expert and snow globe from the close frame, isolating the defendant's furious stare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 지국현 (현재) (Korean 남성, 30대 후반, 갸름한 얼굴, 날카로운 눈매, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피고인석 명패: \"피고인\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.714
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.714
   },
   "violations": {},
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1714
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "프레이밍과 피사체의 시선 방향이 지시사항을 잘 따랐으며, 명패 및 수의의 텍스트('교도소')를 정확하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 1714,
    "verdict_ko": "구도와 표정은 적절하나 수의의 텍스트에 오타('교도스')가 발생했으며 홍조 표현이 다소 부자연스럽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S88sh2_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 지국현 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:994284>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "양 볼의 붉은 기운이 자연스러운 생리적 홍조가 아닌, 인위적으로 칠해진 분장이나 그래픽처럼 표현되어 '절제된 쿨톤' 및 사실성 지시를 위반함.",
     "fix_en": "Blend the artificial red circles on the character's cheeks into a natural, subtle physiological skin flush appropriate for cool lighting. Preserve the character's identity, facial expression, clothing, the courtroom setting, lighting, and framing.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "얼굴 클로즈업인데 상반신과 전경 명패가 크게 들어와 어깨가 하단 가장자리에만 머물지 않는다.",
     "fix_en": "Crop the image tightly into a facial close-up so the character's shoulders rest directly at the bottom edge, minimizing the visible torso and foreground desk. Preserve the character's face, expression, clothing, lighting, and the background courtroom elements.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "시선이 화면 밖 조일남 쪽이 아니라 왼쪽 아래 명패·책상 쪽으로 내려가 있다.",
     "fix_en": "Shift the character's pupils and eyelids to direct his glare diagonally outward and off-screen, rather than looking down toward the desk. Preserve the character's head position, facial expression, clothing, lighting, courtroom background, and framing.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "양 볼의 붉은 기운이 자연스러운 생리적 홍조가 아닌, 인위적으로 칠해진 분장이나 그래픽처럼 표현되어 '절제된 쿨톤' 및 사실성 지시를 위반함.",
     "severity": "major"
    },
    {
     "issue_ko": "얼굴 클로즈업인데 상반신과 전경 명패가 크게 들어와 어깨가 하단 가장자리에만 머물지 않는다.",
     "severity": "major"
    },
    {
     "issue_ko": "시선이 화면 밖 조일남 쪽이 아니라 왼쪽 아래 명패·책상 쪽으로 내려가 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "양볼 홍조가 절제된 쿨톤의 상기가 아니라 과장되게 붉게 칠해져 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "양 볼의 붉은 기운이 자연스러운 생리적 홍조가 아닌, 인위적으로 칠해진 분장이나 그래픽처럼 표현되어 '절제된 쿨톤' 및 사실성 지시를 위반함.",
    "fix_en": "Blend the artificial red circles on the character's cheeks into a natural, subtle physiological skin flush appropriate for cool lighting. Preserve the character's identity, facial expression, clothing, the courtroom setting, lighting, and framing.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "얼굴 클로즈업인데 상반신과 전경 명패가 크게 들어와 어깨가 하단 가장자리에만 머물지 않는다.",
    "fix_en": "Crop the image tightly into a facial close-up so the character's shoulders rest directly at the bottom edge, minimizing the visible torso and foreground desk. Preserve the character's face, expression, clothing, lighting, and the background courtroom elements.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "시선이 화면 밖 조일남 쪽이 아니라 왼쪽 아래 명패·책상 쪽으로 내려가 있다.",
    "fix_en": "Shift the character's pupils and eyelids to direct his glare diagonally outward and off-screen, rather than looking down toward the desk. Preserve the character's head position, facial expression, clothing, lighting, courtroom background, and framing.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S88sh2"
  }
 },
 "S88sh6::cine": {
  "applied": true,
  "fingerprint": "d32017689ca47cef521aaea9b485b4f6f59e0f7ce062ab043dbd8e6556db162d",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S88sh6_sel.png",
  "source_sha256": "3af29cb98382a3c5e52bfbd4666b77f8be61522f891aa8851a0aa5828689e7c4",
  "file": "S88sh6_cine.png",
  "latency_ms": 13308
 },
 "S89sh16::signage": {
  "fp": "7fbcc5caaf70d5dc",
  "inscriptions": []
 },
 "S89sh16": {
  "input_fingerprint": "4821623c4260b49e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 현재 방청석, 고개를 푹 숙인 채 두 손으로 얼굴을 감싸 쥐고 뺨 위로 눈물이 흐른 신경희와 유경자의 상체\n\nLOCATION (lock): In the courtroom public gallery during the verdict. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the side-aisle track slightly above seated eye level, tilting down into a tight two-shot of 신경희 and 유경자 folded inward across adjacent areas of the frame. 신경희 curls over her covered face at left while 유경자 occupies the right with her hands pressed higher against her cheeks; their differing shoulder and elbow angles prevent the shared grief from reading as mirrored posing.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 신경희 in the middle-left of the frame, midground; 유경자 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 방청석 (Occupied during the verdict proceedings) — The seating row is viewed obliquely from the side aisle and slightly above; used as Places both grieving women within the present courtroom audience rather than an isolated emotional void.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination and subdued cool tonality keep the grief intimate and non-sensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom architecture, gallery seating, formal daylight, and the supporters' clothing. Exclude the expert testimony setup and focus on the two women crying in the current gallery.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리); 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 현재 방청석, 고개를 푹 숙인 채 두 손으로 얼굴을 감싸 쥐고 뺨 위로 눈물이 흐른 신경희와 유경자의 상체\n\nLOCATION (lock): In the courtroom public gallery during the verdict. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the side-aisle track slightly above seated eye level, tilting down into a tight two-shot of 신경희 and 유경자 folded inward across adjacent areas of the frame. 신경희 curls over her covered face at left while 유경자 occupies the right with her hands pressed higher against her cheeks; their differing shoulder and elbow angles prevent the shared grief from reading as mirrored posing.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 신경희 in the middle-left of the frame, midground; 유경자 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 방청석 (Occupied during the verdict proceedings) — The seating row is viewed obliquely from the side aisle and slightly above; used as Places both grieving women within the present courtroom audience rather than an isolated emotional void.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination and subdued cool tonality keep the grief intimate and non-sensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom architecture, gallery seating, formal daylight, and the supporters' clothing. Exclude the expert testimony setup and focus on the two women crying in the current gallery.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리); 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 현재 방청석, 고개를 푹 숙인 채 두 손으로 얼굴을 감싸 쥐고 뺨 위로 눈물이 흐른 신경희와 유경자의 상체\n\nLOCATION (lock): In the courtroom public gallery during the verdict. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the side-aisle track slightly above seated eye level, tilting down into a tight two-shot of 신경희 and 유경자 folded inward across adjacent areas of the frame. 신경희 curls over her covered face at left while 유경자 occupies the right with her hands pressed higher against her cheeks; their differing shoulder and elbow angles prevent the shared grief from reading as mirrored posing.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 신경희 in the middle-left of the frame, midground; 유경자 in the middle-right of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 방청석 (Occupied during the verdict proceedings) — The seating row is viewed obliquely from the side aisle and slightly above; used as Places both grieving women within the present courtroom audience rather than an isolated emotional void.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination and subdued cool tonality keep the grief intimate and non-sensational.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom architecture, gallery seating, formal daylight, and the supporters' clothing. Exclude the expert testimony setup and focus on the two women crying in the current gallery.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 신경희 (Korean 여성, 성인, 타원형 얼굴, 검은색 중간 길이 머리); 유경자 (Korean 여성, 60대 후반, 주름진 둥근 얼굴, 검은색·회색 짧은 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "gq": {
   "route": "combined",
   "gap": 0.25,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 1.75,
    "B": 1.5
   },
   "adjusted": {
    "A": 1.75,
    "B": 1.0
   },
   "violations": {
    "B": [
     "[gemini-pro] 유경자의 왼쪽 손가락 개수가 정상보다 많게 묘사된 해부학적 오류",
     "[gemini-pro] 신경희의 손 아래 허공에 지지대 없이 정지된 상태로 떠 있는 다수의 눈물 방울들"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "agreed": false
  },
  "totals": {
   "A": 1750,
   "B": 1000
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 1750,
    "verdict_ko": "눈물이 점선 스티커처럼 평면적으로 묘사된 재질감의 한계가 있으나, 해부학적 오류 없이 지정된 구도와 프레임 배치를 안정적으로 구현했습니다."
   },
   {
    "label": "B",
    "score": 1000,
    "verdict_ko": "공중에 부자연스럽게 떠 있는 눈물 방울과 우측 여성의 다지증 등 치명적인 물리적 및 해부학적 오류가 발생했습니다.  ★위반: [gemini-pro] 유경자의 왼쪽 손가락 개수가 정상보다 많게 묘사된 해부학적 오류 / [gemini-pro] 신경희의 손 아래 허공에 지지대 없이 정지된 상태로 떠 있는 다수의 눈물 방울들"
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S88sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 신경희: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:843594>"
   },
   {
    "label": "CHARACTER REFERENCE — 유경자: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:1007208>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "두 여성의 뺨에 흐르는 눈물이 투명한 액체가 아니라 표면 위에 하얀색 점선이 그려진 그래픽 오버레이처럼 렌더링되어 재질 사실성 지침을 위반함.",
     "fix_en": "Redraw the tears on the cheeks as realistic, glossy, transparent liquid pooling naturally, removing the white dotted lines. Preserve the women, their exact poses, clothing, expressions, and the courtroom setting.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "인물들이 앉아 있는 방청석 벤치가 얇고 평평한 나무 판자 형태로 렌더링되어, 이전 장면 스틸컷(레퍼런스)에 설정된 두껍고 입체적인 법정 가구 디자인과 일치하지 않음.",
     "fix_en": "Thicken the wooden benches and add dimensional paneling to match the heavy courtroom furniture seen in the previous shot. Preserve the characters, their seating positions, lighting, and framing.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "두 여성의 뺨에 흐르는 눈물이 투명한 액체가 아니라 표면 위에 하얀색 점선이 그려진 그래픽 오버레이처럼 렌더링되어 재질 사실성 지침을 위반함.",
     "severity": "major"
    },
    {
     "issue_ko": "인물들이 앉아 있는 방청석 벤치가 얇고 평평한 나무 판자 형태로 렌더링되어, 이전 장면 스틸컷(레퍼런스)에 설정된 두껍고 입체적인 법정 가구 디자인과 일치하지 않음.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "두 여성의 뺨에 흐르는 눈물이 투명한 액체가 아니라 표면 위에 하얀색 점선이 그려진 그래픽 오버레이처럼 렌더링되어 재질 사실성 지침을 위반함.",
    "fix_en": "Redraw the tears on the cheeks as realistic, glossy, transparent liquid pooling naturally, removing the white dotted lines. Preserve the women, their exact poses, clothing, expressions, and the courtroom setting.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "인물들이 앉아 있는 방청석 벤치가 얇고 평평한 나무 판자 형태로 렌더링되어, 이전 장면 스틸컷(레퍼런스)에 설정된 두껍고 입체적인 법정 가구 디자인과 일치하지 않음.",
    "fix_en": "Thicken the wooden benches and add dimensional paneling to match the heavy courtroom furniture seen in the previous shot. Preserve the characters, their seating positions, lighting, and framing.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S88sh6"
  }
 },
 "S89sh16::cine": {
  "applied": true,
  "fingerprint": "7cac4349b245ee2564b8b5b4a0462a05eba6fa26f92c5c91db8f4b2ff7e37758",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S89sh16_sel.png",
  "source_sha256": "b29f7802eeac4256e2cabed9212bf82b68250168be31ab04593dd104271b4aa9",
  "file": "S89sh16_cine.png",
  "latency_ms": 11366
 },
 "S89sh18::signage": {
  "fp": "4c57a00204d14758",
  "inscriptions": [
   {
    "surface_native": "법관석 명패",
    "text_native": "재판장",
    "reason_ko": "대한민국 법정 심리에서 중앙의 부장판사석 앞에 놓이는 공식 직책 명패를 묘사하여 현장감을 더합니다."
   }
  ]
 },
 "S89sh18": {
  "input_fingerprint": "c2d6d9d52cb240be",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 마이크를 향해 흔들림 없는 시선을 던지며 단호하게 입을 벌린 부장판사의 상체\n\nLOCATION (lock): At the raised judges' bench inside the courtroom, facing the silent gallery and parties. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly below 부장판사's eye line and offset to one side of the microphone, complete the dolly in at a firm medium-close three-quarter composition across the bench. The judge occupies the central field with the microphone retained at one edge, speaking steadily past it toward the courtroom rather than into the camera.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마이크 (Positioned for the verdict delivery) — Seen in partial side view at the edge of the three-quarter composition, directed toward 부장판사; used as Edge anchor that clarifies the formal delivery of the sentence without obstructing the judge's face; 법대 (Occupied during sentencing) — Its front edge crosses the lower frame at an oblique angle; used as Provides the elevated institutional plane beneath the judge.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination supports controlled authority with moderate contrast and subdued neutral-to-cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom woodwork, judge's bench, formal daylight, and judicial robe. Exclude the witness and experimental props, isolating the presiding judge delivering the verdict.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 부장판사 (Korean 남성, 중년, 엄숙한 얼굴, 짧게 정돈된 검은색 머리) — wearing: 검사와 판사가 착용하는 흑색 법복. 앞면에 주름이 잡혀 있고 장식이 절제된 묵직한 가운 형태이다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법관석 명패: \"재판장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 마이크를 향해 흔들림 없는 시선을 던지며 단호하게 입을 벌린 부장판사의 상체\n\nLOCATION (lock): At the raised judges' bench inside the courtroom, facing the silent gallery and parties. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly below 부장판사's eye line and offset to one side of the microphone, complete the dolly in at a firm medium-close three-quarter composition across the bench. The judge occupies the central field with the microphone retained at one edge, speaking steadily past it toward the courtroom rather than into the camera.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마이크 (Positioned for the verdict delivery) — Seen in partial side view at the edge of the three-quarter composition, directed toward 부장판사; used as Edge anchor that clarifies the formal delivery of the sentence without obstructing the judge's face; 법대 (Occupied during sentencing) — Its front edge crosses the lower frame at an oblique angle; used as Provides the elevated institutional plane beneath the judge.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination supports controlled authority with moderate contrast and subdued neutral-to-cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom woodwork, judge's bench, formal daylight, and judicial robe. Exclude the witness and experimental props, isolating the presiding judge delivering the verdict.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 부장판사 (Korean 남성, 중년, 엄숙한 얼굴, 짧게 정돈된 검은색 머리) — wearing: 검사와 판사가 착용하는 흑색 법복. 앞면에 주름이 잡혀 있고 장식이 절제된 묵직한 가운 형태이다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법관석 명패: \"재판장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 마이크를 향해 흔들림 없는 시선을 던지며 단호하게 입을 벌린 부장판사의 상체\n\nLOCATION (lock): At the raised judges' bench inside the courtroom, facing the silent gallery and parties. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From slightly below 부장판사's eye line and offset to one side of the microphone, complete the dolly in at a firm medium-close three-quarter composition across the bench. The judge occupies the central field with the microphone retained at one edge, speaking steadily past it toward the courtroom rather than into the camera.\n- FRAMING SCALE: medium shot\n- KEY BACKGROUND ELEMENTS: 마이크 (Positioned for the verdict delivery) — Seen in partial side view at the edge of the three-quarter composition, directed toward 부장판사; used as Edge anchor that clarifies the formal delivery of the sentence without obstructing the judge's face; 법대 (Occupied during sentencing) — Its front edge crosses the lower frame at an oblique angle; used as Provides the elevated institutional plane beneath the judge.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination supports controlled authority with moderate contrast and subdued neutral-to-cool color.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom woodwork, judge's bench, formal daylight, and judicial robe. Exclude the witness and experimental props, isolating the presiding judge delivering the verdict.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 부장판사 (Korean 남성, 중년, 엄숙한 얼굴, 짧게 정돈된 검은색 머리) — wearing: 검사와 판사가 착용하는 흑색 법복. 앞면에 주름이 잡혀 있고 장식이 절제된 묵직한 가운 형태이다. — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 법관석 명패: \"재판장\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "판사의 시선은 마이크를 지나 앞쪽 허공(법정)을 향해 흔들림 없이 고정되어 있으며, 입을 단호하게 벌리고 있음.",
    "built_space": "판사가 착석한 법대가 화면 하단에 위치하고, 뒤쪽으로는 레퍼런스와 일치하는 나무 패널 벽과 창문이 올바른 원근감으로 배치되어 있음. (단, 법대 앞쪽 모서리가 사선이 아닌 수평에 가깝게 연출됨)",
    "entities": "부장판사(중년 남성, 짧은 검은 머리, 엄숙한 표정)와 흑색 법복이 잘 묘사됨. 마이크와 명패(재판장)가 올바르게 존재하며 텍스트도 정확함.",
    "hard_violations": [],
    "physics": "인물의 자세는 보이지 않는 의자에 의해 지탱되며 자연스러움. 마이크는 법대 위에 안정적으로 고정되어 있음."
   },
   {
    "label": "B",
    "direction": "판사의 시선이 마이크 너머 앞쪽을 향하고 있으며, 입을 벌려 판결을 내리는 모습이 확인됨.",
    "built_space": "법대 뒤쪽 배경으로 방청객들이 앉는 긴 나무 의자(pews)가 겹쳐서 나타나며, 이는 높은 단 위에 있어야 할 재판장석의 공간 구조상 불가능한 배치임.",
    "entities": "부장판사의 외모와 법복이 지시사항에 부합함. 명패의 '재판장' 텍스트도 정확히 렌더링됨.",
    "hard_violations": [
     "법대 뒤로 방청석 의자가 늘어서 있어 법정의 건축적 구조와 무대 구성이 물리적으로 불가능함(Built Space violation)."
    ],
    "physics": "마이크 스탠드와 케이블의 형태가 법대 위에서 다소 어색하게 떠 있거나 단절되어 보임. 인물의 착석 상태는 안정적임."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 9,
   "B": 4
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 9,
    "verdict_ko": "명시된 인상착의와 의상, 명패의 텍스트를 정확히 구현했으며, 법정의 구조와 시선 처리도 자연스럽고 사실적으로 연출되었습니다."
   },
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "법대 뒤쪽 공간에 방청석 의자(pews)가 배치되어 법정의 기본 구조와 어긋나는 치명적인 공간 오류가 발생했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S89sh16_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 가장자리에 배치되어야 할 마이크가 프레임 가장자리가 아닌 중앙 부근(명패 뒤쪽)에 위치해 있습니다.",
     "fix_en": "Reposition the microphone so its stem and head sit at the far left edge of the frame, rather than near the center. Preserve the judge's face, posture, robe, the nameplate, the bench, and the background.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "법대 앞모서리가 사선으로 가로지르지 않고 화면 하단과 거의 평행하다",
     "fix_en": "Redraw the lower wooden lip of the bench to taper upward toward the right, creating a slight diagonal line across the bottom frame to simulate an oblique angle. Preserve the judge, the nameplate, the microphone, and the courtroom background.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "명패 글자가 '재판장'이 아니라 사이가 띄어진 '재 판 장'이다",
     "fix_en": "Condense the Korean text on the wooden nameplate to eliminate the wide spaces between the characters. Preserve the bench, the judge's pose and appearance, the microphone, and the background.",
     "severity": "minor",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 가장자리에 배치되어야 할 마이크가 프레임 가장자리가 아닌 중앙 부근(명패 뒤쪽)에 위치해 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "판사의 시선이 마이크가 아니라 화면 왼쪽 법정 쪽을 향한다",
     "severity": "major"
    },
    {
     "issue_ko": "법대 앞모서리가 사선으로 가로지르지 않고 화면 하단과 거의 평행하다",
     "severity": "major"
    },
    {
     "issue_ko": "판사가 화면 중앙이 아니라 오른쪽 필드에 치우쳐 있다",
     "severity": "major"
    },
    {
     "issue_ko": "명패 글자가 '재판장'이 아니라 사이가 띄어진 '재 판 장'이다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 가장자리에 배치되어야 할 마이크가 프레임 가장자리가 아닌 중앙 부근(명패 뒤쪽)에 위치해 있습니다.",
    "fix_en": "Reposition the microphone so its stem and head sit at the far left edge of the frame, rather than near the center. Preserve the judge's face, posture, robe, the nameplate, the bench, and the background.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "법대 앞모서리가 사선으로 가로지르지 않고 화면 하단과 거의 평행하다",
    "fix_en": "Redraw the lower wooden lip of the bench to taper upward toward the right, creating a slight diagonal line across the bottom frame to simulate an oblique angle. Preserve the judge, the nameplate, the microphone, and the courtroom background.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "명패 글자가 '재판장'이 아니라 사이가 띄어진 '재 판 장'이다",
    "fix_en": "Condense the Korean text on the wooden nameplate to eliminate the wide spaces between the characters. Preserve the bench, the judge's pose and appearance, the microphone, and the background.",
    "severity": "minor",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S89sh16"
  }
 },
 "S89sh18::cine": {
  "applied": true,
  "fingerprint": "db10b3e5989a8af97aebd3639a3672de00c9204bc7d11b5b6d003a20df9726f5",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S89sh18_sel.png",
  "source_sha256": "00c821dd32ec08cf2f3ea9d63a66d216ad9849ff4dde2d312f6069aa33738584",
  "file": "S89sh18_cine.png",
  "latency_ms": 10809
 },
 "S89sh20::signage": {
  "fp": "19e36bcde1270e1a",
  "inscriptions": []
 },
 "S89sh20": {
  "input_fingerprint": "0a498ca29f0ca1f5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽을 향해 깊은 신뢰와 안도가 담긴 눈빛으로 시선을 맞추는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): In the courtroom public gallery, looking toward the prosecution side after the verdict. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height just behind and to one side of 장원섭, track into a facial close-up of 전택수, retaining only a soft edge of 장원섭's turned shoulder in the left foreground as the off-axis eyeline reference. 전택수 occupies the center-right, meeting 장원섭's gaze with restrained relief and trust while the audience seating remains softly legible behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, looks toward 전택수 across the courtroom eyeline; 전택수 in the middle-right of the frame, midground, looks toward 장원섭 at the prosecution side.\n- KEY BACKGROUND ELEMENTS: 방청석 (Occupied after the sentence is announced) — The seating row recedes obliquely behind 전택수; used as Soft background context locating 전택수 within the audience after the verdict.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination keeps the relief muted within a subdued cool palette and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom, gallery rows, formal daylight, and the officer's clothing. Exclude the judge and crying supporters from the close frame, isolating the officer's relieved look toward the prosecutor.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽을 향해 깊은 신뢰와 안도가 담긴 눈빛으로 시선을 맞추는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): In the courtroom public gallery, looking toward the prosecution side after the verdict. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height just behind and to one side of 장원섭, track into a facial close-up of 전택수, retaining only a soft edge of 장원섭's turned shoulder in the left foreground as the off-axis eyeline reference. 전택수 occupies the center-right, meeting 장원섭's gaze with restrained relief and trust while the audience seating remains softly legible behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, looks toward 전택수 across the courtroom eyeline; 전택수 in the middle-right of the frame, midground, looks toward 장원섭 at the prosecution side.\n- KEY BACKGROUND ELEMENTS: 방청석 (Occupied after the sentence is announced) — The seating row recedes obliquely behind 전택수; used as Soft background context locating 전택수 within the audience after the verdict.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination keeps the relief muted within a subdued cool palette and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom, gallery rows, formal daylight, and the officer's clothing. Exclude the judge and crying supporters from the close frame, isolating the officer's relieved look toward the prosecutor.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day, with late-night reenactment inserts.\n\nSHOT TEXT (authoritative, Korean): 장원섭 쪽을 향해 깊은 신뢰와 안도가 담긴 눈빛으로 시선을 맞추는 전택수의 얼굴 클로즈업\n\nLOCATION (lock): In the courtroom public gallery, looking toward the prosecution side after the verdict. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At seated eye height just behind and to one side of 장원섭, track into a facial close-up of 전택수, retaining only a soft edge of 장원섭's turned shoulder in the left foreground as the off-axis eyeline reference. 전택수 occupies the center-right, meeting 장원섭's gaze with restrained relief and trust while the audience seating remains softly legible behind him.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: 장원섭 in the middle-left of the frame, foreground, looks toward 전택수 across the courtroom eyeline; 전택수 in the middle-right of the frame, midground, looks toward 장원섭 at the prosecution side.\n- KEY BACKGROUND ELEMENTS: 방청석 (Occupied after the sentence is announced) — The seating row recedes obliquely behind 전택수; used as Soft background context locating 전택수 within the audience after the verdict.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral courtroom illumination keeps the relief muted within a subdued cool palette and naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same courtroom, gallery rows, formal daylight, and the officer's clothing. Exclude the judge and crying supporters from the close frame, isolating the officer's relieved look toward the prosecutor.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "중앙의 판사가 좌측 전경의 경찰관을 바라봄.",
    "built_space": "배경에 법정 방청석과 나무 패널 벽이 보임.",
    "entities": "제외해야 할 판사가 메인으로 등장하고, 지정된 인물인 전택수(경찰관)가 뒷모습으로 전경에 배치됨.",
    "hard_violations": [
     "프롬프트에서 제외를 명시한 인물(판사) 등장",
     "지정된 샷 프레이밍 및 피사체 역할 완전 역전"
    ],
    "physics": "두 인물 모두 안정적으로 앉아 있음."
   },
   {
    "label": "B",
    "direction": "전택수가 좌측 전경의 인물(어깨)을 향해 시선을 맞춤.",
    "built_space": "배경에 법정 방청석 의자들이 알맞게 나열되어 있음.",
    "entities": "전택수의 얼굴과 복장(모자, 재킷)이 캐릭터 레퍼런스와 정확히 일치함.",
    "hard_violations": [],
    "physics": "인물이 자세를 유지하며 안정적으로 앉아 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 3,
   "B": 8
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "프롬프트에서 제외하라고 명시한 판사를 화면 중심에 배치하고, 주인공인 전택수를 전경에 뒷모습으로 배치하여 지시를 완전히 위반했습니다."
   },
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "전택수를 화면 중앙 우측에 클로즈업으로 배치하고 전경에 어깨를 걸쳐 앵글을 구성한 프레이밍 지시를 정확히 수행했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S89sh18_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 좌측(인물의 우측 팔) 소매에 부착된 패치에 알아볼 수 없는 뭉개진 문자가 있습니다.",
     "fix_en": "Blur the illegible text on the left sleeve patch into a solid texture. Maintain the rest of the uniform, the man's face, and the background.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "전택수가 얼굴 클로즈업이 아니라 가슴까지 보이는 오버숄더 중클로즈업으로 잡혀 있다",
     "fix_en": "Crop the frame to a tight facial close-up of the main subject. Maintain the man's facial features, lighting, and the background.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "왼쪽 전경에 장원섭의 어깨 가장자리만 있어야 하는데 뒷머리 전체가 크게 들어와 있다",
     "fix_en": "Replace the back of the head in the left foreground with a soft, out-of-focus shoulder edge. Maintain the main subject and the courtroom background.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "화면 중앙 전택수의 얼굴이 캐릭터 레퍼런스의 얼굴·이목구비와 다르다",
     "fix_en": "Redraw the man's facial features to match the provided character reference exactly. Maintain his uniform, hat, lighting, and the courtroom background.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "전택수 표정이 신뢰와 안도 대신 무표정·침중에 가깝다",
     "fix_en": "Soften the man's eyes and mouth to show subtle relief and trust. Maintain his facial identity, uniform, hat, and the background.",
     "severity": "minor",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 좌측(인물의 우측 팔) 소매에 부착된 패치에 알아볼 수 없는 뭉개진 문자가 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "전택수가 얼굴 클로즈업이 아니라 가슴까지 보이는 오버숄더 중클로즈업으로 잡혀 있다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 전경에 장원섭의 어깨 가장자리만 있어야 하는데 뒷머리 전체가 크게 들어와 있다",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙 전택수의 얼굴이 캐릭터 레퍼런스의 얼굴·이목구비와 다르다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수 표정이 신뢰와 안도 대신 무표정·침중에 가깝다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 5,
  "fix_severity_skipped": [
   {
    "issue_ko": "화면 좌측(인물의 우측 팔) 소매에 부착된 패치에 알아볼 수 없는 뭉개진 문자가 있습니다.",
    "fix_en": "Blur the illegible text on the left sleeve patch into a solid texture. Maintain the rest of the uniform, the man's face, and the background.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "전택수가 얼굴 클로즈업이 아니라 가슴까지 보이는 오버숄더 중클로즈업으로 잡혀 있다",
    "fix_en": "Crop the frame to a tight facial close-up of the main subject. Maintain the man's facial features, lighting, and the background.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "왼쪽 전경에 장원섭의 어깨 가장자리만 있어야 하는데 뒷머리 전체가 크게 들어와 있다",
    "fix_en": "Replace the back of the head in the left foreground with a soft, out-of-focus shoulder edge. Maintain the main subject and the courtroom background.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "화면 중앙 전택수의 얼굴이 캐릭터 레퍼런스의 얼굴·이목구비와 다르다",
    "fix_en": "Redraw the man's facial features to match the provided character reference exactly. Maintain his uniform, hat, lighting, and the courtroom background.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "전택수 표정이 신뢰와 안도 대신 무표정·침중에 가깝다",
    "fix_en": "Soften the man's eyes and mouth to show subtle relief and trust. Maintain his facial identity, uniform, hat, and the background.",
    "severity": "minor",
    "observation_index": 4
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S89sh18"
  }
 },
 "S89sh20::cine": {
  "applied": true,
  "fingerprint": "379aeb4be209f5dae09bd23291c7fa9eede59779433b4f87da7776dfd9fec113",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S89sh20_sel.png",
  "source_sha256": "f98ee37887bf750f26a7bb760b775fcd08d3723c956c2dc86cbee8028a6d9064",
  "file": "S89sh20_cine.png",
  "latency_ms": 12048
 },
 "S90sh5::confined_fp_apt": {
  "applies": false,
  "reason_ko": "이 샷은 차량 내부에서 창밖의 촛불 집회 군중을 바라보는 뷰로, 탑승자의 구체적인 좌석 위치나 차량 내 조종 장치의 정확한 레이아웃이 서사적으로 중요하지 않은 단순 프레이밍 장면이므로 평면도 가이드가 필요하지 않습니다.",
  "input_fingerprint": "50cd5948280a354f"
 },
 "S90sh5::signage": {
  "fp": "dd59fc40bbec1e2c",
  "inscriptions": [
   {
    "surface_native": "피켓",
    "text_native": "평화",
    "reason_ko": "한국의 평화로운 촛불 집회 현장 속 참가자가 들고 있는 피켓에 적힌 문구로, 시대적 배경과 어울리는 평화적인 메시지를 전달하기 위해 필요합니다."
   }
  ]
 },
 "era_assess::ea9ea44791f756f6": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국 승용차 및 택시 뒷좌석 내부",
    "search_terms_native": [
     "yf소나타 뒷좌석",
     "국산 세단 뒷좌석 내부",
     "택시 뒷좌석 시트"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어나 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "2000년대와 2010년대 한국에서 흔히 볼 수 있었던 택시나 세단 차량 뒷좌석의 전형적인 가죽 시트 형태와 내장재 디자인을 살리지 못하고 현대적인 서구권 고급 승용차로 잘못 묘사할 수 있습니다."
   },
   {
    "subject_native": "대한민국 촛불집회 거리 풍경 (2000년대~2010년대)",
    "search_terms_native": [
     "광화문 촛불시위",
     "촛불집회 종이컵",
     "거리 촛불집회"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 영어나 다른 언어로 번역하거나 추가해서는 안 됩니다.",
    "reason_ko": "서구식의 긴 알몸 양초나 교회식 촛불 모임으로 오해하기 쉬우나, 한국의 촛불집회는 바람을 막기 위해 양초 아래에 백색 종이컵을 끼우고 밀집해 있는 고유한 시각적 특징이 있습니다."
   }
  ]
 },
 "era_ref::0d6c22f616eb0e85": {
  "subject": "2000년대~2010년대 한국 승용차 및 택시 뒷좌석 내부",
  "terms": [
   "yf소나타 뒷좌석",
   "국산 세단 뒷좌석 내부",
   "택시 뒷좌석 시트"
  ],
  "queries": [
   [
    "YF쏘나타 뒷좌석 내부 택시 시트",
    "2000년대 2010년대 국산 세단 택시 뒷좌석 내부"
   ],
   [
    "YF쏘나타 뒷좌석 내부 사진",
    "YF쏘나타 택시 뒷좌석 시트 사진",
    "국산 중형 세단 택시 뒷좌석 내부 사진"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://motoristprod.s3.ap-southeast-1.amazonaws.com/private_uploads/sale_photo/image/3265854/thumb_20260331-895119-13fg7g.jpeg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAJLZVGH7SK6OPR6TQ%2F20260331%2Fap-southeast-1%2Fs3%2Faws4_request&X-Amz-Date=20260331T094514Z&X-Amz-Expires=604800&X-Amz-Signature=bab873fe1d1113f764694c2eb334d05303177f5fd9e2d5b55fe65bd6945f5d52&X-Amz-SignedHeaders=host",
  "picked_reason_ko": "2번은 지정 시대의 한국 승용차 뒷좌석을 가장 평범하고 완전하게 보여 주며, 재질·비례·문짝·안전장치와 편의 설비까지 명확히 읽을 수 있다.",
  "sha256": "8ca7df56f76d88f9464c644727e5f5362b6b80394e8567761c2ce681dc48498e",
  "file": "eraref_0d6c22f616eb0e85.png"
 },
 "S90sh5::bgfirst_bg": {
  "input_fingerprint": "d857a39dfcd7b6bc",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노을 빛이 반사된 차창 유리 너머, 손에 촛불을 든 수많은 한국인 남녀들이 거리를 메운 평화로운 풍경\n\nLOCATION (lock): Inside the moving car beside a sunset-lit side window, with the rear seat, glass, and vehicle fittings closely framing the candlelit street outside.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track parallel to the moving car from immediately outside the rear side window at seated head height, looking obliquely through the glass in a wide layered composition. The sunset reflection remains across the window while candle-bearing Korean men and women fill the street beyond, their slow walking phases, shoulder angles, spacing, and candle heights naturally varied rather than synchronized.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: car side window (moving past the street with the car) — The sunset is reflected across the camera-facing surface while the candle-bearing crowd remains visible through the glass; used as Layered optical plane between the camera and the procession; city street (filled with people walking slowly); used as Deep background establishing the scale and peaceful pace of the gathering; candles (lit and carried by the participants); used as Repeated points held throughout the crowd without uniform spacing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Reflected sunset and candlelight provide restrained warmth within an otherwise subdued naturalistic dusk palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 승용차 및 택시 뒷좌석 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 노을 빛이 반사된 차창 유리 너머, 손에 촛불을 든 수많은 한국인 남녀들이 거리를 메운 평화로운 풍경\n\nLOCATION (lock): Inside the moving car beside a sunset-lit side window, with the rear seat, glass, and vehicle fittings closely framing the candlelit street outside.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Track parallel to the moving car from immediately outside the rear side window at seated head height, looking obliquely through the glass in a wide layered composition. The sunset reflection remains across the window while candle-bearing Korean men and women fill the street beyond, their slow walking phases, shoulder angles, spacing, and candle heights naturally varied rather than synchronized.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: car side window (moving past the street with the car) — The sunset is reflected across the camera-facing surface while the candle-bearing crowd remains visible through the glass; used as Layered optical plane between the camera and the procession; city street (filled with people walking slowly); used as Deep background establishing the scale and peaceful pace of the gathering; candles (lit and carried by the participants); used as Repeated points held throughout the crowd without uniform spacing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Reflected sunset and candlelight provide restrained warmth within an otherwise subdued naturalistic dusk palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국 승용차 및 택시 뒷좌석 내부: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S90sh5__bgfirst_bg.png",
  "asset_id": "20456562-9927-464e-95b1-4f4942b69faf",
  "input_asset_ids": [
   "464a8184-1970-48c7-804a-4449b786476c",
   "0c1cefd9-6767-4f28-9c75-05b0eb689eae"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국 승용차 및 택시 뒷좌석 내부",
   "queries": [
    [
     "YF쏘나타 뒷좌석 내부 택시 시트",
     "2000년대 2010년대 국산 세단 택시 뒷좌석 내부"
    ],
    [
     "YF쏘나타 뒷좌석 내부 사진",
     "YF쏘나타 택시 뒷좌석 시트 사진",
     "국산 중형 세단 택시 뒷좌석 내부 사진"
    ]
   ],
   "picked_url": "https://motoristprod.s3.ap-southeast-1.amazonaws.com/private_uploads/sale_photo/image/3265854/thumb_20260331-895119-13fg7g.jpeg?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAJLZVGH7SK6OPR6TQ%2F20260331%2Fap-southeast-1%2Fs3%2Faws4_request&X-Amz-Date=20260331T094514Z&X-Amz-Expires=604800&X-Amz-Signature=bab873fe1d1113f764694c2eb334d05303177f5fd9e2d5b55fe65bd6945f5d52&X-Amz-SignedHeaders=host",
   "sha256": "8ca7df56f76d88f9464c644727e5f5362b6b80394e8567761c2ce681dc48498e",
   "file": "eraref_0d6c22f616eb0e85.png"
  }
 },
 "S90sh5": {
  "input_fingerprint": "99ddd85093d03014",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 노을 빛이 반사된 차창 유리 너머, 손에 촛불을 든 수많은 한국인 남녀들이 거리를 메운 평화로운 풍경\n\nLOCATION (lock): Inside the moving car beside a sunset-lit side window, with the rear seat, glass, and vehicle fittings closely framing the candlelit street outside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track parallel to the moving car from immediately outside the rear side window at seated head height, looking obliquely through the glass in a wide layered composition. The sunset reflection remains across the window while candle-bearing Korean men and women fill the street beyond, their slow walking phases, shoulder angles, spacing, and candle heights naturally varied rather than synchronized.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: car side window (moving past the street with the car) — The sunset is reflected across the camera-facing surface while the candle-bearing crowd remains visible through the glass; used as Layered optical plane between the camera and the procession; city street (filled with people walking slowly); used as Deep background establishing the scale and peaceful pace of the gathering; candles (lit and carried by the participants); used as Repeated points held throughout the crowd without uniform spacing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Reflected sunset and candlelight provide restrained warmth within an otherwise subdued naturalistic dusk palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candle-bearing crowd continues filling the road around Taeksu's slowly moving car.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 수많은 한국인 남녀들 right now — 수많은 한국인 남녀들's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 수많은 한국인 남녀들 and no one else. The framing stays on the object.\n\nNo people appear unless the shot text itself says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피켓: \"평화\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective, every panel, control, seat, mirror, window and fixture stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: how many steering wheels or control surfaces, how many seats and which way each faces, where each mirror sits and what it could reflect from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second rim, sliding a seat, turning a mirror or growing a new panel is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: which seat the body occupies, which side of the wheel the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 노을 빛이 반사된 차창 유리 너머, 손에 촛불을 든 수많은 한국인 남녀들이 거리를 메운 평화로운 풍경\n\nLOCATION (lock): Inside the moving car beside a sunset-lit side window, with the rear seat, glass, and vehicle fittings closely framing the candlelit street outside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track parallel to the moving car from immediately outside the rear side window at seated head height, looking obliquely through the glass in a wide layered composition. The sunset reflection remains across the window while candle-bearing Korean men and women fill the street beyond, their slow walking phases, shoulder angles, spacing, and candle heights naturally varied rather than synchronized.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: car side window (moving past the street with the car) — The sunset is reflected across the camera-facing surface while the candle-bearing crowd remains visible through the glass; used as Layered optical plane between the camera and the procession; city street (filled with people walking slowly); used as Deep background establishing the scale and peaceful pace of the gathering; candles (lit and carried by the participants); used as Repeated points held throughout the crowd without uniform spacing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Reflected sunset and candlelight provide restrained warmth within an otherwise subdued naturalistic dusk palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candle-bearing crowd continues filling the road around Taeksu's slowly moving car.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 수많은 한국인 남녀들 right now — 수많은 한국인 남녀들's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 수많은 한국인 남녀들 and no one else. The framing stays on the object.\n\nNo people appear unless the shot text itself says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피켓: \"평화\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 노을 빛이 반사된 차창 유리 너머, 손에 촛불을 든 수많은 한국인 남녀들이 거리를 메운 평화로운 풍경\n\nLOCATION (lock): Inside the moving car beside a sunset-lit side window, with the rear seat, glass, and vehicle fittings closely framing the candlelit street outside. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Track parallel to the moving car from immediately outside the rear side window at seated head height, looking obliquely through the glass in a wide layered composition. The sunset reflection remains across the window while candle-bearing Korean men and women fill the street beyond, their slow walking phases, shoulder angles, spacing, and candle heights naturally varied rather than synchronized.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: car side window (moving past the street with the car) — The sunset is reflected across the camera-facing surface while the candle-bearing crowd remains visible through the glass; used as Layered optical plane between the camera and the procession; city street (filled with people walking slowly); used as Deep background establishing the scale and peaceful pace of the gathering; candles (lit and carried by the participants); used as Repeated points held throughout the crowd without uniform spacing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Reflected sunset and candlelight provide restrained warmth within an otherwise subdued naturalistic dusk palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candle-bearing crowd continues filling the road around Taeksu's slowly moving car.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 수많은 한국인 남녀들 right now — 수많은 한국인 남녀들's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 수많은 한국인 남녀들 and no one else. The framing stays on the object.\n\nNo people appear unless the shot text itself says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 피켓: \"평화\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S90sh5__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched panels, seats, controls and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S90sh5.png"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L10B01.png"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 750,
      "verdict_ko": "차량 내부에서 창문 밖을 내다보는 구도를 취했으나, 화면 우측 좌석이 차량 측면을 향해 배치된 물리적으로 불가능한 실내 구조를 생성하여 심각한 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 구조 (우측 좌석이 전방이 아닌 측면 창문을 향해 있음) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 호명하지 않은 운전자를 전경에 발명함 / [openrouter:x-ai/grok-4.6] 약속된 카메라 위치(뒷옆유리 바로 바깥)를 실내 시점으로 바꿈"
     },
     {
      "label": "B",
      "score": 917,
      "verdict_ko": "지정된 차량 내부 위치를 위반하고 외부에서 촬영된 구도이며, 창문에 반사된 피켓의 텍스트가 좌우 반전되지 않은 물리적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 현상 (창문에 반사된 피켓의 '평화' 텍스트가 좌우 반전되지 않고 정방향으로 보임) / [gemini-pro] 지정된 카메라 위치(LOCATION lock) 위반 (차량 내부가 아닌 외부에서 촬영됨) / [openrouter:x-ai/grok-4.6] 외부에서 옆유리를 보면 있어야 할 차실이 소거되고 유리 직후에 거리 군중이 있어 물리적으로 불가능한 연출"
     }
    ],
    "all_candidates_fail": true,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.667
     },
     "adjusted": {
      "A": 0.75,
      "B": 0.917
     },
     "violations": {
      "A": [
       "[gemini-pro] 물리적으로 불가능한 구조 (우측 좌석이 전방이 아닌 측면 창문을 향해 있음)",
       "[openrouter:x-ai/grok-4.6] 샷 텍스트가 호명하지 않은 운전자를 전경에 발명함",
       "[openrouter:x-ai/grok-4.6] 약속된 카메라 위치(뒷옆유리 바로 바깥)를 실내 시점으로 바꿈"
      ],
      "B": [
       "[gemini-pro] 물리적으로 불가능한 현상 (창문에 반사된 피켓의 '평화' 텍스트가 좌우 반전되지 않고 정방향으로 보임)",
       "[gemini-pro] 지정된 카메라 위치(LOCATION lock) 위반 (차량 내부가 아닌 외부에서 촬영됨)",
       "[openrouter:x-ai/grok-4.6] 외부에서 옆유리를 보면 있어야 할 차실이 소거되고 유리 직후에 거리 군중이 있어 물리적으로 불가능한 연출"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.5,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 750,
      "verdict_ko": "차량 내부에서 창문 밖을 내다보는 구도를 취했으나, 화면 우측 좌석이 차량 측면을 향해 배치된 물리적으로 불가능한 실내 구조를 생성하여 심각한 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 구조 (우측 좌석이 전방이 아닌 측면 창문을 향해 있음) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 호명하지 않은 운전자를 전경에 발명함 / [openrouter:x-ai/grok-4.6] 약속된 카메라 위치(뒷옆유리 바로 바깥)를 실내 시점으로 바꿈"
     },
     {
      "label": "B",
      "score": 917,
      "verdict_ko": "지정된 차량 내부 위치를 위반하고 외부에서 촬영된 구도이며, 창문에 반사된 피켓의 텍스트가 좌우 반전되지 않은 물리적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 현상 (창문에 반사된 피켓의 '평화' 텍스트가 좌우 반전되지 않고 정방향으로 보임) / [gemini-pro] 지정된 카메라 위치(LOCATION lock) 위반 (차량 내부가 아닌 외부에서 촬영됨) / [openrouter:x-ai/grok-4.6] 외부에서 옆유리를 보면 있어야 할 차실이 소거되고 유리 직후에 거리 군중이 있어 물리적으로 불가능한 연출"
     }
    ],
    "all_candidates_fail": true
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 750,
      "verdict_ko": "차량 내부 구조와 인물들의 배치를 물리적 오류 없이 구현했으나, 지시문에 명시된 '유리에 반사된 노을'과 유리창 레이어 효과를 누락하고 창문이 열린 상태로 묘사하여 아쉽습니다.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 운전자를 창작함 / [openrouter:x-ai/grok-4.6] 지정 카메라는 차 바로 바깥 뒷옆창 머리 높이인데 실내 뒷좌석 시점으로 찍힘"
     },
     {
      "label": "A",
      "score": 1000,
      "verdict_ko": "차창 겉면에 반사된 풍경을 묘사했음에도 '평화' 피켓의 글씨가 좌우 반전되지 않아 물리적, 광학적으로 불가능한 치명적 오류(Hard Violation)가 발생했습니다.  ★위반: [gemini-pro] 반사된 풍경임에도 '평화' 텍스트가 좌우 반전되지 않고 정방향으로 읽히는 광학적/물리적 오류 / [gemini-pro] 차량 내부를 관통해 반대편을 보는 것이라면 차량 내부 구조가 전혀 없는 공간적 오류"
     }
    ],
    "all_candidates_fail": false,
    "dual": {
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ],
     "normalized": {
      "A": 1.5,
      "B": 1.25
     },
     "adjusted": {
      "A": 1.0,
      "B": 0.75
     },
     "violations": {
      "A": [
       "[gemini-pro] 반사된 풍경임에도 '평화' 텍스트가 좌우 반전되지 않고 정방향으로 읽히는 광학적/물리적 오류",
       "[gemini-pro] 차량 내부를 관통해 반대편을 보는 것이라면 차량 내부 구조가 전혀 없는 공간적 오류"
      ],
      "B": [
       "[openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 운전자를 창작함",
       "[openrouter:x-ai/grok-4.6] 지정 카메라는 차 바로 바깥 뒷옆창 머리 높이인데 실내 뒷좌석 시점으로 찍힘"
      ]
     },
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "agreed": false
    },
    "gq": {
     "route": "combined",
     "gap": 0.75,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 750,
      "verdict_ko": "차량 내부 구조와 인물들의 배치를 물리적 오류 없이 구현했으나, 지시문에 명시된 '유리에 반사된 노을'과 유리창 레이어 효과를 누락하고 창문이 열린 상태로 묘사하여 아쉽습니다.  ★위반: [openrouter:x-ai/grok-4.6] 샷 텍스트에 없는 운전자를 창작함 / [openrouter:x-ai/grok-4.6] 지정 카메라는 차 바로 바깥 뒷옆창 머리 높이인데 실내 뒷좌석 시점으로 찍힘"
     },
     {
      "label": "B",
      "score": 1000,
      "verdict_ko": "차창 겉면에 반사된 풍경을 묘사했음에도 '평화' 피켓의 글씨가 좌우 반전되지 않아 물리적, 광학적으로 불가능한 치명적 오류(Hard Violation)가 발생했습니다.  ★위반: [gemini-pro] 반사된 풍경임에도 '평화' 텍스트가 좌우 반전되지 않고 정방향으로 읽히는 광학적/물리적 오류 / [gemini-pro] 차량 내부를 관통해 반대편을 보는 것이라면 차량 내부 구조가 전혀 없는 공간적 오류"
     }
    ],
    "all_candidates_fail": false
   },
   "combined": {
    "totals": {
     "A": 1500,
     "B": 1917
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "totals": {
   "A": 1500,
   "B": 1917
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 750,
    "verdict_ko": "차량 내부에서 창문 밖을 내다보는 구도를 취했으나, 화면 우측 좌석이 차량 측면을 향해 배치된 물리적으로 불가능한 실내 구조를 생성하여 심각한 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 구조 (우측 좌석이 전방이 아닌 측면 창문을 향해 있음) / [openrouter:x-ai/grok-4.6] 샷 텍스트가 호명하지 않은 운전자를 전경에 발명함 / [openrouter:x-ai/grok-4.6] 약속된 카메라 위치(뒷옆유리 바로 바깥)를 실내 시점으로 바꿈"
   },
   {
    "label": "B",
    "score": 917,
    "verdict_ko": "지정된 차량 내부 위치를 위반하고 외부에서 촬영된 구도이며, 창문에 반사된 피켓의 텍스트가 좌우 반전되지 않은 물리적 오류가 있습니다.  ★위반: [gemini-pro] 물리적으로 불가능한 현상 (창문에 반사된 피켓의 '평화' 텍스트가 좌우 반전되지 않고 정방향으로 보임) / [gemini-pro] 지정된 카메라 위치(LOCATION lock) 위반 (차량 내부가 아닌 외부에서 촬영됨) / [openrouter:x-ai/grok-4.6] 외부에서 옆유리를 보면 있어야 할 차실이 소거되고 유리 직후에 거리 군중이 있어 물리적으로 불가능한 연출"
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L10B01.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "차 실내·뒷좌석이 창과 함께 거리를 프레이밍하지 않고, 바깥에서 본 옆유리 바로 뒤에 군중이 실내 공간을 채워 창 너머 시선이 성립하지 않는다.",
     "fix_en": "Redraw the space behind the glass to show the car's dark interior seating and opposite window framing the street outside. Maintain the car's exterior shell and the sunset reflection on the glass.",
     "severity": "critical",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "좌측 상단의 '평화' 피켓을 지탱하는 막대나 잡고 있는 손이 보이지 않아 피켓이 군중 위에 떠 있음",
     "fix_en": "Add a wooden pole extending down from the sign, gripped by an anatomically correct hand from the crowd. Maintain the sign's text, the people, and the lighting.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "앞줄 좌측에서 두 번째 줄무늬 셔츠를 입은 여성과 우측 끝 남성이 촛불을 쥐고 있는 손가락 형태가 심하게 뭉개지고 해부학적으로 기형적임",
     "fix_en": "Redraw the hands of the striped-shirt woman and the far-right man so they anatomically correctly grip their candles. Maintain their faces, clothing, the candles, and the crowd.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "자연스러운 걷기 묘사가 아닌, 앞줄의 좌측 남성, 줄무늬 셔츠 여성, 중앙의 안경 쓴 남성 등이 일렬로 서서 카메라 렌즈를 정면으로 응시하고 있음",
     "fix_en": "Turn the eyes and head angles of the front row people so they look forward along the street rather than into the lens. Maintain their facial features, clothing, and positions.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "창 너머 배경에 참조 장소의 하천·난간·지면 등불 줄이 없고 다른 돌담과 도로가 보인다.",
     "fix_en": "Replace the background stone wall with the location reference's river, railing, and row of ground-level lanterns. Maintain the crowd, the car, and the lighting.",
     "severity": "major",
     "observation_index": 5
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "창문에 비친 상(또는 창문 너머 풍경)임에도 '평화' 피켓의 글씨가 좌우 반전되지 않았고, 유리창 너머로 차량 내부가 전혀 보이지 않아 광학적으로 불가능한 구조임",
     "severity": "critical"
    },
    {
     "issue_ko": "좌측 상단의 '평화' 피켓을 지탱하는 막대나 잡고 있는 손이 보이지 않아 피켓이 군중 위에 떠 있음",
     "severity": "critical"
    },
    {
     "issue_ko": "자연스러운 걷기 묘사가 아닌, 앞줄의 좌측 남성, 줄무늬 셔츠 여성, 중앙의 안경 쓴 남성 등이 일렬로 서서 카메라 렌즈를 정면으로 응시하고 있음",
     "severity": "major"
    },
    {
     "issue_ko": "앞줄 좌측에서 두 번째 줄무늬 셔츠를 입은 여성과 우측 끝 남성이 촛불을 쥐고 있는 손가락 형태가 심하게 뭉개지고 해부학적으로 기형적임",
     "severity": "critical"
    },
    {
     "issue_ko": "차 실내·뒷좌석이 창과 함께 거리를 프레이밍하지 않고, 바깥에서 본 옆유리 바로 뒤에 군중이 실내 공간을 채워 창 너머 시선이 성립하지 않는다.",
     "severity": "critical"
    },
    {
     "issue_ko": "창 너머 배경에 참조 장소의 하천·난간·지면 등불 줄이 없고 다른 돌담과 도로가 보인다.",
     "severity": "major"
    },
    {
     "issue_ko": "앞줄 남녀가 화면을 향해 걷고 다수가 렌즈를 정면 응시해 평화로운 행진이 아니라 단체 초상처럼 보인다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "자연스러운 걷기 묘사가 아닌, 앞줄의 좌측 남성, 줄무늬 셔츠 여성, 중앙의 안경 쓴 남성 등이 일렬로 서서 카메라 렌즈를 정면으로 응시하고 있음",
    "fix_en": "Turn the eyes and head angles of the front row people so they look forward along the street rather than into the lens. Maintain their facial features, clothing, and positions.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "창 너머 배경에 참조 장소의 하천·난간·지면 등불 줄이 없고 다른 돌담과 도로가 보인다.",
    "fix_en": "Replace the background stone wall with the location reference's river, railing, and row of ground-level lanterns. Maintain the crowd, the car, and the lighting.",
    "severity": "major",
    "observation_index": 5
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Redraw the space behind the glass to show the car's dark interior seating and opposite window framing the street outside. Maintain the car's exterior shell and the sunset reflection on the glass.\n- Add a wooden pole extending down from the sign, gripped by an anatomically correct hand from the crowd. Maintain the sign's text, the people, and the lighting.\n- Redraw the hands of the striped-shirt woman and the far-right man so they anatomically correctly grip their candles. Maintain their faces, clothing, the candles, and the crowd.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "차량 내부의 디테일이 생략되어 유리를 투과하는 시선인지 반사된 장면인지 광학적으로 다소 모호하지만, 지정된 카메라 위치, 촛불을 든 군중, 노을 반사, 정확한 피켓 텍스트 등 프롬프트의 요구사항을 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "차량 뒷좌석 시트 위에 사람들이 미니어처처럼 서서 걷고 있는 물리적으로 불가능한 구도가 형성되어 치명적인 오류(Hard Violation)가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 차량 외부에서 뒷좌석 창문을 바라보고 있으며, 창문에 비치거나 투과된 사람들은 카메라 쪽을 향해 걷고 있습니다. 피켓의 글씨 역시 카메라를 정확히 향하고 있습니다.",
      "built_space": "차량 외부의 도어 프레임과 외부 손잡이가 보입니다. 창문 유리에는 노을과 군중이 겹쳐 보이나, 창문 너머로 당연히 보여야 할 차량 내부(좌석 등)의 공간감은 묘사되지 않았습니다.",
      "entities": "촛불을 든 한국인 남녀 무리, '평화'라고 정확히 적힌 피켓, 노을 진 하늘 등 프롬프트가 요구한 개체들이 모두 확인됩니다.",
      "hard_violations": [],
      "physics": "사람들은 땅을 딛고 자연스럽게 걷고 있으며, 손으로 촛불과 피켓의 막대를 정상적으로 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "카메라는 차량 외부에서 뒷좌석 창문을 향하고 있으며, 사람들은 카메라 쪽으로 걷고 있고 피켓의 텍스트도 정면을 향합니다.",
      "built_space": "차량 외부 도어와 창문, 그리고 창문 너머로 차량 내부의 뒷좌석 시트와 헤드레스트가 뚜렷하게 묘사되어 있습니다.",
      "entities": "촛불을 든 한국인 남녀, '평화'가 적힌 피켓, 노을이 지는 배경이 등장합니다.",
      "hard_violations": [
       "물리적으로 불가능한 연출(invented physically impossible staging): 군중이 차량 밖 거리가 아닌 차량 내부의 뒷좌석 시트 위에 서서 걷고 있으며, 사람들의 크기가 좌석 크기에 비해 터무니없이 작게 축소되어 있음"
      ],
      "physics": "사람들이 디뎌야 할 거리가 아니라 차량 내부의 푹신한 시트 위에 떠 있거나 서 있으며, 스케일이 전혀 맞지 않는 상태로 연출되었습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "차량 내부의 디테일이 생략되어 유리를 투과하는 시선인지 반사된 장면인지 광학적으로 다소 모호하지만, 지정된 카메라 위치, 촛불을 든 군중, 노을 반사, 정확한 피켓 텍스트 등 프롬프트의 요구사항을 충실히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "차량 뒷좌석 시트 위에 사람들이 미니어처처럼 서서 걷고 있는 물리적으로 불가능한 구도가 형성되어 치명적인 오류(Hard Violation)가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 차량 외부에서 뒷좌석 창문을 바라보고 있으며, 창문에 비치거나 투과된 사람들은 카메라 쪽을 향해 걷고 있습니다. 피켓의 글씨 역시 카메라를 정확히 향하고 있습니다.",
      "built_space": "차량 외부의 도어 프레임과 외부 손잡이가 보입니다. 창문 유리에는 노을과 군중이 겹쳐 보이나, 창문 너머로 당연히 보여야 할 차량 내부(좌석 등)의 공간감은 묘사되지 않았습니다.",
      "entities": "촛불을 든 한국인 남녀 무리, '평화'라고 정확히 적힌 피켓, 노을 진 하늘 등 프롬프트가 요구한 개체들이 모두 확인됩니다.",
      "hard_violations": [],
      "physics": "사람들은 땅을 딛고 자연스럽게 걷고 있으며, 손으로 촛불과 피켓의 막대를 정상적으로 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "카메라는 차량 외부에서 뒷좌석 창문을 향하고 있으며, 사람들은 카메라 쪽으로 걷고 있고 피켓의 텍스트도 정면을 향합니다.",
      "built_space": "차량 외부 도어와 창문, 그리고 창문 너머로 차량 내부의 뒷좌석 시트와 헤드레스트가 뚜렷하게 묘사되어 있습니다.",
      "entities": "촛불을 든 한국인 남녀, '평화'가 적힌 피켓, 노을이 지는 배경이 등장합니다.",
      "hard_violations": [
       "물리적으로 불가능한 연출(invented physically impossible staging): 군중이 차량 밖 거리가 아닌 차량 내부의 뒷좌석 시트 위에 서서 걷고 있으며, 사람들의 크기가 좌석 크기에 비해 터무니없이 작게 축소되어 있음"
      ],
      "physics": "사람들이 디뎌야 할 거리가 아니라 차량 내부의 푹신한 시트 위에 떠 있거나 서 있으며, 스케일이 전혀 맞지 않는 상태로 연출되었습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "창문에 노을이 반사되는 디테일은 있으나, 사람들이 거리가 아닌 자동차 뒷좌석 위를 걷는 미니어처 크기로 렌더링되어 물리적 스케일과 공간 설정에 치명적인 오류가 있습니다."
     },
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "B와 마찬가지로 거대한 자동차 뒷좌석 쿠션 위를 걷는 미니어처 사람들로 묘사되어 물리적으로 불가능한 공간 왜곡이 발생한 실패작입니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "카메라는 차량 외부에서 뒷좌석 측면 창문을 향하고 있으며, 창문 너머의 사람들은 앞을 향해 걷고 있음. '평화' 피켓은 정방향으로 잘 보임.",
      "built_space": "자동차 외부 문틀과 창문이 보이고 유리에 노을과 도시 풍경이 반사되어 있음. 그러나 창문 너머로 보여야 할 거리가 아니라 차량 내부의 뒷좌석 시트가 바닥으로 묘사됨.",
      "entities": "촛불을 든 수많은 한국인 남녀들과 '평화'라는 한글 피켓이 프롬프트에 맞게 묘사됨.",
      "hard_violations": [
       "Physically impossible staging (미니어처 크기의 사람들이 거리 대신 자동차 뒷좌석 쿠션 위를 걷고 있음)"
      ],
      "physics": "사람들의 발이 자동차 뒷좌석 직물 위에 닿아 그림자를 드리우고 있으며, 주변 차량 시트와 헤드레스트에 비해 사람들이 손가락만한 크기로 축소된 물리적 불가능 상태임."
     },
     {
      "label": "A",
      "direction": "카메라는 차량 외부에서 뒷좌석 측면 창문을 향하고 있으며, 사람들은 촛불과 피켓을 들고 정면을 향해 걷고 있음.",
      "built_space": "자동차 외부 창문 프레임이 보이며, 내부에는 거대한 크기의 뒷좌석 시트와 헤드레스트가 보임.",
      "entities": "촛불을 든 남녀 무리와 '평화'가 적힌 피켓이 존재함.",
      "hard_violations": [
       "Physically impossible staging (미니어처 크기의 사람들이 거리 대신 자동차 뒷좌석 쿠션 위를 걷고 있음)"
      ],
      "physics": "사람들이 거리가 아닌 거대한 자동차 뒷좌석 쿠션 위를 걷고 있으며, 공간과 사람 간의 크기 비율이 완전히 무너진 채 시트 위에 물리적으로 놓여 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 0,
      "verdict_ko": "창문에 노을이 반사되는 디테일은 있으나, 사람들이 거리가 아닌 자동차 뒷좌석 위를 걷는 미니어처 크기로 렌더링되어 물리적 스케일과 공간 설정에 치명적인 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 0,
      "verdict_ko": "B와 마찬가지로 거대한 자동차 뒷좌석 쿠션 위를 걷는 미니어처 사람들로 묘사되어 물리적으로 불가능한 공간 왜곡이 발생한 실패작입니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 차량 외부에서 뒷좌석 측면 창문을 향하고 있으며, 창문 너머의 사람들은 앞을 향해 걷고 있음. '평화' 피켓은 정방향으로 잘 보임.",
      "built_space": "자동차 외부 문틀과 창문이 보이고 유리에 노을과 도시 풍경이 반사되어 있음. 그러나 창문 너머로 보여야 할 거리가 아니라 차량 내부의 뒷좌석 시트가 바닥으로 묘사됨.",
      "entities": "촛불을 든 수많은 한국인 남녀들과 '평화'라는 한글 피켓이 프롬프트에 맞게 묘사됨.",
      "hard_violations": [
       "Physically impossible staging (미니어처 크기의 사람들이 거리 대신 자동차 뒷좌석 쿠션 위를 걷고 있음)"
      ],
      "physics": "사람들의 발이 자동차 뒷좌석 직물 위에 닿아 그림자를 드리우고 있으며, 주변 차량 시트와 헤드레스트에 비해 사람들이 손가락만한 크기로 축소된 물리적 불가능 상태임."
     },
     {
      "label": "B",
      "direction": "카메라는 차량 외부에서 뒷좌석 측면 창문을 향하고 있으며, 사람들은 촛불과 피켓을 들고 정면을 향해 걷고 있음.",
      "built_space": "자동차 외부 창문 프레임이 보이며, 내부에는 거대한 크기의 뒷좌석 시트와 헤드레스트가 보임.",
      "entities": "촛불을 든 남녀 무리와 '평화'가 적힌 피켓이 존재함.",
      "hard_violations": [
       "Physically impossible staging (미니어처 크기의 사람들이 거리 대신 자동차 뒷좌석 쿠션 위를 걷고 있음)"
      ],
      "physics": "사람들이 거리가 아닌 거대한 자동차 뒷좌석 쿠션 위를 걷고 있으며, 공간과 사람 간의 크기 비율이 완전히 무너진 채 시트 위에 물리적으로 놓여 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 8,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S90sh5__bgfirst_bg.png",
   "bg_asset_id": "20456562-9927-464e-95b1-4f4942b69faf",
   "bg_record_key": "S90sh5::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S90sh5::cine": {
  "applied": true,
  "fingerprint": "3f4a642c61bb36f164da8d4822c853545332c405d2c6ae76afc97a62f5ec6899",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S90sh5_sel.png",
  "source_sha256": "559a7e2cf9585d542ccdd796755a4315442c049f9c9448a035b31ab4ac42e93e",
  "file": "S90sh5_cine.png",
  "latency_ms": 12825
 },
 "S90sh7::confined_fp_apt": {
  "applies": false,
  "reason_ko": "단순한 자동차 뒷좌석 탑승 장면으로, 복잡한 제어 장치나 좌석 배치가 중요한 공간이 아니며 방향성이 엄격히 요구되는 객체도 없으므로 평면도 배치의 도움이 필요하지 않습니다.",
  "input_fingerprint": "6542c56321bf85b8"
 },
 "S90sh7::signage": {
  "fp": "d6ec83b41034065d",
  "inscriptions": [
   {
    "surface_native": "집회 참가자가 들고 있는 피켓",
    "text_native": "진실을 밝혀라",
    "reason_ko": "촛불집회 현장의 분위기와 인물의 슬픈 감정을 뒷받침하기 위해 창밖 군중이 들고 있는 시위 피켓의 문구가 필요합니다."
   }
  ]
 },
 "era_assess::4b7812f60cb6a2cf": {
  "subjects": [
   {
    "subject_native": "2000년대~2010년대 한국의 촛불집회 풍경",
    "search_terms_native": [
     "촛불집회 종이컵",
     "광화문 촛불시위",
     "촛불문화제"
    ],
    "language_lock_native": "모든 검색어는 반드시 한국어로만 작성해야 하며, 다른 언어로 번역하거나 추가하지 마십시오.",
    "reason_ko": "한국의 촛불집회는 종이컵에 끼운 양초나 LED 촛불을 든 군중의 독특한 시각적 특징이 있어 일반적인 서양식 촛불 행사와 다릅니다."
   },
   {
    "subject_native": "2000년대 한국 자동차 뒷좌석 내부",
    "search_terms_native": [
     "2000년대 국산차 내부",
     "EF쏘나타 뒷좌석",
     "그랜저 내부 시트"
    ],
    "language_lock_native": "검색은 반드시 한국어로만 수행되어야 하며, 영어를 포함한 다른 언어로 번역해서는 안 됩니다.",
    "reason_ko": "2000년대 초반 한국 중형 세단의 시트 재질, 가죽 도어 트림, 윈도우 버튼 등의 구성은 서구식 차량 인테리어와 확연히 다릅니다."
   }
  ]
 },
 "era_ref::fe576d50b182c2d2": {
  "subject": "2000년대~2010년대 한국의 촛불집회 풍경",
  "terms": [
   "촛불집회 종이컵",
   "광화문 촛불시위",
   "촛불문화제"
  ],
  "queries": [
   [
    "2000년대 2010년대 한국 촛불집회 종이컵 광화문 촛불시위 촛불문화제",
    "광화문 촛불집회 종이컵 촛불문화제 시민 풍경"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://www.popco.net/zboard/data2/com_freeboard/2016/12/11/1018889730584c286cb917b.JPG",
  "picked_reason_ko": "촛불집회의 핵심인 야간 도심의 대규모 군중과 촛불 물결을 가장 선명하고 대표적으로 보여주는 사진이다.",
  "sha256": "92fa72355418e69101177cfe2a91644edce26555a9ca0ff454b516f56f5be3ea",
  "file": "eraref_fe576d50b182c2d2.png"
 },
 "S90sh7::bgfirst_bg": {
  "input_fingerprint": "a825d1df73308d55",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차창 밖의 촛불들을 물끄러미 응시하며 한 손등이 눈물 젖은 볼에 맞닿아 있는 심옥의 옆얼굴\n\nLOCATION (lock): Inside the car's rear passenger compartment, beside the window overlooking the candlelight gathering.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From inside the rear compartment on 민정 (현재)'s side, finish the slow dolly-in slightly below 심옥's eye level at a close three-quarter profile. 심옥 occupies one side of the frame with restrained look room toward the side window, her attention fixed outside as the back of one hand rests against her tear-wet cheek and the candle procession stays softly legible beyond the glass.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: rear side window (beside 심옥 in the moving car) — The candle-bearing people pass beyond the glass along 심옥's line of sight; used as Carries 심옥's outward sightline and separates her from the passing procession; candle procession (passing slowly outside); used as Soft moving background retained at the edge of the close framing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk illumination, sunset reflection, and the candles outside shape a restrained warm edge against the subdued car interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 촛불집회 풍경: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 차창 밖의 촛불들을 물끄러미 응시하며 한 손등이 눈물 젖은 볼에 맞닿아 있는 심옥의 옆얼굴\n\nLOCATION (lock): Inside the car's rear passenger compartment, beside the window overlooking the candlelight gathering.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: From inside the rear compartment on 민정 (현재)'s side, finish the slow dolly-in slightly below 심옥's eye level at a close three-quarter profile. 심옥 occupies one side of the frame with restrained look room toward the side window, her attention fixed outside as the back of one hand rests against her tear-wet cheek and the candle procession stays softly legible beyond the glass.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: rear side window (beside 심옥 in the moving car) — The candle-bearing people pass beyond the glass along 심옥's line of sight; used as Carries 심옥's outward sightline and separates her from the passing procession; candle procession (passing slowly outside); used as Soft moving background retained at the edge of the close framing.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk illumination, sunset reflection, and the candles outside shape a restrained warm edge against the subdued car interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대~2010년대 한국의 촛불집회 풍경: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S90sh7__bgfirst_bg.png",
  "asset_id": "27b5e480-9cac-4cc1-a84d-c6b220dbbcb9",
  "input_asset_ids": [
   "37baac42-2abe-41d6-82fd-a5b37c8a07dd",
   "0c1cefd9-6767-4f28-9c75-05b0eb689eae"
  ],
  "era_research": {
   "subject": "2000년대~2010년대 한국의 촛불집회 풍경",
   "queries": [
    [
     "2000년대 2010년대 한국 촛불집회 종이컵 광화문 촛불시위 촛불문화제",
     "광화문 촛불집회 종이컵 촛불문화제 시민 풍경"
    ]
   ],
   "picked_url": "https://www.popco.net/zboard/data2/com_freeboard/2016/12/11/1018889730584c286cb917b.JPG",
   "sha256": "92fa72355418e69101177cfe2a91644edce26555a9ca0ff454b516f56f5be3ea",
   "file": "eraref_fe576d50b182c2d2.png"
  }
 },
 "S90sh7": {
  "input_fingerprint": "6083aefecd24f0bd",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 차창 밖의 촛불들을 물끄러미 응시하며 한 손등이 눈물 젖은 볼에 맞닿아 있는 심옥의 옆얼굴\n\nLOCATION (lock): Inside the car's rear passenger compartment, beside the window overlooking the candlelight gathering. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From inside the rear compartment on 민정 (현재)'s side, finish the slow dolly-in slightly below 심옥's eye level at a close three-quarter profile. 심옥 occupies one side of the frame with restrained look room toward the side window, her attention fixed outside as the back of one hand rests against her tear-wet cheek and the candle procession stays softly legible beyond the glass.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: rear side window (beside 심옥 in the moving car) — The candle-bearing people pass beyond the glass along 심옥's line of sight; used as Carries 심옥's outward sightline and separates her from the passing procession; candle procession (passing slowly outside); used as Soft moving background retained at the edge of the close framing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk illumination, sunset reflection, and the candles outside shape a restrained warm edge against the subdued car interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candlelit crowd remains visible outside the moving car as Sim-ok wipes tears from her cheek.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 집회 참가자가 들고 있는 피켓: \"진실을 밝혀라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot inside a tight, built interior. The FIRST attached image (SHOT BACKGROUND) is the finished empty interior of this shot, and in a space this cramped its geometry is the truth of the shot — keep it EXACTLY: its camera, perspective, every panel, control, seat, mirror, window and fixture stay untouched, in the same place, at the same angle, in the same number.\n\nBefore you place anyone, count what the background shows: how many steering wheels or control surfaces, how many seats and which way each faces, where each mirror sits and what it could reflect from this camera. Those counts and placements are what you must still be able to make after the people are in. Adding a second rim, sliding a seat, turning a mirror or growing a new panel is a failure even when the person looks right.\n\nThe SECOND attached image (LAYOUT SKETCH) tells you where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Its background lines are not decoration — they are the same structure seen in line form, so use them to place each person correctly with respect to it: which seat the body occupies, which side of the wheel the hands are on, what the body passes in front of and what it passes behind. Where sketch and background disagree about the structure itself, the background wins.\n\nThe CHARACTER REFERENCE photographs show the real people.\n\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. A person may cover part of the structure — that is expected, and covering is not redrawing. What the body hides stays hidden; what remains visible stays exactly as the background had it. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 차창 밖의 촛불들을 물끄러미 응시하며 한 손등이 눈물 젖은 볼에 맞닿아 있는 심옥의 옆얼굴\n\nLOCATION (lock): Inside the car's rear passenger compartment, beside the window overlooking the candlelight gathering. The shot takes place here — the FIRST attached image (SHOT BACKGROUND) is this exact place, already built: its ground, structures, horizon, materials and lighting are the finished truth of this location and must not be redesigned or replaced. No location photograph is attached — read the place from that image alone, and add no scenery, structure, vehicle or fixture that it does not already show. This lock governs the place only; the figures in the shot follow the staging and pose instructions.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From inside the rear compartment on 민정 (현재)'s side, finish the slow dolly-in slightly below 심옥's eye level at a close three-quarter profile. 심옥 occupies one side of the frame with restrained look room toward the side window, her attention fixed outside as the back of one hand rests against her tear-wet cheek and the candle procession stays softly legible beyond the glass.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: rear side window (beside 심옥 in the moving car) — The candle-bearing people pass beyond the glass along 심옥's line of sight; used as Carries 심옥's outward sightline and separates her from the passing procession; candle procession (passing slowly outside); used as Soft moving background retained at the edge of the close framing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk illumination, sunset reflection, and the candles outside shape a restrained warm edge against the subdued car interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candlelit crowd remains visible outside the moving car as Sim-ok wipes tears from her cheek.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 차창 밖의 촛불들을 물끄러미 응시하며 한 손등이 눈물 젖은 볼에 맞닿아 있는 심옥의 옆얼굴\n\nLOCATION (lock): Inside the car's rear passenger compartment, beside the window overlooking the candlelight gathering. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From inside the rear compartment on 민정 (현재)'s side, finish the slow dolly-in slightly below 심옥's eye level at a close three-quarter profile. 심옥 occupies one side of the frame with restrained look room toward the side window, her attention fixed outside as the back of one hand rests against her tear-wet cheek and the candle procession stays softly legible beyond the glass.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: rear side window (beside 심옥 in the moving car) — The candle-bearing people pass beyond the glass along 심옥's line of sight; used as Carries 심옥's outward sightline and separates her from the passing procession; candle procession (passing slowly outside); used as Soft moving background retained at the edge of the close framing.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk illumination, sunset reflection, and the candles outside shape a restrained warm edge against the subdued car interior.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The candlelit crowd remains visible outside the moving car as Sim-ok wipes tears from her cheek.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 집회 참가자가 들고 있는 피켓: \"진실을 밝혀라\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S90sh7__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement AND the structure they sit inside — the sketched panels, seats, controls and openings are the same ones the background photograph shows, drawn as lines; read them to place each body correctly against that structure, and where the two disagree about the structure itself the background photograph wins)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S90sh7.png"
    },
    {
     "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:866898>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L10B01.png"
    },
    {
     "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:866898>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "필수 지시사항인 피켓 텍스트(\"진실을 밝혀라\")를 정확히 렌더링하고 인물과 배경의 시네마틱한 대비를 잘 살렸으나, 손등이 뺨에 닿는 디테일은 다소 부정확함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "필수적인 피켓 텍스트가 완전히 누락되었으며, 요구된 손 자세(손등이 뺨에 닿음)도 정확하게 구현되지 않음."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "심옥의 시선이 창밖의 피켓과 촛불을 든 군중을 향함.",
      "built_space": "자동차 뒷좌석 내부의 창틀, 유리창, 시트가 올바른 위치에 표현됨.",
      "entities": "심옥의 외모와 의상이 레퍼런스와 일치하며, 창밖에 \"진실을 밝혀라\"라고 적힌 피켓이 정확하게 나타남.",
      "hard_violations": [],
      "physics": "손이 얼굴을 자연스럽게 받치고 있으나, 프롬프트에 명시된 '손등이 닿는' 자세가 아닌 손가락 측면이 닿아 있음."
     },
     {
      "label": "A",
      "direction": "심옥의 시선이 창밖 멀리 있는 촛불 행렬을 향함.",
      "built_space": "자동차 뒷좌석 내부 구조가 인물과 올바른 비례로 배치됨.",
      "entities": "심옥의 캐릭터 레퍼런스는 잘 반영되었으나, 프롬프트가 요구한 피켓 및 텍스트가 전혀 존재하지 않음.",
      "hard_violations": [],
      "physics": "손이 뺨을 만지고 있으나 손등이 아닌 손바닥과 손가락 안쪽이 닿아 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "필수 지시사항인 피켓 텍스트(\"진실을 밝혀라\")를 정확히 렌더링하고 인물과 배경의 시네마틱한 대비를 잘 살렸으나, 손등이 뺨에 닿는 디테일은 다소 부정확함."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "필수적인 피켓 텍스트가 완전히 누락되었으며, 요구된 손 자세(손등이 뺨에 닿음)도 정확하게 구현되지 않음."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "심옥의 시선이 창밖의 피켓과 촛불을 든 군중을 향함.",
      "built_space": "자동차 뒷좌석 내부의 창틀, 유리창, 시트가 올바른 위치에 표현됨.",
      "entities": "심옥의 외모와 의상이 레퍼런스와 일치하며, 창밖에 \"진실을 밝혀라\"라고 적힌 피켓이 정확하게 나타남.",
      "hard_violations": [],
      "physics": "손이 얼굴을 자연스럽게 받치고 있으나, 프롬프트에 명시된 '손등이 닿는' 자세가 아닌 손가락 측면이 닿아 있음."
     },
     {
      "label": "A",
      "direction": "심옥의 시선이 창밖 멀리 있는 촛불 행렬을 향함.",
      "built_space": "자동차 뒷좌석 내부 구조가 인물과 올바른 비례로 배치됨.",
      "entities": "심옥의 캐릭터 레퍼런스는 잘 반영되었으나, 프롬프트가 요구한 피켓 및 텍스트가 전혀 존재하지 않음.",
      "hard_violations": [],
      "physics": "손이 뺨을 만지고 있으나 손등이 아닌 손바닥과 손가락 안쪽이 닿아 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "불가능한 구도 왜곡(우측 뺨 노출)의 치명적 오류가 있으나, 필수 지시사항인 피켓의 텍스트를 완벽하게 구현하여 B보다 낫습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "불가능한 구도 왜곡 오류를 공유하며, 핵심인 피켓 텍스트가 완전히 누락되어 우선순위에서 밀립니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "심옥의 시선은 창밖 촛불 집회와 피켓을 향함.",
      "built_space": "차량 뒷좌석. 인물 왼쪽에서 창밖을 향하는 카메라 위치이나, 우측 뺨이 보이는 불가능한 공간 왜곡 발생.",
      "entities": "심옥(외모 일치), '진실을 밝혀라' 피켓(완벽 일치). 단, 손등이 아닌 손바닥/손가락이 뺨에 닿음.",
      "hard_violations": [
       "physically impossible staging: 카메라가 차량 내부(인물 왼쪽)에서 우측 창밖을 향하는데 인물의 우측 뺨이 보이는 불가능한 구도"
      ],
      "physics": "좌석이 몸을, 팔이 손을 자연스럽게 지탱함."
     },
     {
      "label": "B",
      "direction": "심옥의 시선은 창밖 촛불 행렬을 향함.",
      "built_space": "차량 뒷좌석. A와 동일하게 카메라 위치상 보일 수 없는 우측 뺨이 노출된 공간 왜곡 발생.",
      "entities": "심옥(외모 일치). 필수 요소인 피켓 및 텍스트 전체 누락. 손바닥이 뺨에 닿음.",
      "hard_violations": [
       "physically impossible staging: 인물 왼쪽에서 우측 창밖을 보는 시점임에도 인물의 우측 뺨이 렌더링된 불가능한 구도"
      ],
      "physics": "좌석이 몸을, 팔이 손을 지탱함."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 4,
      "verdict_ko": "불가능한 구도 왜곡(우측 뺨 노출)의 치명적 오류가 있으나, 필수 지시사항인 피켓의 텍스트를 완벽하게 구현하여 B보다 낫습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "불가능한 구도 왜곡 오류를 공유하며, 핵심인 피켓 텍스트가 완전히 누락되어 우선순위에서 밀립니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "B",
      "direction": "심옥의 시선은 창밖 촛불 집회와 피켓을 향함.",
      "built_space": "차량 뒷좌석. 인물 왼쪽에서 창밖을 향하는 카메라 위치이나, 우측 뺨이 보이는 불가능한 공간 왜곡 발생.",
      "entities": "심옥(외모 일치), '진실을 밝혀라' 피켓(완벽 일치). 단, 손등이 아닌 손바닥/손가락이 뺨에 닿음.",
      "hard_violations": [
       "physically impossible staging: 카메라가 차량 내부(인물 왼쪽)에서 우측 창밖을 향하는데 인물의 우측 뺨이 보이는 불가능한 구도"
      ],
      "physics": "좌석이 몸을, 팔이 손을 자연스럽게 지탱함."
     },
     {
      "label": "A",
      "direction": "심옥의 시선은 창밖 촛불 행렬을 향함.",
      "built_space": "차량 뒷좌석. A와 동일하게 카메라 위치상 보일 수 없는 우측 뺨이 노출된 공간 왜곡 발생.",
      "entities": "심옥(외모 일치). 필수 요소인 피켓 및 텍스트 전체 누락. 손바닥이 뺨에 닿음.",
      "hard_violations": [
       "physically impossible staging: 인물 왼쪽에서 우측 창밖을 보는 시점임에도 인물의 우측 뺨이 렌더링된 불가능한 구도"
      ],
      "physics": "좌석이 몸을, 팔이 손을 지탱함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 11
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "B",
    "direction": "심옥의 시선이 창밖의 피켓과 촛불을 든 군중을 향함.",
    "built_space": "자동차 뒷좌석 내부의 창틀, 유리창, 시트가 올바른 위치에 표현됨.",
    "entities": "심옥의 외모와 의상이 레퍼런스와 일치하며, 창밖에 \"진실을 밝혀라\"라고 적힌 피켓이 정확하게 나타남.",
    "hard_violations": [],
    "physics": "손이 얼굴을 자연스럽게 받치고 있으나, 프롬프트에 명시된 '손등이 닿는' 자세가 아닌 손가락 측면이 닿아 있음."
   },
   {
    "label": "A",
    "direction": "심옥의 시선이 창밖 멀리 있는 촛불 행렬을 향함.",
    "built_space": "자동차 뒷좌석 내부 구조가 인물과 올바른 비례로 배치됨.",
    "entities": "심옥의 캐릭터 레퍼런스는 잘 반영되었으나, 프롬프트가 요구한 피켓 및 텍스트가 전혀 존재하지 않음.",
    "hard_violations": [],
    "physics": "손이 뺨을 만지고 있으나 손등이 아닌 손바닥과 손가락 안쪽이 닿아 있음."
   }
  ],
  "totals": {
   "A": 7,
   "B": 11
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "필수 지시사항인 피켓 텍스트(\"진실을 밝혀라\")를 정확히 렌더링하고 인물과 배경의 시네마틱한 대비를 잘 살렸으나, 손등이 뺨에 닿는 디테일은 다소 부정확함."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "필수적인 피켓 텍스트가 완전히 누락되었으며, 요구된 손 자세(손등이 뺨에 닿음)도 정확하게 구현되지 않음."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/episodes/f5372927-bbec-405d-ad2c-100d587f5373/images/background_chain/L10B01.png"
   },
   {
    "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:866898>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트는 '손등'이 볼에 닿아야 한다고 명시했으나, 사진에서는 손바닥과 손가락 안쪽이 볼에 닿아 있습니다.",
     "fix_en": "Change the hand pose so that the back of the hand rests against the cheek, rather than the palm and fingers. Keep the woman's face, expression, tears, clothing, car interior, and the background crowd exactly the same.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "창밖의 촛불 행렬이 자연스럽게 지나가는 모습이 아니라, 참가자들이 카메라(창문)를 향해 일렬로 나란히 서서 정면을 향해 정적인 포즈를 취하고 있습니다.",
     "fix_en": "Redraw the crowd outside the window so they are walking past in profile or semi-profile as a passing procession, rather than standing still and facing the car directly. Maintain the woman, her pose, tears, lighting, and the car interior unchanged.",
     "severity": "major",
     "observation_index": 1
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트는 '손등'이 볼에 닿아야 한다고 명시했으나, 사진에서는 손바닥과 손가락 안쪽이 볼에 닿아 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "창밖의 촛불 행렬이 자연스럽게 지나가는 모습이 아니라, 참가자들이 카메라(창문)를 향해 일렬로 나란히 서서 정면을 향해 정적인 포즈를 취하고 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "심옥의 손이 손등이 아니라 손바닥·손가락이 볼에 닿아 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "창밖 집회가 멀리 지나가는 행렬이 아니라 차창에 바짝 붙어 정면으로 서 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "프롬프트는 '손등'이 볼에 닿아야 한다고 명시했으나, 사진에서는 손바닥과 손가락 안쪽이 볼에 닿아 있습니다.",
    "fix_en": "Change the hand pose so that the back of the hand rests against the cheek, rather than the palm and fingers. Keep the woman's face, expression, tears, clothing, car interior, and the background crowd exactly the same.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "창밖의 촛불 행렬이 자연스럽게 지나가는 모습이 아니라, 참가자들이 카메라(창문)를 향해 일렬로 나란히 서서 정면을 향해 정적인 포즈를 취하고 있습니다.",
    "fix_en": "Redraw the crowd outside the window so they are walking past in profile or semi-profile as a passing procession, rather than standing still and facing the car directly. Maintain the woman, her pose, tears, lighting, and the car interior unchanged.",
    "severity": "major",
    "observation_index": 1
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S90sh7__bgfirst_bg.png",
   "bg_asset_id": "27b5e480-9cac-4cc1-a84d-c6b220dbbcb9",
   "bg_record_key": "S90sh7::bgfirst_bg",
   "chain_winner": false,
   "authority": "plate"
  },
  "ref_mode": "플레이트+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S90sh5"
  }
 },
 "S90sh7::cine": {
  "applied": true,
  "fingerprint": "9d0e3959d2e051c3545ae88a36b42c9cfafcb59ede97e5031abe7518dd16592a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S90sh7_sel.png",
  "source_sha256": "0e9ded4ddc3a521ec2f10dde9313c8809ec91d62008cb2ca39bb59742e4004c7",
  "file": "S90sh7_cine.png",
  "latency_ms": 13374
 },
 "S91sh3::signage": {
  "fp": "cdd9b6921da8aec7",
  "inscriptions": []
 },
 "era_assess::805f34f5ad8389d8": {
  "subjects": [
   {
    "subject_native": "2000년대 한국의 납골당 및 추모공원 야외 주차장",
    "search_terms_native": [
     "납골당 외관 주차장",
     "봉안당 야외 주차장",
     "추모공원 주차장"
    ],
    "language_lock_native": "검색 시 반드시 한국어 검색어만을 사용하고, 다른 언어의 단어를 혼용하거나 번역하여 검색하지 마십시오.",
    "reason_ko": "2000년대 한국의 추모공원 및 납골당은 특유의 산악 지형 결합형 배치, 석조 외관, 그리고 한국식 주차 구획선과 조경 스타일을 띠고 있어, 일반적인 공동묘지나 서구식 주차장과 구분되는 고유의 시각적 특징이 있습니다."
   }
  ]
 },
 "era_ref::ca1aad082a069e28": {
  "subject": "2000년대 한국의 납골당 및 추모공원 야외 주차장",
  "terms": [
   "납골당 외관 주차장",
   "봉안당 야외 주차장",
   "추모공원 주차장"
  ],
  "queries": [
   [
    "2000년대 한국 납골당 외관 야외 주차장",
    "2000년대 한국 봉안당 추모공원 야외 주차장"
   ]
  ],
  "candidates": 4,
  "picked_index": 1,
  "picked_url": "https://www.chungahpark.co.kr/img/park/1409613991067_287.%EB%B9%84%EC%A3%BC%EC%B0%A8%EC%9E%A5.jpg",
  "picked_reason_ko": "후보 1은 2000년대 한국 추모시설의 평범한 야외 주차장을 가장 넓고 명료하게 보여 주어 포장, 주차 구획, 수목 배치와 시설 건물의 관계를 읽기 좋은 참고 사진이다.",
  "sha256": "ee07e510884b6b19929f1e996cec4c6455d1d142e3f01409c8a6db1ebf915859",
  "file": "eraref_ca1aad082a069e28.png"
 },
 "groupbg::납골시설 주차장": {
  "input_fingerprint": "dd13010fc705e311",
  "meta": {
   "model": "gpt-image-2",
   "size": "1536x864",
   "pack": "11.202607220237",
   "contract": "bgfirst_full_v3",
   "group_sig": {
    "key": "납골시설 주차장",
    "tags": [
     "S91sh3"
    ]
   },
   "context_sig": "043e27696c154a3d",
   "era_research_sha": "ee07e510884b6b19929f1e996cec4c6455d1d142e3f01409c8a6db1ebf915859"
  },
  "prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: In the columbarium's exterior parking area, near the path leading to the building entrance.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n납골당 앞 주차장: 붉은 저녁노을이 짙게 깔려 차분한 분위기의 추모 시설 야외 주차 공간이다. (특징: 포장된 주차 구획선 위로 길게 늘어지는 그림자; 하늘 전체를 물들인 붉고 주황빛 도는 노을빛)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 납골당 앞 주차장 - 해질녘\n- 납골당으로 걸어가는 심옥과 민정을 배웅하는 택수.\n\nTIME OF DAY (lock): sunset.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 납골당 및 추모공원 야외 주차장: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create ONE empty live-action location background photograph — NO PEOPLE, no figures, no body parts, no silhouettes, no shadows or reflections of people anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods and produce in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch of a shot that happens at this location: use it ONLY as spatial evidence — what this place contains, how its ground, structures and landmarks are arranged and proportioned. Ignore the sketched people and arrows entirely, and do NOT copy its line style: render a fully photographic, physically plausible real place that fits THE LOCATION text below.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nTHE LOCATION — South Korea, 2001–2017; all people are Korean unless stated: In the columbarium's exterior parking area, near the path leading to the building entrance.\n\nLOCATION DETAIL (descriptive evidence about this place from the production's location records — evidence, not a staging order): use it to understand what kind of place this is and what it permanently contains. Carry over ONLY its enduring physical features — layout, structures, ground and materials, fixed equipment and installations. Do NOT treat any lighting, weather or time-of-day wording, momentary object states, depicted events or subjective impressions in it as fixed facts of the place; the TIME OF DAY lock below is the sole authority for time and lighting.\n납골당 앞 주차장: 붉은 저녁노을이 짙게 깔려 차분한 분위기의 추모 시설 야외 주차 공간이다. (특징: 포장된 주차 구획선 위로 길게 늘어지는 그림자; 하늘 전체를 물들인 붉고 주황빛 도는 노을빛)\n\nSCENE EVIDENCE (verbatim quotes from the screenplay about this place — treat them as evidence of what the location physically contains and looks like; stage the PLACE those moments happen in, but do NOT depict the momentary actions, people or staged props themselves):\n- 납골당 앞 주차장 - 해질녘\n- 납골당으로 걸어가는 심옥과 민정을 배웅하는 택수.\n\nTIME OF DAY (lock): sunset.\n\nRender ONE photorealistic empty location photograph, 16:9, neutral enough that every shot of this place can be staged from it later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 납골당 및 추모공원 야외 주차장: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_납골시설_주차장_210875.png",
  "asset_id": "5f540a15-e154-47a9-b2e5-24c7c2b9b95b",
  "input_asset_ids": [
   "6e0a06ee-76c4-406f-9569-79eaf0f59559"
  ],
  "origin_tag": "S91sh3",
  "place_text": "In the columbarium's exterior parking area, near the path leading to the building entrance.",
  "origin_inputs": {
   "place_text": "In the columbarium's exterior parking area, near the path leading to the building entrance.",
   "time_of_day_en": "sunset",
   "conti_asset_id": "6e0a06ee-76c4-406f-9569-79eaf0f59559"
  },
  "era_research": {
   "subject": "2000년대 한국의 납골당 및 추모공원 야외 주차장",
   "terms": [
    "납골당 외관 주차장",
    "봉안당 야외 주차장",
    "추모공원 주차장"
   ],
   "queries": [
    [
     "2000년대 한국 납골당 외관 야외 주차장",
     "2000년대 한국 봉안당 추모공원 야외 주차장"
    ]
   ],
   "candidates": 4,
   "picked_index": 1,
   "picked_url": "https://www.chungahpark.co.kr/img/park/1409613991067_287.%EB%B9%84%EC%A3%BC%EC%B0%A8%EC%9E%A5.jpg",
   "picked_reason_ko": "후보 1은 2000년대 한국 추모시설의 평범한 야외 주차장을 가장 넓고 명료하게 보여 주어 포장, 주차 구획, 수목 배치와 시설 건물의 관계를 읽기 좋은 참고 사진이다.",
   "sha256": "ee07e510884b6b19929f1e996cec4c6455d1d142e3f01409c8a6db1ebf915859",
   "file": "eraref_ca1aad082a069e28.png",
   "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/eraref_ca1aad082a069e28.png"
  }
 },
 "era_assess::317b39deed90c64f": {
  "subjects": [
   {
    "subject_native": "2000년대 한국의 봉안당(납골당) 건물 외관 및 주차장",
    "search_terms_native": [
     "봉안당 외관",
     "납골당 주차장",
     "추모공원 건물 입구",
     "한국 납골당"
    ],
    "language_lock_native": "검색어는 반드시 한국어로만 작성해야 하며 영어나 다른 언어로 번역하거나 혼용하지 마십시오.",
    "reason_ko": "한국의 봉안당(납골당)은 서구식 납골당이나 일반 현대식 건물과 달리 특유의 석조 건축 양식, 단정한 조경, 특유의 입구 표지판 및 주차장 양식을 가지고 있어 고유의 실물 참조가 필요합니다."
   }
  ]
 },
 "era_ref::6078b523bc0daa99": {
  "subject": "2000년대 한국의 봉안당(납골당) 건물 외관 및 주차장",
  "terms": [
   "봉안당 외관",
   "납골당 주차장",
   "추모공원 건물 입구",
   "한국 납골당"
  ],
  "queries": [
   [
    "2000년대 한국 봉안당 납골당 건물 외관 주차장",
    "한국 추모공원 납골당 건물 입구 주차장"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://www.goifuneral.co.kr/_next/image/?q=75&url=https%3A%2F%2Fwww.15774129.go.kr%2FBCUser%2Ffacilitypic%2F1346833912142_7000000631_0.jpg&w=3840",
  "picked_reason_ko": "주차장 전체는 보이지 않아도 선택에 지장이 없으며, 봉안당 본체의 일상적인 외관과 재료·구조를 가장 깨끗하고 명확하게 확인할 수 있다.",
  "sha256": "fbe1b48636475f2dc5c30f07d4cefda72fb0f83a2630cd4ae7d01075fe65ad96",
  "file": "eraref_6078b523bc0daa99.png"
 },
 "S91sh3::bgfirst_bg": {
  "input_fingerprint": "e5246c71c54718f1",
  "prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 전택수의 손을 위아래로 단단하게 맞잡은 심옥과 민정의 포개진 손 클로즈업\n\nLOCATION (lock): In the columbarium's exterior parking area, near the path leading to the building entrance.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold statically just above hand height and slightly to the side, looking down obliquely at the three layered hands in an isolated insert. 전택수's hand lies between 심옥's firm upper clasp and 민정 (현재)'s supporting hand below, with only partial forearms and a small amount of parking-area context left around the gesture.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 심옥 hand enclosing from above in the middle-center of the frame, foreground; 전택수 hand held in the middle in the middle-center of the frame, foreground; 민정 hand supporting from below in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: ossuary parking area (at dusk); used as Minimal contextual plane around the joined hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk light renders the layered hands with subdued contrast and a restrained, humane tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 봉안당(납골당) 건물 외관 및 주차장: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "effective_prompt": "Create the EMPTY BACKGROUND PLATE for one film shot — NO PEOPLE, no figures, no body parts, no sketch lines, no arrows anywhere.\n\"Empty\" means no people only: KEEP the location's inherent occupants and stock that define the place — animals in an animal shelter, pen or farm, goods in a market, moored boats in a harbour — unless the shot text explicitly removes them.\nThe FIRST attached image is a thin-line storyboard sketch: use ONLY its camera angle, horizon, perspective and the placement/size of buildings and set masses — ignore the sketched people and arrows entirely. The SECOND attached image (LOCATION PHOTOGRAPH) is the real place: take its architecture, materials, signage and fixed features, and RE-PROJECT them into the sketch's camera. If the photograph's camera differs from the sketch's, the sketch's camera wins.\nHUMAN-SCALE CALIBRATION: derive every structure's true size from human-scale elements — a door ≈ 2m, a window ≈ 1–1.5m wide, one storey ≈ 2.5–3m; never inflate a small structure or shrink a large one.\n\nSHOT TEXT this background must serve (Korean): 전택수의 손을 위아래로 단단하게 맞잡은 심옥과 민정의 포개진 손 클로즈업\n\nLOCATION (lock): In the columbarium's exterior parking area, near the path leading to the building entrance.\n\nTIME OF DAY (lock): sunset.\n\nCAMERA & FRAME (follow exactly — this is the composition authority for this still; the reference images supply identity and place, never the framing):\n- CAMERA: Hold statically just above hand height and slightly to the side, looking down obliquely at the three layered hands in an isolated insert. 전택수's hand lies between 심옥's firm upper clasp and 민정 (현재)'s supporting hand below, with only partial forearms and a small amount of parking-area context left around the gesture.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 심옥 hand enclosing from above in the middle-center of the frame, foreground; 전택수 hand held in the middle in the middle-center of the frame, foreground; 민정 hand supporting from below in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: ossuary parking area (at dusk); used as Minimal contextual plane around the joined hands.\nCompose the frame exactly as specified above — camera angle, subject scale and screen placement. Keep true physical scale between people and background elements: fixtures and distant objects occupy only the small screen area their real size and distance dictate; never enlarge a background object into a foreground presence.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk light renders the layered hands with subdued contrast and a restrained, humane tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nRender ONE photorealistic empty location photograph, 16:9, that this shot can be staged inside later. Signage and other writing that permanently belongs to this place may appear, in the place's native language and period style; no captions, watermarks or overlay text.\n\nPERIOD REFERENCE — 2000년대 한국의 봉안당(납골당) 건물 외관 및 주차장: a researched photograph of the real\nthing. Copy its era-accurate form, proportions, materials, fittings\nand styling exactly. It decides ONLY what this thing looks like —\nnever the composition, framing, lighting, people or their placement.",
  "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S91sh3__bgfirst_bg.png",
  "asset_id": "f74ffde5-de75-489d-ac6a-5cbe021113d4",
  "input_asset_ids": [
   "6e0a06ee-76c4-406f-9569-79eaf0f59559",
   "5f540a15-e154-47a9-b2e5-24c7c2b9b95b"
  ],
  "era_research": {
   "subject": "2000년대 한국의 봉안당(납골당) 건물 외관 및 주차장",
   "queries": [
    [
     "2000년대 한국 봉안당 납골당 건물 외관 주차장",
     "한국 추모공원 납골당 건물 입구 주차장"
    ]
   ],
   "picked_url": "https://www.goifuneral.co.kr/_next/image/?q=75&url=https%3A%2F%2Fwww.15774129.go.kr%2FBCUser%2Ffacilitypic%2F1346833912142_7000000631_0.jpg&w=3840",
   "sha256": "fbe1b48636475f2dc5c30f07d4cefda72fb0f83a2630cd4ae7d01075fe65ad96",
   "file": "eraref_6078b523bc0daa99.png"
  }
 },
 "S91sh3": {
  "input_fingerprint": "16a07b79d5200633",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 전택수의 손을 위아래로 단단하게 맞잡은 심옥과 민정의 포개진 손 클로즈업\n\nLOCATION (lock): In the columbarium's exterior parking area, near the path leading to the building entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Hold statically just above hand height and slightly to the side, looking down obliquely at the three layered hands in an isolated insert. 전택수's hand lies between 심옥's firm upper clasp and 민정 (현재)'s supporting hand below, with only partial forearms and a small amount of parking-area context left around the gesture.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 심옥 hand enclosing from above in the middle-center of the frame, foreground; 전택수 hand held in the middle in the middle-center of the frame, foreground; 민정 hand supporting from below in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: ossuary parking area (at dusk); used as Minimal contextual plane around the joined hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk light renders the layered hands with subdued contrast and a restrained, humane tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sim-ok and Minjeong clasp Taeksu's hand firmly together in response to his apology.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리); 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리); 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Stage the shot. The FIRST attached image (SHOT BACKGROUND) is the finished empty background of this shot — keep it EXACTLY: its camera, perspective, architecture, lighting and every fixed feature stay untouched. The SECOND attached image (LAYOUT SKETCH) tells you ONLY where the people go: each sketched person's position, screen size, pose and the gaze/motion arrows. Ignore the sketch's background lines. The CHARACTER REFERENCE photographs show the real people.\nPlace the real people into the background at exactly the sketched positions, sizes and poses, following the arrow directions. No sketch lines or arrows may remain.\n\nCreate ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 전택수의 손을 위아래로 단단하게 맞잡은 심옥과 민정의 포개진 손 클로즈업\n\nLOCATION (lock): In the columbarium's exterior parking area, near the path leading to the building entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Hold statically just above hand height and slightly to the side, looking down obliquely at the three layered hands in an isolated insert. 전택수's hand lies between 심옥's firm upper clasp and 민정 (현재)'s supporting hand below, with only partial forearms and a small amount of parking-area context left around the gesture.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 심옥 hand enclosing from above in the middle-center of the frame, foreground; 전택수 hand held in the middle in the middle-center of the frame, foreground; 민정 hand supporting from below in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: ossuary parking area (at dusk); used as Minimal contextual plane around the joined hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk light renders the layered hands with subdued contrast and a restrained, humane tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sim-ok and Minjeong clasp Taeksu's hand firmly together in response to his apology.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리); 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리); 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 전택수의 손을 위아래로 단단하게 맞잡은 심옥과 민정의 포개진 손 클로즈업\n\nLOCATION (lock): In the columbarium's exterior parking area, near the path leading to the building entrance. The shot takes place here — the attached LOCATION PHOTOGRAPH shows the exact spot.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Hold statically just above hand height and slightly to the side, looking down obliquely at the three layered hands in an isolated insert. 전택수's hand lies between 심옥's firm upper clasp and 민정 (현재)'s supporting hand below, with only partial forearms and a small amount of parking-area context left around the gesture.\n- FRAMING SCALE: insert close-up on a detail\n- FRAME LAYOUT: 심옥 hand enclosing from above in the middle-center of the frame, foreground; 전택수 hand held in the middle in the middle-center of the frame, foreground; 민정 hand supporting from below in the lower-center of the frame, foreground.\n- KEY BACKGROUND ELEMENTS: ossuary parking area (at dusk); used as Minimal contextual plane around the joined hands.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural dusk light renders the layered hands with subdued contrast and a restrained, humane tone.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Sim-ok and Minjeong clasp Taeksu's hand firmly together in response to his apology.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 심옥 (Korean 여성, 50대 중반, 세월감 있는 얼굴, 검은색 단발머리); 민정 (현재) (Korean 여성, 30대 초반, 차분한 타원형 얼굴, 검은색 중간 길이 머리); 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "roll_refs": {
   "A": [
    {
     "label": "SHOT BACKGROUND",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S91sh3__bgfirst_bg.png"
    },
    {
     "label": "LAYOUT SKETCH (people placement only)",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/conti/conti_S91sh3.png"
    },
    {
     "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:866898>"
    },
    {
     "label": "CHARACTER REFERENCE — 민정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:803682>"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ],
   "B": [
    {
     "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
     "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_납골시설_주차장_210875.png"
    },
    {
     "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:866898>"
    },
    {
     "label": "CHARACTER REFERENCE — 민정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:803682>"
    },
    {
     "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
     "path": "<bytes:917702>"
    }
   ]
  },
  "judge_flip": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "심옥(베이지색 스웨터), 전택수(경찰복), 민정(어두운 코트)의 소매를 정확히 반영하여 세 사람의 손이 위아래로 포개진 샷 텍스트의 구도를 완벽하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "민정의 손이 누락되었으며, 전택수가 두 손으로 심옥의 손을 잡는 잘못된 구조를 연출하여 프롬프트의 지시를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "세 사람의 손이 화면 중앙을 향해 모여 서로를 단단히 맞잡고 있음.",
      "built_space": "아웃포커싱된 해 질 녘의 야외 주차장 공간이 배경으로 확인됨.",
      "entities": "위쪽의 베이지색 소매(심옥), 가운데의 남색 재킷과 체크 셔츠(전택수), 아래쪽의 어두운 코트 소매(민정)가 모두 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "위아래의 손이 가운데 손을 물리적으로 단단히 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "손들이 화면 중앙을 향해 뻗어 교차하고 있음.",
      "built_space": "흐리게 처리된 야외 건축물과 길목이 보임.",
      "entities": "왼쪽의 베이지색 소매(심옥)를 오른쪽의 남색 재킷 소매(전택수) 두 개가 위아래로 잡고 있으며, 민정은 존재하지 않음.",
      "hard_violations": [
       "지시된 인물(민정) 누락",
       "잘못된 신체 구조 및 배치(전택수의 두 손이 심옥의 한 손을 잡고 있음)"
      ],
      "physics": "오른쪽 인물의 두 손이 왼쪽 인물의 손을 붙잡아 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "심옥(베이지색 스웨터), 전택수(경찰복), 민정(어두운 코트)의 소매를 정확히 반영하여 세 사람의 손이 위아래로 포개진 샷 텍스트의 구도를 완벽하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "민정의 손이 누락되었으며, 전택수가 두 손으로 심옥의 손을 잡는 잘못된 구조를 연출하여 프롬프트의 지시를 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "세 사람의 손이 화면 중앙을 향해 모여 서로를 단단히 맞잡고 있음.",
      "built_space": "아웃포커싱된 해 질 녘의 야외 주차장 공간이 배경으로 확인됨.",
      "entities": "위쪽의 베이지색 소매(심옥), 가운데의 남색 재킷과 체크 셔츠(전택수), 아래쪽의 어두운 코트 소매(민정)가 모두 레퍼런스와 일치함.",
      "hard_violations": [],
      "physics": "위아래의 손이 가운데 손을 물리적으로 단단히 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "손들이 화면 중앙을 향해 뻗어 교차하고 있음.",
      "built_space": "흐리게 처리된 야외 건축물과 길목이 보임.",
      "entities": "왼쪽의 베이지색 소매(심옥)를 오른쪽의 남색 재킷 소매(전택수) 두 개가 위아래로 잡고 있으며, 민정은 존재하지 않음.",
      "hard_violations": [
       "지시된 인물(민정) 누락",
       "잘못된 신체 구조 및 배치(전택수의 두 손이 심옥의 한 손을 잡고 있음)"
      ],
      "physics": "오른쪽 인물의 두 손이 왼쪽 인물의 손을 붙잡아 지지하고 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "세 인물의 의상과 손이 포개진 순서(심옥 위, 전택수 중간, 민정 아래)를 요구사항에 맞춰 정확하게 구현한 우수한 컷입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "민정이 누락되었으며, 전택수가 심옥의 손을 잡고 있어 손을 포개는 핵심 지시를 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라가 약간 위에서 아래로 비스듬히 포개진 세 손을 향함.",
      "built_space": "배경은 해질녘의 흐릿한 야외 주차장 구역으로 보임.",
      "entities": "위쪽 베이지 스웨터(심옥), 가운데 남색 재킷과 체크 셔츠(전택수), 아래쪽 짙은 회색 코트(민정) 소매 모두 일치함.",
      "hard_violations": [],
      "physics": "세 사람의 손이 물리적으로 자연스럽게 포개져 서로를 지탱함."
     },
     {
      "label": "B",
      "direction": "카메라가 마주잡은 두 사람의 손을 정면에서 약간 위로 향함.",
      "built_space": "배경은 주차장이 아닌 건물 사이의 야외 통로로 보임.",
      "entities": "왼쪽 베이지 스웨터(심옥)의 손을 오른쪽 남색 재킷(전택수)의 두 손이 감싸고 있으며 민정은 없음.",
      "hard_violations": [
       "등장인물 누락 (민정 없음)",
       "손이 포개진 순서와 주체 위반 (심옥과 민정이 전택수를 잡는 것이 아니라 전택수가 심옥을 잡음)"
      ],
      "physics": "오른쪽 남색 소매에서 나온 두 손의 구조와 연결이 다소 불분명함."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "세 인물의 의상과 손이 포개진 순서(심옥 위, 전택수 중간, 민정 아래)를 요구사항에 맞춰 정확하게 구현한 우수한 컷입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "민정이 누락되었으며, 전택수가 심옥의 손을 잡고 있어 손을 포개는 핵심 지시를 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "카메라가 약간 위에서 아래로 비스듬히 포개진 세 손을 향함.",
      "built_space": "배경은 해질녘의 흐릿한 야외 주차장 구역으로 보임.",
      "entities": "위쪽 베이지 스웨터(심옥), 가운데 남색 재킷과 체크 셔츠(전택수), 아래쪽 짙은 회색 코트(민정) 소매 모두 일치함.",
      "hard_violations": [],
      "physics": "세 사람의 손이 물리적으로 자연스럽게 포개져 서로를 지탱함."
     },
     {
      "label": "A",
      "direction": "카메라가 마주잡은 두 사람의 손을 정면에서 약간 위로 향함.",
      "built_space": "배경은 주차장이 아닌 건물 사이의 야외 통로로 보임.",
      "entities": "왼쪽 베이지 스웨터(심옥)의 손을 오른쪽 남색 재킷(전택수)의 두 손이 감싸고 있으며 민정은 없음.",
      "hard_violations": [
       "등장인물 누락 (민정 없음)",
       "손이 포개진 순서와 주체 위반 (심옥과 민정이 전택수를 잡는 것이 아니라 전택수가 심옥을 잡음)"
      ],
      "physics": "오른쪽 남색 소매에서 나온 두 손의 구조와 연결이 다소 불분명함."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 6,
     "B": 15
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   }
  },
  "readings": [
   {
    "label": "B",
    "direction": "세 사람의 손이 화면 중앙을 향해 모여 서로를 단단히 맞잡고 있음.",
    "built_space": "아웃포커싱된 해 질 녘의 야외 주차장 공간이 배경으로 확인됨.",
    "entities": "위쪽의 베이지색 소매(심옥), 가운데의 남색 재킷과 체크 셔츠(전택수), 아래쪽의 어두운 코트 소매(민정)가 모두 레퍼런스와 일치함.",
    "hard_violations": [],
    "physics": "위아래의 손이 가운데 손을 물리적으로 단단히 지탱하고 있음."
   },
   {
    "label": "A",
    "direction": "손들이 화면 중앙을 향해 뻗어 교차하고 있음.",
    "built_space": "흐리게 처리된 야외 건축물과 길목이 보임.",
    "entities": "왼쪽의 베이지색 소매(심옥)를 오른쪽의 남색 재킷 소매(전택수) 두 개가 위아래로 잡고 있으며, 민정은 존재하지 않음.",
    "hard_violations": [
     "지시된 인물(민정) 누락",
     "잘못된 신체 구조 및 배치(전택수의 두 손이 심옥의 한 손을 잡고 있음)"
    ],
    "physics": "오른쪽 인물의 두 손이 왼쪽 인물의 손을 붙잡아 지지하고 있음."
   }
  ],
  "totals": {
   "A": 6,
   "B": 15
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "심옥(베이지색 스웨터), 전택수(경찰복), 민정(어두운 코트)의 소매를 정확히 반영하여 세 사람의 손이 위아래로 포개진 샷 텍스트의 구도를 완벽하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "민정의 손이 누락되었으며, 전택수가 두 손으로 심옥의 손을 잡는 잘못된 구조를 연출하여 프롬프트의 지시를 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "LOCATION PHOTOGRAPH — the exact place of this shot: its architecture, materials, fixed features and lighting mood are spatial truth; stage the moment inside this place. Never copy its camera framing.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/groupbg_납골시설_주차장_210875.png"
   },
   {
    "label": "CHARACTER REFERENCE — 심옥: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:866898>"
   },
   {
    "label": "CHARACTER REFERENCE — 민정 (현재): the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:803682>"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프레임 우측에 심옥(베이지색 소매)과 민정(어두운 코트 소매) 외에 출처를 알 수 없는 어두운 소매의 팔이 하나 더 추가되어, 프레임 내에 총 4개의 팔이 등장합니다.",
     "fix_en": "Remove the extra arm in the dark grey sleeve on the middle right, replacing it with the dark coat sleeve of the hand below it. Keep the people present, their positions, their clothing, the set, the lighting, and the framing unchanged.",
     "severity": "critical",
     "observation_index": 0,
     "needs_regeneration": true
    },
    {
     "issue_ko": "전택수의 손(좌측, 남색 재킷과 체크 셔츠 소매)이 가운데 위치하여 위아래로 잡히는 구조가 아니라, 다른 손을 측면에서 감싸 쥐고 있어 지시된 손의 층위와 역할이 완전히 틀렸습니다.",
     "fix_en": "Modify the fingers of the left hand (navy jacket with plaid cuff) so they extend flat into the center of the clasped hands rather than wrapping around the outside. Keep the people present, their positions, their clothing, the set, the lighting, and the framing unchanged.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "베이지 스웨터 쪽 위 손(심옥이어야 함)이 중년 여성이 아닌 굵고 큰 남성 손으로 보임",
     "fix_en": "Slim and soften the top hand emerging from the beige sweater so it reads as a middle-aged woman's hand. Keep the people present, their positions, their clothing, the set, the lighting, and the framing unchanged.",
     "severity": "critical",
     "observation_index": 2
    },
    {
     "issue_ko": "오른쪽 앉아 있는 몸통과 허벅지가 크게 나와 팔뚝만 남기는 손 인서트가 아님",
     "fix_en": "Remove the torso, suit jacket, and leg visible on the right side of the frame, replacing them with the blurred grey asphalt of the parking lot background. Keep the people present, their arm positions, their clothing, the set, the lighting, and the framing unchanged.",
     "severity": "major",
     "observation_index": 4
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프레임 우측에 심옥(베이지색 소매)과 민정(어두운 코트 소매) 외에 출처를 알 수 없는 어두운 소매의 팔이 하나 더 추가되어, 프레임 내에 총 4개의 팔이 등장합니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수의 손(좌측, 남색 재킷과 체크 셔츠 소매)이 가운데 위치하여 위아래로 잡히는 구조가 아니라, 다른 손을 측면에서 감싸 쥐고 있어 지시된 손의 층위와 역할이 완전히 틀렸습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "베이지 스웨터 쪽 위 손(심옥이어야 함)이 중년 여성이 아닌 굵고 큰 남성 손으로 보임",
     "severity": "critical"
    },
    {
     "issue_ko": "세 겹이어야 할 맞잡기에 주먹 옆뒤로 네 번째 손이 더 들어가 있음",
     "severity": "major"
    },
    {
     "issue_ko": "오른쪽 앉아 있는 몸통과 허벅지가 크게 나와 팔뚝만 남기는 손 인서트가 아님",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 3
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "오른쪽 앉아 있는 몸통과 허벅지가 크게 나와 팔뚝만 남기는 손 인서트가 아님",
    "fix_en": "Remove the torso, suit jacket, and leg visible on the right side of the frame, replacing them with the blurred grey asphalt of the parking lot background. Keep the people present, their arm positions, their clothing, the set, the lighting, and the framing unchanged.",
    "severity": "major",
    "observation_index": 4
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Remove the extra arm in the dark grey sleeve on the middle right, replacing it with the dark coat sleeve of the hand below it. Keep the people present, their positions, their clothing, the set, the lighting, and the framing unchanged.\n- Modify the fingers of the left hand (navy jacket with plaid cuff) so they extend flat into the center of the clasped hands rather than wrapping around the outside. Keep the people present, their positions, their clothing, the set, the lighting, and the framing unchanged.\n- Slim and soften the top hand emerging from the beige sweater so it reads as a middle-aged woman's hand. Keep the people present, their positions, their clothing, the set, the lighting, and the framing unchanged.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 샷 크기를 완벽하게 따랐으며, 위부터 심옥(베이지색 소매), 전택수(남색 재킷), 민정(어두운 코트) 순서로 손이 포개지는 동작과 의상을 정확히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "프롬프트가 요구한 손 클로즈업과 포개진 동작, 노을 진 주차장 배경을 완전히 무시하고 레퍼런스 인물들을 흰 배경에 전신으로 세워둔 치명적인 실패작입니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 화면 중앙에 위치한 세 사람의 포개진 손을 비스듬히 내려다보고 있습니다.",
      "built_space": "야외 주차장의 흐릿한 배경이 보이며, 노을빛이 손에 닿고 있습니다.",
      "entities": "세 사람의 팔과 손만 보이며, 각 소매의 질감과 색상(베이지색 스웨터, 체크 셔츠와 남색 재킷, 어두운 코트)이 인물 설정과 일치합니다.",
      "hard_violations": [],
      "physics": "세 사람의 손이 공중에서 서로를 단단히 붙잡고 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 카메라 정면을 향해 시선을 두고 있습니다.",
      "built_space": "아무런 구조물이나 배경 묘사가 없는 하얀색 공간입니다.",
      "entities": "심옥, 민정, 전택수가 전신으로 등장하며, 레퍼런스와 의상 및 외형이 일치합니다.",
      "hard_violations": [
       "프롬프트가 지시한 손 클로즈업(insert close-up) 대신 전신 샷으로 렌더링함.",
       "지시된 야외 주차장 배경이 아닌 빈 스튜디오 배경으로 렌더링함.",
       "캐릭터 간의 동작(손 맞잡기)이 전혀 이루어지지 않음."
      ],
      "physics": "세 사람이 하얀 바닥에 두 발을 딛고 서 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "지시된 클로즈업 샷 크기를 완벽하게 따랐으며, 위부터 심옥(베이지색 소매), 전택수(남색 재킷), 민정(어두운 코트) 순서로 손이 포개지는 동작과 의상을 정확히 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "프롬프트가 요구한 손 클로즈업과 포개진 동작, 노을 진 주차장 배경을 완전히 무시하고 레퍼런스 인물들을 흰 배경에 전신으로 세워둔 치명적인 실패작입니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 화면 중앙에 위치한 세 사람의 포개진 손을 비스듬히 내려다보고 있습니다.",
      "built_space": "야외 주차장의 흐릿한 배경이 보이며, 노을빛이 손에 닿고 있습니다.",
      "entities": "세 사람의 팔과 손만 보이며, 각 소매의 질감과 색상(베이지색 스웨터, 체크 셔츠와 남색 재킷, 어두운 코트)이 인물 설정과 일치합니다.",
      "hard_violations": [],
      "physics": "세 사람의 손이 공중에서 서로를 단단히 붙잡고 지탱하고 있습니다."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 카메라 정면을 향해 시선을 두고 있습니다.",
      "built_space": "아무런 구조물이나 배경 묘사가 없는 하얀색 공간입니다.",
      "entities": "심옥, 민정, 전택수가 전신으로 등장하며, 레퍼런스와 의상 및 외형이 일치합니다.",
      "hard_violations": [
       "프롬프트가 지시한 손 클로즈업(insert close-up) 대신 전신 샷으로 렌더링함.",
       "지시된 야외 주차장 배경이 아닌 빈 스튜디오 배경으로 렌더링함.",
       "캐릭터 간의 동작(손 맞잡기)이 전혀 이루어지지 않음."
      ],
      "physics": "세 사람이 하얀 바닥에 두 발을 딛고 서 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 9,
      "verdict_ko": "요청된 세 사람의 손이 포개진 클로즈업 샷을 정확히 구현했으며, 각 인물의 의상 소매 디테일과 해질녘 주차장 배경의 분위기까지 완벽하게 살려냈습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "프레이밍, 장소, 행동 지시를 완전히 무시하고 스튜디오 전신 샷을 생성하여 샷 텍스트의 요구사항을 전혀 충족하지 못했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "카메라는 비스듬히 아래를 향해 포개진 손들을 비추고 있으며, 세 손이 서로를 단단히 맞잡고 있음.",
      "built_space": "배경은 흐릿하지만 아스팔트 바닥과 건물의 형태가 보여 주차장 환경과 일치함.",
      "entities": "베이지색 소매의 손(심옥)이 위에서, 남색 재킷과 체크 셔츠 소매의 손(전택수)이 중간에, 짙은 회색 코트와 남색 니트 소매의 손(민정)이 아래에서 받치고 있어 레퍼런스 의상과 정확히 일치함.",
      "hard_violations": [],
      "physics": "세 사람의 손이 물리적으로 자연스럽게 포개져 서로를 지탱하고 있음."
     },
     {
      "label": "A",
      "direction": "세 인물 모두 정면의 카메라를 똑바로 응시하고 있음.",
      "built_space": "아무런 구조물이나 배경이 없는 하얀색 스튜디오 공간임.",
      "entities": "심옥, 민정, 전택수가 레퍼런스의 얼굴과 의상 그대로 서 있으나, 지시된 특정 신체 부위(손)가 아닌 전신이 노출됨.",
      "hard_violations": [
       "클로즈업 샷 지시를 위반하고 전신 샷을 생성함 (Staging violation)",
       "해질녘 야외 주차장 배경 지시를 위반하고 하얀 배경을 생성함 (Location violation)",
       "손을 맞잡는 행동 지시를 위반하고 차렷 자세로 서 있음"
      ],
      "physics": "세 인물 모두 바닥에 두 발을 딛고 서 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 9,
      "verdict_ko": "요청된 세 사람의 손이 포개진 클로즈업 샷을 정확히 구현했으며, 각 인물의 의상 소매 디테일과 해질녘 주차장 배경의 분위기까지 완벽하게 살려냈습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "프레이밍, 장소, 행동 지시를 완전히 무시하고 스튜디오 전신 샷을 생성하여 샷 텍스트의 요구사항을 전혀 충족하지 못했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 비스듬히 아래를 향해 포개진 손들을 비추고 있으며, 세 손이 서로를 단단히 맞잡고 있음.",
      "built_space": "배경은 흐릿하지만 아스팔트 바닥과 건물의 형태가 보여 주차장 환경과 일치함.",
      "entities": "베이지색 소매의 손(심옥)이 위에서, 남색 재킷과 체크 셔츠 소매의 손(전택수)이 중간에, 짙은 회색 코트와 남색 니트 소매의 손(민정)이 아래에서 받치고 있어 레퍼런스 의상과 정확히 일치함.",
      "hard_violations": [],
      "physics": "세 사람의 손이 물리적으로 자연스럽게 포개져 서로를 지탱하고 있음."
     },
     {
      "label": "B",
      "direction": "세 인물 모두 정면의 카메라를 똑바로 응시하고 있음.",
      "built_space": "아무런 구조물이나 배경이 없는 하얀색 스튜디오 공간임.",
      "entities": "심옥, 민정, 전택수가 레퍼런스의 얼굴과 의상 그대로 서 있으나, 지시된 특정 신체 부위(손)가 아닌 전신이 노출됨.",
      "hard_violations": [
       "클로즈업 샷 지시를 위반하고 전신 샷을 생성함 (Staging violation)",
       "해질녘 야외 주차장 배경 지시를 위반하고 하얀 배경을 생성함 (Location violation)",
       "손을 맞잡는 행동 지시를 위반하고 차렷 자세로 서 있음"
      ],
      "physics": "세 인물 모두 바닥에 두 발을 딛고 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 16,
     "B": 4
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "bgfirst": {
   "bg_path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S91sh3__bgfirst_bg.png",
   "bg_asset_id": "f74ffde5-de75-489d-ac6a-5cbe021113d4",
   "bg_record_key": "S91sh3::bgfirst_bg",
   "chain_winner": false,
   "authority": "groupbg",
   "group_key": "납골시설 주차장",
   "groupbg_asset_id": "5f540a15-e154-47a9-b2e5-24c7c2b9b95b"
  },
  "ref_mode": "그룹 배경+엔티티 (2택1: 무콘티 승)",
  "share_plan": {
   "ref_plan": "background"
  }
 },
 "S91sh3::cine": {
  "applied": true,
  "fingerprint": "cb4ffa6416349575406cf6310ee4fa21572ad71e3f8dc9a6fd8f5d13628d6819",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S91sh3_sel.png",
  "source_sha256": "057020f68abddb9e0bc1167e498a9f4f2d80519357e0d7283e43018568775972",
  "file": "S91sh3_cine.png",
  "latency_ms": 12040
 },
 "S91sh4::signage": {
  "fp": "633bb247a80a2ab5",
  "inscriptions": [
   {
    "surface_native": "유리 안 명패",
    "text_native": "故 김선영",
    "reason_ko": "납골당 봉안단 유리 안쪽에 고인의 신원을 표시하는 명패가 있어야 장면의 사실성을 높일 수 있습니다."
   }
  ]
 },
 "S91sh4": {
  "input_fingerprint": "92fedff5e6052ba5",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 납골당 유리 안쪽, 환하게 웃고 있는 김선영의 얼굴이 담긴 액자 속 평면 사진 클로즈업\n\nLOCATION (lock): Inside the columbarium, directly before the glass-fronted memorial niche containing the framed photograph. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Place the camera just outside the ossuary glass at face height and close range, slightly off the framed photograph's frontal axis. 김선영's brightly smiling face occupies most of the composition while the frame edge and intervening glass remain visible enough to establish that this is a flat memorial photograph rather than a living portrait.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: ossuary glass (fixed in front of the photograph) — The framed photograph of 김선영 is visible through the glass; used as Intervening plane through which the memorial photograph is viewed; framed photograph (positioned inside the ossuary) — The image-bearing front face is visible at a slight oblique angle and shows 김선영 smiling brightly; used as Primary focal surface carrying 김선영's smiling image.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the ossuary is kept subdued and neutral, allowing the brightness of 김선영's photographed expression to carry the emotional contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's bright portrait remains fixed inside the columbarium niche.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유리 안 명패: \"故 김선영\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 납골당 유리 안쪽, 환하게 웃고 있는 김선영의 얼굴이 담긴 액자 속 평면 사진 클로즈업\n\nLOCATION (lock): Inside the columbarium, directly before the glass-fronted memorial niche containing the framed photograph. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Place the camera just outside the ossuary glass at face height and close range, slightly off the framed photograph's frontal axis. 김선영's brightly smiling face occupies most of the composition while the frame edge and intervening glass remain visible enough to establish that this is a flat memorial photograph rather than a living portrait.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: ossuary glass (fixed in front of the photograph) — The framed photograph of 김선영 is visible through the glass; used as Intervening plane through which the memorial photograph is viewed; framed photograph (positioned inside the ossuary) — The image-bearing front face is visible at a slight oblique angle and shows 김선영 smiling brightly; used as Primary focal surface carrying 김선영's smiling image.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the ossuary is kept subdued and neutral, allowing the brightness of 김선영's photographed expression to carry the emotional contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's bright portrait remains fixed inside the columbarium niche.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유리 안 명패: \"故 김선영\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 납골당 유리 안쪽, 환하게 웃고 있는 김선영의 얼굴이 담긴 액자 속 평면 사진 클로즈업\n\nLOCATION (lock): Inside the columbarium, directly before the glass-fronted memorial niche containing the framed photograph. The shot takes place here — the attached LOCATION STRUCTURE PHOTOGRAPH is the single authority for this exact place — its fixed structure and permanent site details are LOCKED to it. No separate location photograph exists for this place. Build everything else strictly from the location text above and the shot text; the layout sketch (when attached) governs framing and placement only, and the shot text governs time of day, lighting and action.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Place the camera just outside the ossuary glass at face height and close range, slightly off the framed photograph's frontal axis. 김선영's brightly smiling face occupies most of the composition while the frame edge and intervening glass remain visible enough to establish that this is a flat memorial photograph rather than a living portrait.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: ossuary glass (fixed in front of the photograph) — The framed photograph of 김선영 is visible through the glass; used as Intervening plane through which the memorial photograph is viewed; framed photograph (positioned inside the ossuary) — The image-bearing front face is visible at a slight oblique angle and shows 김선영 smiling brightly; used as Primary focal surface carrying 김선영's smiling image.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Natural ambient illumination appropriate to the ossuary is kept subdued and neutral, allowing the brightness of 김선영's photographed expression to carry the emotional contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Seonyeong's bright portrait remains fixed inside the columbarium niche.\n\nNO PEOPLE IN THIS SHOT: the shot text shows only the place and its state — no living person or any body part appears in frame, unless the shot text itself explicitly says so.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 유리 안 명패: \"故 김선영\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "gq": {
   "route": "combined",
   "gap": 0.667,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "dual": {
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "normalized": {
    "A": 2.0,
    "B": 1.333
   },
   "adjusted": {
    "A": 1.25,
    "B": 0.583
   },
   "violations": {
    "A": [
     "[gemini-pro] 명시된 카메라 위치(납골당 실내)에서는 발생할 수 없는 반사/배경(레퍼런스의 야외 건물과 주차장이 유리에 비침).",
     "[openrouter:x-ai/grok-4.6] 약정된 납골당 봉안함 유리 바로 밖 얼굴 높이가 아니라 건물 외부가 보이는 창가에 카메라를 둠",
     "[openrouter:x-ai/grok-4.6] 유리 앞 봉안 니치가 아닌 창턱 액자와 주차장·건물 외관을 발명함"
    ],
    "B": [
     "[gemini-pro] 명시된 카메라 위치(납골당 실내)에서는 발생할 수 없는 반사/배경(레퍼런스의 야외 건물과 일몰이 유리에 비침).",
     "[openrouter:x-ai/grok-4.6] 약정된 납골당 봉안함 유리 바로 밖 얼굴 높이가 아니라 건물 외부가 보이는 창가에 카메라를 둠",
     "[openrouter:x-ai/grok-4.6] 유리 앞 봉안 니치가 아닌 창턱 액자와 창밖 건물·노을 전망을 발명함"
    ]
   },
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "agreed": false
  },
  "totals": {
   "B": 583,
   "A": 1250
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 583,
    "verdict_ko": "명패의 형태는 적절하나, 프롬프트가 '납골당 실내'를 명시했음에도 레퍼런스의 야외 건물이 유리에 반사되거나 배경으로 나타나는 심각한 공간적 오류가 있습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(납골당 실내)에서는 발생할 수 없는 반사/배경(레퍼런스의 야외 건물과 일몰이 유리에 비침). / [openrouter:x-ai/grok-4.6] 약정된 납골당 봉안함 유리 바로 밖 얼굴 높이가 아니라 건물 외부가 보이는 창가에 카메라를 둠 / [openrouter:x-ai/grok-4.6] 유리 앞 봉안 니치가 아닌 창턱 액자와 창밖 건물·노을 전망을 발명함"
   },
   {
    "label": "A",
    "score": 1250,
    "verdict_ko": "텍스트와 인물의 표정은 잘 표현되었으나, B와 마찬가지로 실내 위치 설정과 모순되게 야외 건물이 배경에 등장하는 물리적/공간적 오류가 발생했습니다.  ★위반: [gemini-pro] 명시된 카메라 위치(납골당 실내)에서는 발생할 수 없는 반사/배경(레퍼런스의 야외 건물과 주차장이 유리에 비침). / [openrouter:x-ai/grok-4.6] 약정된 납골당 봉안함 유리 바로 밖 얼굴 높이가 아니라 건물 외부가 보이는 창가에 카메라를 둠 / [openrouter:x-ai/grok-4.6] 유리 앞 봉안 니치가 아닌 창턱 액자와 주차장·건물 외관을 발명함"
   }
  ],
  "refs": [
   {
    "label": "LOCATION STRUCTURE PHOTOGRAPH — the confirmed photograph of this exact place and its fixed structure: it is the SINGLE authority for the location, the structure's shape, proportions, materials, colors, openings and every permanent site detail. Never copy its camera framing, time of day or lighting — the shot text is the authority for those.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/background_chain/seed_bg_memorial_ossuary_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "장소가 납골당 내부 유리 안치단이 아니라 건물 외벽과 주차장이 보이는 창가이다.",
     "fix_en": "Replace the bright outdoor building, parking lot, and sky in the background with the dim, neutral indoor environment of a columbarium, featuring adjacent marble memorial niches. Preserve the woman's portrait, the wooden frame, and the metal nameplate exactly as they are.",
     "severity": "critical",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "납골당 유리가 카메라와 사진 사이에 없고 액자 뒤 창문 유리로만 보인다.",
     "fix_en": "Render a continuous pane of glass in the extreme foreground, placing it directly between the camera and the wooden frame, with subtle interior reflections to show the portrait is enclosed inside a niche. Preserve the woman's portrait, the wooden frame, and the metal nameplate.",
     "severity": "critical",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "김선영 얼굴이 화면 대부분을 차지하지 않고 우측·상단 건물 배경이 크게 드러난다.",
     "fix_en": "Scale up the framed portrait so that the woman's smiling face fills the majority of the composition, significantly reducing the amount of visible background. Preserve her face, expression, and clothing.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "납골당 내부 공간임에도 불구하고 유리 표면에 건물 외관(참조 사진)이 반사되어 보이는 것은 물리적으로 불가능한 반사입니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "장소가 납골당 내부 유리 안치단이 아니라 건물 외벽과 주차장이 보이는 창가이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "액자가 안치단 유리 안쪽에 고정되지 않고 창틀 선반 위 탁상 액자로 세워져 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "납골당 유리가 카메라와 사진 사이에 없고 액자 뒤 창문 유리로만 보인다.",
     "severity": "critical"
    },
    {
     "issue_ko": "김선영 얼굴이 화면 대부분을 차지하지 않고 우측·상단 건물 배경이 크게 드러난다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 4
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "김선영 얼굴이 화면 대부분을 차지하지 않고 우측·상단 건물 배경이 크게 드러난다.",
    "fix_en": "Scale up the framed portrait so that the woman's smiling face fills the majority of the composition, significantly reducing the amount of visible background. Preserve her face, expression, and clothing.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 2,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the bright outdoor building, parking lot, and sky in the background with the dim, neutral indoor environment of a columbarium, featuring adjacent marble memorial niches. Preserve the woman's portrait, the wooden frame, and the metal nameplate exactly as they are.\n- Render a continuous pane of glass in the extreme foreground, placing it directly between the camera and the wooden frame, with subtle interior reflections to show the portrait is enclosed inside a niche. Preserve the woman's portrait, the wooden frame, and the metal nameplate.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "명패에 한자가 쓰이고 배경에 임의의 텍스트가 등장하는 단점이 있으나, 실내 납골당 안치단이라는 공간적 설정과 유리 반사를 물리적 모순 없이 타당하게 구현하여 더 나은 결과물입니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "명패의 한글 텍스트는 완벽하게 렌더링되었으나, 실내 납골당 샷임에도 배경에 기준 사진의 건물 외관 전체가 노출되는 치명적인 물리적 공간 연출 오류(Hard Violation)가 발생했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "사진 속 여성의 시선과 미소는 정면(카메라)을 향하고 있음.",
      "built_space": "실내 납골당 안치단 내부에 액자가 배치됨. 배경으로 다른 안치단들이 늘어선 실내 공간이 보이며, 앞면 유리에 천장 조명이 반사되어 맺힘.",
      "entities": "환하게 웃고 있는 여성의 영정 사진이 보임. 명패에는 지정된 한글 대신 한자 '故 金善英'이 적혀 있으며, 배경 안치단에 의미를 알 수 없는 문자 흔적들이 묘사됨.",
      "hard_violations": [],
      "physics": "액자는 안치단 바닥면에 놓여 있으며 뒤쪽 스탠드에 의해 안정적으로 지탱됨."
     },
     {
      "label": "A",
      "direction": "사진 속 여성의 시선과 미소는 정면(카메라)을 향하고 있음.",
      "built_space": "유리창 안쪽에 액자가 배치되어 있으나, 배경으로 기준 사진의 실외 건물 전체와 주차장이 선명하게 나타나 실내 위치 설정과 정면으로 배치됨.",
      "entities": "환하게 웃고 있는 여성의 영정 사진. 명패에 지정된 텍스트 '故 김선영'이 정확하게 렌더링됨.",
      "hard_violations": [
       "물리적으로 불가능한 무대 연출 (실내 납골당 안치단 내부를 비추는 샷에서 건물 외관 전체와 주차장이 배경으로 보임)"
      ],
      "physics": "액자는 선반 위에 놓여 스탠드로 지지되어 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "명패에 한자가 쓰이고 배경에 임의의 텍스트가 등장하는 단점이 있으나, 실내 납골당 안치단이라는 공간적 설정과 유리 반사를 물리적 모순 없이 타당하게 구현하여 더 나은 결과물입니다."
     },
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "명패의 한글 텍스트는 완벽하게 렌더링되었으나, 실내 납골당 샷임에도 배경에 기준 사진의 건물 외관 전체가 노출되는 치명적인 물리적 공간 연출 오류(Hard Violation)가 발생했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "사진 속 여성의 시선과 미소는 정면(카메라)을 향하고 있음.",
      "built_space": "실내 납골당 안치단 내부에 액자가 배치됨. 배경으로 다른 안치단들이 늘어선 실내 공간이 보이며, 앞면 유리에 천장 조명이 반사되어 맺힘.",
      "entities": "환하게 웃고 있는 여성의 영정 사진이 보임. 명패에는 지정된 한글 대신 한자 '故 金善英'이 적혀 있으며, 배경 안치단에 의미를 알 수 없는 문자 흔적들이 묘사됨.",
      "hard_violations": [],
      "physics": "액자는 안치단 바닥면에 놓여 있으며 뒤쪽 스탠드에 의해 안정적으로 지탱됨."
     },
     {
      "label": "A",
      "direction": "사진 속 여성의 시선과 미소는 정면(카메라)을 향하고 있음.",
      "built_space": "유리창 안쪽에 액자가 배치되어 있으나, 배경으로 기준 사진의 실외 건물 전체와 주차장이 선명하게 나타나 실내 위치 설정과 정면으로 배치됨.",
      "entities": "환하게 웃고 있는 여성의 영정 사진. 명패에 지정된 텍스트 '故 김선영'이 정확하게 렌더링됨.",
      "hard_violations": [
       "물리적으로 불가능한 무대 연출 (실내 납골당 안치단 내부를 비추는 샷에서 건물 외관 전체와 주차장이 배경으로 보임)"
      ],
      "physics": "액자는 선반 위에 놓여 스탠드로 지지되어 있음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 6,
      "verdict_ko": "실내 납골당의 공간감을 논리적으로 구현했으나, 명패의 텍스트가 요청된 한글 대신 한자로 표기된 점이 감점 요인입니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요청된 한글 텍스트는 정확히 반영했으나, 실내 씬임에도 불구하고 유리에 건물 외부가 반사되는 치명적인 공간 논리 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "카메라는 유리 너머로 정면을 향해 놓인 여성의 인물 사진 액자를 바라보고 있습니다.",
      "built_space": "납골당 내부 안치단 안쪽에 액자가 놓여 있으며, 유리 표면에는 카메라 뒤편에 있는 다른 안치단 벽면이 실내 환경에 맞게 반사되어 보입니다.",
      "entities": "밝게 웃는 여성의 사진이 보이지만, 명패의 글씨가 '故 김선영'이 아닌 한자 '故 金善英'으로 잘못 출력되었습니다.",
      "hard_violations": [],
      "physics": "액자는 안치단 바닥면에 안정적으로 세워져 있습니다."
     },
     {
      "label": "B",
      "direction": "카메라는 유리 너머로 정면을 향해 놓인 여성의 인물 사진 액자를 바라보고 있습니다.",
      "built_space": "유리에 레퍼런스 사진 속 건물 외부와 주차장이 반사되어 나타나는데, 실내 납골당 내부를 촬영하는 카메라 위치에서는 발생할 수 없는 불가능한 반사입니다.",
      "entities": "밝게 웃는 여성의 사진이 보이며, 명패에 지정된 텍스트 '故 김선영'이 정확하게 표기되어 있습니다.",
      "hard_violations": [
       "실내 카메라 위치에서 물리적으로 불가능한 반사(실내 유리에 외부 건물과 주차장이 비침)"
      ],
      "physics": "액자는 안치단 바닥면에 안정적으로 세워져 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 6,
      "verdict_ko": "실내 납골당의 공간감을 논리적으로 구현했으나, 명패의 텍스트가 요청된 한글 대신 한자로 표기된 점이 감점 요인입니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "요청된 한글 텍스트는 정확히 반영했으나, 실내 씬임에도 불구하고 유리에 건물 외부가 반사되는 치명적인 공간 논리 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "카메라는 유리 너머로 정면을 향해 놓인 여성의 인물 사진 액자를 바라보고 있습니다.",
      "built_space": "납골당 내부 안치단 안쪽에 액자가 놓여 있으며, 유리 표면에는 카메라 뒤편에 있는 다른 안치단 벽면이 실내 환경에 맞게 반사되어 보입니다.",
      "entities": "밝게 웃는 여성의 사진이 보이지만, 명패의 글씨가 '故 김선영'이 아닌 한자 '故 金善英'으로 잘못 출력되었습니다.",
      "hard_violations": [],
      "physics": "액자는 안치단 바닥면에 안정적으로 세워져 있습니다."
     },
     {
      "label": "A",
      "direction": "카메라는 유리 너머로 정면을 향해 놓인 여성의 인물 사진 액자를 바라보고 있습니다.",
      "built_space": "유리에 레퍼런스 사진 속 건물 외부와 주차장이 반사되어 나타나는데, 실내 납골당 내부를 촬영하는 카메라 위치에서는 발생할 수 없는 불가능한 반사입니다.",
      "entities": "밝게 웃는 여성의 사진이 보이며, 명패에 지정된 텍스트 '故 김선영'이 정확하게 표기되어 있습니다.",
      "hard_violations": [
       "실내 카메라 위치에서 물리적으로 불가능한 반사(실내 유리에 외부 건물과 주차장이 비침)"
      ],
      "physics": "액자는 안치단 바닥면에 안정적으로 세워져 있습니다."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 12
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "seed-bg만 (배경 전용)",
  "share_plan": {
   "ref_plan": "background"
  },
  "lane_policy": "ab_select_bypass:bg_only"
 },
 "S91sh4::cine": {
  "applied": true,
  "fingerprint": "476e5827d71bc90e8e119e47054659ab93a6b28fbdb9b085a49e2276f031b882",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S91sh4_sel.png",
  "source_sha256": "89d7ae1c03066de8abf054fa43a9da33be70ae20a32ce76850fa3f69d95f3989",
  "file": "S91sh4_cine.png",
  "latency_ms": 11293
 },
 "S92sh4::signage": {
  "fp": "538625c579ec72fb",
  "inscriptions": []
 },
 "S92sh4": {
  "input_fingerprint": "51fcdc1d4c4e59aa",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 노을 지는 강변에 나란히 서서 붉게 물든 수면을 응시하는 전택수와 장원섭의 전경\n\nLOCATION (lock): On the quiet open riverbank at sunset, beside the parked cars and facing the reddened water. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Cut to a slightly wider rear three-quarter view from near shoulder height, behind and to one side of 전택수 and 장원섭, keeping both full figures and the river in their shared sightline. 전택수 settles his weight toward the view while 장원섭 holds one foot slightly advanced beside him, their parallel orientation unified without making their silhouettes mechanically identical.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward river and sunset; 장원섭 in the middle-right of the frame, midground, looks toward river and sunset; river and sunset in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: river (overlooked from the riverbank) — The water reflects the red tones of the sunset within the men's shared view; used as Shared background vista receiving both men's sightlines; riverbank (quiet at dusk); used as Ground plane establishing the men's side-by-side placement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun gives the river and the two figures restrained warm-red dusk tones while contrast remains naturalistic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 노을 지는 강변에 나란히 서서 붉게 물든 수면을 응시하는 전택수와 장원섭의 전경\n\nLOCATION (lock): On the quiet open riverbank at sunset, beside the parked cars and facing the reddened water. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Cut to a slightly wider rear three-quarter view from near shoulder height, behind and to one side of 전택수 and 장원섭, keeping both full figures and the river in their shared sightline. 전택수 settles his weight toward the view while 장원섭 holds one foot slightly advanced beside him, their parallel orientation unified without making their silhouettes mechanically identical.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward river and sunset; 장원섭 in the middle-right of the frame, midground, looks toward river and sunset; river and sunset in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: river (overlooked from the riverbank) — The water reflects the red tones of the sunset within the men's shared view; used as Shared background vista receiving both men's sightlines; riverbank (quiet at dusk); used as Ground plane establishing the men's side-by-side placement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun gives the river and the two figures restrained warm-red dusk tones while contrast remains naturalistic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 노을 지는 강변에 나란히 서서 붉게 물든 수면을 응시하는 전택수와 장원섭의 전경\n\nLOCATION (lock): On the quiet open riverbank at sunset, beside the parked cars and facing the reddened water. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Cut to a slightly wider rear three-quarter view from near shoulder height, behind and to one side of 전택수 and 장원섭, keeping both full figures and the river in their shared sightline. 전택수 settles his weight toward the view while 장원섭 holds one foot slightly advanced beside him, their parallel orientation unified without making their silhouettes mechanically identical.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-left of the frame, midground, looks toward river and sunset; 장원섭 in the middle-right of the frame, midground, looks toward river and sunset; river and sunset in the upper-center of the frame, background.\n- KEY BACKGROUND ELEMENTS: river (overlooked from the riverbank) — The water reflects the red tones of the sunset within the men's shared view; used as Shared background vista receiving both men's sightlines; riverbank (quiet at dusk); used as Ground plane establishing the men's side-by-side placement.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: The setting sun gives the river and the two figures restrained warm-red dusk tones while contrast remains naturalistic.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 장원섭 (Korean 남성, 40대 초반, 갸름한 얼굴, 정돈된 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "두 남자가 앞쪽의 붉은 강물을 향해 시선을 두고 있음.",
    "built_space": "흙으로 된 강변과 배경의 콘크리트 둑, 흐릿한 차량들이 보임.",
    "entities": "전택수(왼쪽)는 뒷면에 'SALICS'가 적힌 남색 재킷을 입고 있으며 발목 아래가 잘림. 장원섭(오른쪽)은 짙은 색 재킷을 입고 발목 아래가 잘림.",
    "hard_violations": [],
    "physics": "두 사람 모두 바닥에 서 있으나 프레임에 의해 발이 잘려 지지면이 완전히 보이지 않음."
   },
   {
    "label": "B",
    "direction": "두 남자가 붉게 물든 강물을 응시하고 있음.",
    "built_space": "강변의 흙바닥, 뚜렷한 콘크리트 둑 위로 주차된 차량들이 보임.",
    "entities": "전택수(왼쪽)는 모자 뒷면에 '수사'가 적혀 있고 전신이 프레임에 들어옴. 장원섭(오른쪽)은 참조와 일치하는 회색 정장을 입고 전신이 보임.",
    "hard_violations": [],
    "physics": "두 사람 모두 강변 흙바닥에 두 발을 딛고 안정적으로 서 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "두 인물의 전신을 모두 프레임에 담아 촬영 규모 지시를 정확히 따랐으며, 배경의 주차된 차량과 주변 환경을 훌륭하게 구현했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "인물들의 발목이 잘려 전신을 유지하라는 프레이밍 지시를 위반했으며, 전택수의 재킷 뒷면에 잘못된 텍스트가 생성되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S20sh5_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 장원섭: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:872599>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "전택수가 착용한 모자의 뒷면에 앞면에 있어야 할 '수사'라는 텍스트가 잘못 생성되어 있습니다.",
     "fix_en": "Remove the yellow text from the back of the left man's cap, replacing it with solid navy blue fabric. Preserve both men, their clothing, the riverbank, and the sunset lighting.",
     "severity": "major",
     "observation_index": 0
    },
    {
     "issue_ko": "두 사람이 거의 대칭적인 직립 자세로 서 있어 전택수의 체중 이동과 장원섭의 한 발 앞선 자세가 드러나지 않는다.",
     "fix_en": "Redraw the men's stances: shift the left man's weight noticeably to one side, and advance one of the right man's feet slightly forward. Preserve their identities, clothing, the river setting, and the framing.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "전택수가 착용한 모자의 뒷면에 앞면에 있어야 할 '수사'라는 텍스트가 잘못 생성되어 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "프롬프트에 장원섭이 한 발을 약간 앞으로 내밀고 서 있다고 지시되었으나, 두 발을 평행하게 모으고 똑바로 서 있습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 전택수 모자 뒷면에 '수사' 글자가 보여 잠긴 이전 의상(앞면 위치)과 다르다.",
     "severity": "major"
    },
    {
     "issue_ko": "두 사람이 거의 대칭적인 직립 자세로 서 있어 전택수의 체중 이동과 장원섭의 한 발 앞선 자세가 드러나지 않는다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 2,
  "fix_severity_skipped": [
   {
    "issue_ko": "전택수가 착용한 모자의 뒷면에 앞면에 있어야 할 '수사'라는 텍스트가 잘못 생성되어 있습니다.",
    "fix_en": "Remove the yellow text from the back of the left man's cap, replacing it with solid navy blue fabric. Preserve both men, their clothing, the riverbank, and the sunset lighting.",
    "severity": "major",
    "observation_index": 0
   },
   {
    "issue_ko": "두 사람이 거의 대칭적인 직립 자세로 서 있어 전택수의 체중 이동과 장원섭의 한 발 앞선 자세가 드러나지 않는다.",
    "fix_en": "Redraw the men's stances: shift the left man's weight noticeably to one side, and advance one of the right man's feet slightly forward. Preserve their identities, clothing, the river setting, and the framing.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S20sh5"
  }
 },
 "S92sh4::cine": {
  "applied": true,
  "fingerprint": "aac4561e71d333c0d37defb0e40a2c775bf87870a88145c8da778cbc201d8890",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S92sh4_sel.png",
  "source_sha256": "0752756b0051c792c2d1b6f42358cb5f2020217c88aab2b36400176502169274",
  "file": "S92sh4_cine.png",
  "latency_ms": 11703
 },
 "S92sh7::signage": {
  "fp": "7c8bbb89a52c4f53",
  "inscriptions": []
 },
 "S92sh7": {
  "input_fingerprint": "84bb759ce828524d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 지는 태양을 향해 시선을 고정한 채 씁쓸한 표정에 잠긴 전택수의 감상적인 얼굴\n\nLOCATION (lock): At the riverbank's edge, facing the setting sun over the river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the lateral dolly-in close beside 전택수, slightly below eye level and offset from his forward sightline, settling on his three-quarter profile. His face occupies most of the frame without becoming an extreme close-up, with restrained look room toward the setting sun as his eyes remain fixed there and his expression settles into quiet bitterness and reflection.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: river (visible near dusk) — The river carries subdued sunset tones behind 전택수 without replacing his face as the focal subject; used as Soft contextual background behind 전택수's profile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-sun light shapes 전택수's profile with restrained warmth and moderate naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sunset riverbank, red reflections, quiet water, and the officer's clothing. Exclude the prosecutor from the close frame and isolate the officer looking toward the setting sun.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 지는 태양을 향해 시선을 고정한 채 씁쓸한 표정에 잠긴 전택수의 감상적인 얼굴\n\nLOCATION (lock): At the riverbank's edge, facing the setting sun over the river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the lateral dolly-in close beside 전택수, slightly below eye level and offset from his forward sightline, settling on his three-quarter profile. His face occupies most of the frame without becoming an extreme close-up, with restrained look room toward the setting sun as his eyes remain fixed there and his expression settles into quiet bitterness and reflection.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: river (visible near dusk) — The river carries subdued sunset tones behind 전택수 without replacing his face as the focal subject; used as Soft contextual background behind 전택수's profile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-sun light shapes 전택수's profile with restrained warmth and moderate naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sunset riverbank, red reflections, quiet water, and the officer's clothing. Exclude the prosecutor from the close frame and isolate the officer looking toward the setting sun.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): sunset.\n\nSHOT TEXT (authoritative, Korean): 지는 태양을 향해 시선을 고정한 채 씁쓸한 표정에 잠긴 전택수의 감상적인 얼굴\n\nLOCATION (lock): At the riverbank's edge, facing the setting sun over the river. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the lateral dolly-in close beside 전택수, slightly below eye level and offset from his forward sightline, settling on his three-quarter profile. His face occupies most of the frame without becoming an extreme close-up, with restrained look room toward the setting sun as his eyes remain fixed there and his expression settles into quiet bitterness and reflection.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: river (visible near dusk) — The river carries subdued sunset tones behind 전택수 without replacing his face as the focal subject; used as Soft contextual background behind 전택수's profile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Setting-sun light shapes 전택수's profile with restrained warmth and moderate naturalistic contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same sunset riverbank, red reflections, quiet water, and the officer's clothing. Exclude the prosecutor from the close frame and isolate the officer looking toward the setting sun.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "남성이 화면 우측 밖의 해질녘 강가를 향해 시선을 고정하고 있음.",
    "built_space": "좌측에 경사진 콘크리트 제방이 있고 우측에 강물이 위치해 이전 샷의 공간적 특징과 일치함.",
    "entities": "전택수의 얼굴 특징, 남색 점퍼, 체크 셔츠, '수사' 모자 등 레퍼런스와 일치하며 좌측 가슴의 텍스트 위치도 자연스러움.",
    "hard_violations": [],
    "physics": "화면 밖 지면에 안정적으로 서서 체중을 지탱하고 있음."
   },
   {
    "label": "B",
    "direction": "남성이 화면 우측 밖의 해질녘 강가를 향해 시선을 고정하고 있음.",
    "built_space": "좌측 제방과 우측 강물의 배치가 이전 샷의 구조와 일치함.",
    "entities": "인물의 얼굴과 모자는 레퍼런스와 유사하나, 점퍼의 우측 가슴에 레퍼런스에 없는 'POLICE' 문구가 추가됨.",
    "hard_violations": [],
    "physics": "화면 밖 지면에 서서 체중을 지탱하고 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 7,
   "B": 5
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 7,
    "verdict_ko": "이전 샷의 배경 및 인물 외형(의상, 얼굴)을 완벽하게 재현했으며, 지는 해를 향한 씁쓸하고 감상적인 표정과 조명 분위기를 매우 영화적으로 연출함."
   },
   {
    "label": "B",
    "score": 5,
    "verdict_ko": "배경과 인물의 기본적인 배치는 좋으나, 레퍼런스와 달리 점퍼 우측 가슴에 'POLICE' 문구가 잘못 중복되어 추가되었으며 표정의 깊이가 다소 부족함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S92sh4_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "자켓 왼쪽 가슴에 있어야 할 'POLICE' 문구가 화면 하단으로 치우쳐 잘못된 위치에 렌더링되었으며, 'OLICE'로 잘려 보입니다.",
     "fix_en": "Remove the misplaced 'OLICE' text from the lower right of the jacket and place 'POLICE' correctly on the left chest. Preserve the character's face, cap, plaid shirt, the rest of the jacket, background river, sunset lighting, and framing.",
     "severity": "major",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "자켓 왼쪽 가슴에 있어야 할 'POLICE' 문구가 화면 하단으로 치우쳐 잘못된 위치에 렌더링되었으며, 'OLICE'로 잘려 보입니다.",
     "severity": "major"
    },
    {
     "issue_ko": "모자 정면의 '수사' 글자 위에 있어야 할 작은 영문 텍스트가 누락되었습니다.",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 0
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "자켓 왼쪽 가슴에 있어야 할 'POLICE' 문구가 화면 하단으로 치우쳐 잘못된 위치에 렌더링되었으며, 'OLICE'로 잘려 보입니다.",
    "fix_en": "Remove the misplaced 'OLICE' text from the lower right of the jacket and place 'POLICE' correctly on the left chest. Preserve the character's face, cap, plaid shirt, the rest of the jacket, background river, sunset lighting, and framing.",
    "severity": "major",
    "observation_index": 0
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S92sh4"
  }
 },
 "S92sh7::cine": {
  "applied": true,
  "fingerprint": "e0057253b45f5567a5c3233747ba88900f0ff2b8ca7f7ef1cfab65225205a1a9",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S92sh7_sel.png",
  "source_sha256": "25a471c0595454e071338b380192fb9a3e09e139c2aec55b925e70239e5bdfad",
  "file": "S92sh7_cine.png",
  "latency_ms": 12211
 },
 "S93sh1::signage": {
  "fp": "97862c8eb954047c",
  "inscriptions": [
   {
    "surface_native": "벽면 아크릴 현판",
    "text_native": "나주경찰서",
    "reason_ko": "나주경찰서장실이라는 구체적인 공간 배경을 시각적으로 명시하기 위해 벽면에 경찰서 명칭을 표시합니다."
   }
  ]
 },
 "S93sh1": {
  "input_fingerprint": "933102a94f98418e",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서장실 소파에 앉은 나주경찰서장을 중심으로, 양옆에 둘러앉은 전택수, 강주철, 서의용, 나상혁의 전경\n\nLOCATION (lock): In the sofa meeting area of the police chief's office, arranged around a low table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin a slow diagonal dolly-in from the open side of the seating area, slightly above seated eye level, establishing 나주경찰서장 at the head of the sofa and the four men distributed to either side. 전택수 leans into his challenge toward the chief, while 강주철, 서의용, and 나상혁 occupy distinct seated moments—forward-braced, tightly attentive, and drawn back uneasily—so the group reads as individuals rather than a posed lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: sofa seating area (occupied by the five men) — The open side faces the camera diagonally, with the chief at the head and the four officers extending to either side; used as Defines the head position of the chief and the opposing seated arrangement; table (positioned inside the seating arrangement) — Its near edge angles across the lower foreground toward the chief; used as Separates the camera-side foreground from the seated group; newspaper (placed on the table) — The printed front page lies visible on the tabletop, though the headline is not yet isolated; used as Secondary story clue awaiting the later downward move.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light appropriate to the office maintains moderate contrast and restrained procedural realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station-chief office, sofa arrangement, table, and daytime light. Exclude the earlier laptop broadcast and add the five officials seated around the chief.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper remains on the table in front of the seated group in the chief's office.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 아크릴 현판: \"나주경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서장실 소파에 앉은 나주경찰서장을 중심으로, 양옆에 둘러앉은 전택수, 강주철, 서의용, 나상혁의 전경\n\nLOCATION (lock): In the sofa meeting area of the police chief's office, arranged around a low table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin a slow diagonal dolly-in from the open side of the seating area, slightly above seated eye level, establishing 나주경찰서장 at the head of the sofa and the four men distributed to either side. 전택수 leans into his challenge toward the chief, while 강주철, 서의용, and 나상혁 occupy distinct seated moments—forward-braced, tightly attentive, and drawn back uneasily—so the group reads as individuals rather than a posed lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: sofa seating area (occupied by the five men) — The open side faces the camera diagonally, with the chief at the head and the four officers extending to either side; used as Defines the head position of the chief and the opposing seated arrangement; table (positioned inside the seating arrangement) — Its near edge angles across the lower foreground toward the chief; used as Separates the camera-side foreground from the seated group; newspaper (placed on the table) — The printed front page lies visible on the tabletop, though the headline is not yet isolated; used as Secondary story clue awaiting the later downward move.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light appropriate to the office maintains moderate contrast and restrained procedural realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station-chief office, sofa arrangement, table, and daytime light. Exclude the earlier laptop broadcast and add the five officials seated around the chief.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper remains on the table in front of the seated group in the chief's office.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 아크릴 현판: \"나주경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 서장실 소파에 앉은 나주경찰서장을 중심으로, 양옆에 둘러앉은 전택수, 강주철, 서의용, 나상혁의 전경\n\nLOCATION (lock): In the sofa meeting area of the police chief's office, arranged around a low table. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Begin a slow diagonal dolly-in from the open side of the seating area, slightly above seated eye level, establishing 나주경찰서장 at the head of the sofa and the four men distributed to either side. 전택수 leans into his challenge toward the chief, while 강주철, 서의용, and 나상혁 occupy distinct seated moments—forward-braced, tightly attentive, and drawn back uneasily—so the group reads as individuals rather than a posed lineup.\n- FRAMING SCALE: wide shot\n- KEY BACKGROUND ELEMENTS: sofa seating area (occupied by the five men) — The open side faces the camera diagonally, with the chief at the head and the four officers extending to either side; used as Defines the head position of the chief and the opposing seated arrangement; table (positioned inside the seating arrangement) — Its near edge angles across the lower foreground toward the chief; used as Separates the camera-side foreground from the seated group; newspaper (placed on the table) — The printed front page lies visible on the tabletop, though the headline is not yet isolated; used as Secondary story clue awaiting the later downward move.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient light appropriate to the office maintains moderate contrast and restrained procedural realism.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same station-chief office, sofa arrangement, table, and daytime light. Exclude the earlier laptop broadcast and add the five officials seated around the chief.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper remains on the table in front of the seated group in the chief's office.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 아크릴 현판: \"나주경찰서\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "A",
    "direction": "인물들이 서로를 향해 시선을 교환하며 대화하는 방향성을 보임.",
    "built_space": "경찰서장실 소파 공간. 5명의 남성이 중앙 테이블을 중심으로 둘러앉아 있음.",
    "entities": "나주경찰서장(반팔 경찰복)은 레퍼런스와 일치하나 중앙 상석이 아닌 우측에 앉아 있음. 나머지 4명(전택수, 서의용, 강주철, 나상혁)은 레퍼런스와 전혀 다른 복장과 외모(정장 및 긴팔 경찰복 등)를 한 임의의 인물들임. 테이블 위에 신문이 있고, 벽면에 '나주경찰서' 아크릴 현판이 존재함.",
    "hard_violations": [
     "이전 샷에서 고정(locked)된 필수 캐릭터 전택수 누락",
     "서의용, 강주철, 나상혁 누락 및 프롬프트에 없는 임의의 인물들 추가",
     "나주경찰서장이 소파의 중앙 상석(head of the sofa)에 위치하지 않음"
    ],
    "physics": "모든 인물이 소파에 정상적으로 앉아 있으며 지지 상태에 물리적 오류 없음."
   },
   {
    "label": "B",
    "direction": "인물들의 시선이 좌측 또는 중앙의 인물을 향해 집중되어 있음.",
    "built_space": "경찰서장실 소파 공간. 5명의 남성이 중앙 테이블을 중심으로 둘러앉아 있음.",
    "entities": "강주철(주황색/회색 재킷), 나주경찰서장(반팔 경찰복), 나상혁(남색 해링턴 재킷)의 외형은 레퍼런스와 일치함. 그러나 나주경찰서장이 중앙이 아닌 좌측에 앉아 있음. 전택수와 서의용은 보이지 않으며, 회색 정장과 베이지색 카디건을 입은 임의의 인물로 대체됨. 테이블 위 신문에 텍스트가 명확하게 보이며, 벽면에 '나주경찰서' 현판이 있음.",
    "hard_violations": [
     "이전 샷에서 고정(locked)된 필수 캐릭터 전택수 누락",
     "서의용 누락 및 프롬프트에 없는 임의의 인물들 추가",
     "나주경찰서장이 소파의 중앙 상석(head of the sofa)에 위치하지 않음"
    ],
    "physics": "모든 인물이 소파에 정상적으로 안착되어 있으며 물리적 결함은 없음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 3,
   "A": 2
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 3,
    "verdict_ko": "강주철과 나상혁 등 일부 인물의 외형은 일치하나, 필수 고정 캐릭터인 전택수와 서의용이 누락되어 프롬프트 지시를 심각하게 위반함."
   },
   {
    "label": "A",
    "score": 2,
    "verdict_ko": "고정 캐릭터인 전택수를 비롯해 4명의 지정 인물이 누락되고 임의의 인물들로 대체되어 프롬프트를 크게 위반함."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S33sh7_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "CHARACTER REFERENCE — 강주철: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:695371>"
   },
   {
    "label": "CHARACTER REFERENCE — 나상혁: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:747016>"
   },
   {
    "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832602>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "프롬프트에 지정된 5명의 인물 중 '전택수'와 '서의용'이 화면에 완전히 누락되었으며, 대신 레퍼런스에 존재하지 않는 회색 정장과 베이지색 가디건을 입은 두 명의 인물이 우측 소파에 임의로 생성되었습니다.",
     "fix_en": "Replace the man in the grey suit with 전택수 wearing his navy police jacket and cap, and replace the man in the beige cardigan with 서의용 wearing his black leather jacket and beanie. Preserve the other three men, their positions and clothing, the set, the lighting, the table with the newspaper, and the framing.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "나주경찰서장이 소파의 중심(상석)에 앉고 나머지 4명이 양옆에 둘러앉아야 한다는 배치 지시와 달리, 서장이 좌측 소파에 다른 인물과 나란히 앉아 있어 구도가 틀렸습니다.",
     "fix_en": "Relocate the chief to the head of the sofa and arrange the four men on either side.",
     "severity": "major",
     "observation_index": 1,
     "needs_regeneration": true
    },
    {
     "issue_ko": "이전 샷(PREVIOUS SHOT STILL)의 고정된 공간 특징을 유지해야 한다는 지시와 달리, 레퍼런스 사진에 존재하지 않던 대형 목재 책장과 다른 형태의 벽면 구조가 배경에 임의로 추가되었습니다.",
     "fix_en": "Restore the light grey walls and beige sofa from the reference, removing the wooden bookshelves.",
     "severity": "major",
     "observation_index": 2,
     "needs_regeneration": true
    },
    {
     "issue_ko": "벽면 현판에 지정된 텍스트('나주경찰서') 외에는 어떤 텍스트도 추가하지 말라는 지시를 어기고, 한글 텍스트 바로 위에 의미를 알 수 없는 작은 알파벳 문구들이 임의로 생성되었습니다.",
     "fix_en": "Remove the small illegible text above '나주경찰서', leaving a bare acrylic surface.",
     "severity": "major",
     "observation_index": 3
    },
    {
     "issue_ko": "전택수가 서장 쪽으로 몸을 기울여 따지는 자세가 없고 나머지 인물도 앞으로 버티거나 긴장하거나 불안하게 물러나지 않은 채 모두 손을 모아 비슷하게 앉아 있다",
     "fix_en": "Adjust the men's postures to lean, brace, or pull back uneasily instead of sitting uniformly with hands together.",
     "severity": "major",
     "observation_index": 7
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "프롬프트에 지정된 5명의 인물 중 '전택수'와 '서의용'이 화면에 완전히 누락되었으며, 대신 레퍼런스에 존재하지 않는 회색 정장과 베이지색 가디건을 입은 두 명의 인물이 우측 소파에 임의로 생성되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "나주경찰서장이 소파의 중심(상석)에 앉고 나머지 4명이 양옆에 둘러앉아야 한다는 배치 지시와 달리, 서장이 좌측 소파에 다른 인물과 나란히 앉아 있어 구도가 틀렸습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "이전 샷(PREVIOUS SHOT STILL)의 고정된 공간 특징을 유지해야 한다는 지시와 달리, 레퍼런스 사진에 존재하지 않던 대형 목재 책장과 다른 형태의 벽면 구조가 배경에 임의로 추가되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "벽면 현판에 지정된 텍스트('나주경찰서') 외에는 어떤 텍스트도 추가하지 말라는 지시를 어기고, 한글 텍스트 바로 위에 의미를 알 수 없는 작은 알파벳 문구들이 임의로 생성되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "전택수(수사 모자·남색 점퍼)가 없고 화면 가운데-오른쪽 회색 양복 남성이 다른 얼굴로 앉아 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "서의용(가죽 점퍼·비니)이 없고 맨 오른쪽 베이지 가디건 남성이 다른 인물로 앉아 있다",
     "severity": "critical"
    },
    {
     "issue_ko": "나주경찰서장이 소파 머리 중심이 아니라 왼쪽 암체어에 앉아 네 명이 양옆으로 둘러앉은 배치가 아니다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 서장 쪽으로 몸을 기울여 따지는 자세가 없고 나머지 인물도 앞으로 버티거나 긴장하거나 불안하게 물러나지 않은 채 모두 손을 모아 비슷하게 앉아 있다",
     "severity": "major"
    },
    {
     "issue_ko": "이전 스틸의 베이지 소파·창가 집무 책상·태극기와 달리 갈색 가죽 소파와 책장 위주로 공간이 바뀌었다",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 4,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "나주경찰서장이 소파의 중심(상석)에 앉고 나머지 4명이 양옆에 둘러앉아야 한다는 배치 지시와 달리, 서장이 좌측 소파에 다른 인물과 나란히 앉아 있어 구도가 틀렸습니다.",
    "fix_en": "Relocate the chief to the head of the sofa and arrange the four men on either side.",
    "severity": "major",
    "observation_index": 1,
    "needs_regeneration": true
   },
   {
    "issue_ko": "이전 샷(PREVIOUS SHOT STILL)의 고정된 공간 특징을 유지해야 한다는 지시와 달리, 레퍼런스 사진에 존재하지 않던 대형 목재 책장과 다른 형태의 벽면 구조가 배경에 임의로 추가되었습니다.",
    "fix_en": "Restore the light grey walls and beige sofa from the reference, removing the wooden bookshelves.",
    "severity": "major",
    "observation_index": 2,
    "needs_regeneration": true
   },
   {
    "issue_ko": "벽면 현판에 지정된 텍스트('나주경찰서') 외에는 어떤 텍스트도 추가하지 말라는 지시를 어기고, 한글 텍스트 바로 위에 의미를 알 수 없는 작은 알파벳 문구들이 임의로 생성되었습니다.",
    "fix_en": "Remove the small illegible text above '나주경찰서', leaving a bare acrylic surface.",
    "severity": "major",
    "observation_index": 3
   },
   {
    "issue_ko": "전택수가 서장 쪽으로 몸을 기울여 따지는 자세가 없고 나머지 인물도 앞으로 버티거나 긴장하거나 불안하게 물러나지 않은 채 모두 손을 모아 비슷하게 앉아 있다",
    "fix_en": "Adjust the men's postures to lean, brace, or pull back uneasily instead of sitting uniformly with hands together.",
    "severity": "major",
    "observation_index": 7
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 7,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the man in the grey suit with 전택수 wearing his navy police jacket and cap, and replace the man in the beige cardigan with 서의용 wearing his black leather jacket and beanie. Preserve the other three men, their positions and clothing, the set, the lighting, the table with the newspaper, and the framing.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 서장실 배경과 착석 구도를 잘 따랐으나, 필수 인물(전택수 등)이 누락되고 프롬프트에 없는 인물들이 다수 추가되어 실패함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "로케이션(서장실)을 완전히 무시했고, 전원 착석 지시를 어기고 기립한 상태로 연출하였으며 인물 구성도 틀림."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "인물들이 서로를 향해 시선을 두고 대화에 집중하고 있음.",
      "built_space": "이전 샷과 일치하는 밝은 서장실. 소파와 낮고 넓은 테이블이 올바르게 배치됨.",
      "entities": "강주철(주황색 재킷), 경찰서장(경찰복), 나상혁(남색 재킷)은 등장하나, 회색 정장과 베이지색 가디건을 입은 두 명의 창조된 인물이 포함됨. 전택수와 서의용 누락. 테이블 위 신문과 벽면 '나주경찰서' 현판이 정확함.",
      "hard_violations": [
       "지시되지 않은 창조된 인물 2명 추가"
      ],
      "physics": "모든 인물이 소파에 앉아 자연스럽게 지지받고 있음."
     },
     {
      "label": "B",
      "direction": "5명 모두 정면의 카메라 렌즈를 응시하고 있음.",
      "built_space": "지정된 밝은 서장실이 아닌 어둡고 낡은 창고 형태의 공간.",
      "entities": "서의용(가죽 재킷), 강주철(주황색 재킷) 외 3명은 완전히 창조된 인물임. 서장, 전택수, 나상혁 누락.",
      "hard_violations": [
       "지시되지 않은 창조된 인물 3명 추가",
       "착석 지시를 어기고 모두 기립한 상태(잘못된 스테이징)"
      ],
      "physics": "바닥을 딛고 서서 몸을 지지함."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지정된 서장실 배경과 착석 구도를 잘 따랐으나, 필수 인물(전택수 등)이 누락되고 프롬프트에 없는 인물들이 다수 추가되어 실패함."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "로케이션(서장실)을 완전히 무시했고, 전원 착석 지시를 어기고 기립한 상태로 연출하였으며 인물 구성도 틀림."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "인물들이 서로를 향해 시선을 두고 대화에 집중하고 있음.",
      "built_space": "이전 샷과 일치하는 밝은 서장실. 소파와 낮고 넓은 테이블이 올바르게 배치됨.",
      "entities": "강주철(주황색 재킷), 경찰서장(경찰복), 나상혁(남색 재킷)은 등장하나, 회색 정장과 베이지색 가디건을 입은 두 명의 창조된 인물이 포함됨. 전택수와 서의용 누락. 테이블 위 신문과 벽면 '나주경찰서' 현판이 정확함.",
      "hard_violations": [
       "지시되지 않은 창조된 인물 2명 추가"
      ],
      "physics": "모든 인물이 소파에 앉아 자연스럽게 지지받고 있음."
     },
     {
      "label": "B",
      "direction": "5명 모두 정면의 카메라 렌즈를 응시하고 있음.",
      "built_space": "지정된 밝은 서장실이 아닌 어둡고 낡은 창고 형태의 공간.",
      "entities": "서의용(가죽 재킷), 강주철(주황색 재킷) 외 3명은 완전히 창조된 인물임. 서장, 전택수, 나상혁 누락.",
      "hard_violations": [
       "지시되지 않은 창조된 인물 3명 추가",
       "착석 지시를 어기고 모두 기립한 상태(잘못된 스테이징)"
      ],
      "physics": "바닥을 딛고 서서 몸을 지지함."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 5,
      "verdict_ko": "지정된 서장실 공간, 소파에 앉은 구도, 현판 등은 우수하나, 필수 고정 인물(전택수)이 누락되고 미지정 인물이 추가되었습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "장소와 조명이 완전히 다르며, 소파에 앉은 구도를 무시하고 일렬로 선 채 미지정 인물까지 추가되어 지시를 크게 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "인물들이 서로를 바라보며 대화를 나누는 시선.",
      "built_space": "낮의 경찰서장실. 소파와 낮은 테이블이 배치되었고 벽에 '나주경찰서' 아크릴 현판이 있음.",
      "entities": "강주철, 나주경찰서장, 나상혁은 참조와 일치. 전택수와 서의용 누락. 회색 정장과 베이지 가디건을 입은 미지정 인물 2명 추가됨.",
      "hard_violations": [
       "이전 샷에 고정된 필수 인물(전택수) 누락",
       "명시되지 않은 허위 인물 2명 추가"
      ],
      "physics": "소파에 자연스럽게 앉아 체중이 지지됨."
     },
     {
      "label": "A",
      "direction": "모든 인물이 카메라 렌즈 정면을 응시함.",
      "built_space": "소파가 없는 어두운 실내, 나무 테이블 뒤에 일렬로 배치됨.",
      "entities": "서의용, 강주철은 참조와 유사. 경찰서장은 긴소매와 모자 등 제복 오류. 전택수 누락. 미지정 인물 2명 추가됨. 현판 없음.",
      "hard_violations": [
       "소파에 앉은 구도를 무시하고 전원 기립 (구도/배치 위반)",
       "명시되지 않은 허위 인물 2명 추가",
       "필수 인물 누락",
       "지정된 장소(경찰서장실) 및 조명 완전 불일치"
      ],
      "physics": "바닥에 두 발로 지지하며 서 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 5,
      "verdict_ko": "지정된 서장실 공간, 소파에 앉은 구도, 현판 등은 우수하나, 필수 고정 인물(전택수)이 누락되고 미지정 인물이 추가되었습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "장소와 조명이 완전히 다르며, 소파에 앉은 구도를 무시하고 일렬로 선 채 미지정 인물까지 추가되어 지시를 크게 위반했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "인물들이 서로를 바라보며 대화를 나누는 시선.",
      "built_space": "낮의 경찰서장실. 소파와 낮은 테이블이 배치되었고 벽에 '나주경찰서' 아크릴 현판이 있음.",
      "entities": "강주철, 나주경찰서장, 나상혁은 참조와 일치. 전택수와 서의용 누락. 회색 정장과 베이지 가디건을 입은 미지정 인물 2명 추가됨.",
      "hard_violations": [
       "이전 샷에 고정된 필수 인물(전택수) 누락",
       "명시되지 않은 허위 인물 2명 추가"
      ],
      "physics": "소파에 자연스럽게 앉아 체중이 지지됨."
     },
     {
      "label": "B",
      "direction": "모든 인물이 카메라 렌즈 정면을 응시함.",
      "built_space": "소파가 없는 어두운 실내, 나무 테이블 뒤에 일렬로 배치됨.",
      "entities": "서의용, 강주철은 참조와 유사. 경찰서장은 긴소매와 모자 등 제복 오류. 전택수 누락. 미지정 인물 2명 추가됨. 현판 없음.",
      "hard_violations": [
       "소파에 앉은 구도를 무시하고 전원 기립 (구도/배치 위반)",
       "명시되지 않은 허위 인물 2명 추가",
       "필수 인물 누락",
       "지정된 장소(경찰서장실) 및 조명 완전 불일치"
      ],
      "physics": "바닥에 두 발로 지지하며 서 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 9,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": true,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "needs_reshoot": true,
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S33sh7"
  }
 },
 "S93sh1::cine": {
  "applied": true,
  "fingerprint": "3db920daf4dafe9fba5e1399fb70c06f31c0ff2a5de8cb178a0b8675839211d6",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S93sh1_sel.png",
  "source_sha256": "73b8d021f0b420b4c847fcf63cf8515b789d471f145ac78653faabc6ab1dee01",
  "file": "S93sh1_cine.png",
  "latency_ms": 9649
 },
 "S93sh3::signage": {
  "fp": "2a01c7acc090df27",
  "inscriptions": [
   {
    "surface_native": "벽면 액자",
    "text_native": "신뢰받는 경찰",
    "reason_ko": "경찰서장실 소파 구역 뒤편 벽면에 경찰 표어가 적힌 액자를 배치하여 이곳이 경찰서장실임을 분명히 하고 서장의 권위와 대조되는 상황의 아이러니를 보여줍니다."
   }
  ]
 },
 "S93sh3": {
  "input_fingerprint": "743dd76ef7c1950d",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수를 향해 매서운 눈총을 쏘아보며 목에 핏대를 세운 나주경찰서장의 얼굴\n\nLOCATION (lock): At the head seat of the sofa meeting area inside the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly-in from just outside 전택수's shoulder, slightly below the seated chief's eye line, retaining only a narrow soft edge of 전택수 while tightening on 나주경찰서장's low three-quarter face. The chief braces forward at the sofa head, neck strained and eyes locked on 전택수, making the accusation read as directed confrontation rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: sofa head position (occupied by the police chief) — The chief-facing section remains behind his shoulders while most of the seating area is excluded; used as Limited contextual edge anchoring the chief's seated position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral office daylight keeps the chief's strained expression clear without heightened or sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나주경찰서장 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office, sofa seating, daylight, and chief's clothing. Exclude the other officials from the close frame and isolate the chief's angry glare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper remains on the table while the chief glares at Taeksu.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 액자: \"신뢰받는 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수를 향해 매서운 눈총을 쏘아보며 목에 핏대를 세운 나주경찰서장의 얼굴\n\nLOCATION (lock): At the head seat of the sofa meeting area inside the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly-in from just outside 전택수's shoulder, slightly below the seated chief's eye line, retaining only a narrow soft edge of 전택수 while tightening on 나주경찰서장's low three-quarter face. The chief braces forward at the sofa head, neck strained and eyes locked on 전택수, making the accusation read as directed confrontation rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: sofa head position (occupied by the police chief) — The chief-facing section remains behind his shoulders while most of the seating area is excluded; used as Limited contextual edge anchoring the chief's seated position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral office daylight keeps the chief's strained expression clear without heightened or sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나주경찰서장 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office, sofa seating, daylight, and chief's clothing. Exclude the other officials from the close frame and isolate the chief's angry glare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper remains on the table while the chief glares at Taeksu.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 액자: \"신뢰받는 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수를 향해 매서운 눈총을 쏘아보며 목에 핏대를 세운 나주경찰서장의 얼굴\n\nLOCATION (lock): At the head seat of the sofa meeting area inside the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Finish the dolly-in from just outside 전택수's shoulder, slightly below the seated chief's eye line, retaining only a narrow soft edge of 전택수 while tightening on 나주경찰서장's low three-quarter face. The chief braces forward at the sofa head, neck strained and eyes locked on 전택수, making the accusation read as directed confrontation rather than a look into the lens.\n- FRAMING SCALE: close-up\n- KEY BACKGROUND ELEMENTS: sofa head position (occupied by the police chief) — The chief-facing section remains behind his shoulders while most of the seating area is excluded; used as Limited contextual edge anchoring the chief's seated position.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral office daylight keeps the chief's strained expression clear without heightened or sensational contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 나주경찰서장 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same office, sofa seating, daylight, and chief's clothing. Exclude the other officials from the close frame and isolate the chief's angry glare.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper remains on the table while the chief glares at Taeksu.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 벽면 액자: \"신뢰받는 경찰\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "initial_roll_all_fail": true,
  "readings": [
   {
    "label": "B",
    "direction": "메인 피사체(회색 정장 남성)가 왼쪽 전경에 있는 인물(비니 쓴 남성)을 향해 매서운 눈빛을 겨냥하고 있습니다.",
    "built_space": "경찰서장 집무실 배경으로 소파, 목재 캐비닛, 벽면 간판이 배치되어 있으며, 인물은 소파 상석에 올바르게 앉아 있습니다.",
    "entities": "지시된 나주경찰서장(제복 입은 인물) 대신 이전 컷의 회색 정장 인물을 메인 피사체로 잘못 생성했습니다. 전택수(비니와 주황색 점퍼)는 전경에 일부 잘려서 나타납니다. 벽면 간판 텍스트 '신뢰받는 경찰'은 정확하게 렌더링되었습니다.",
    "hard_violations": [
     "지시된 나주경찰서장의 신원(제복)을 무시하고, 이전 컷에서 금지된 다른 인물(회색 정장)의 얼굴과 의상을 메인 피사체로 복제함 (invented people/wrong identity)."
    ],
    "physics": "메인 피사체는 소파에 엉덩이를 대고 앞으로 몸을 기울인 자세를 안정적으로 유지하고 있습니다."
   },
   {
    "label": "A",
    "direction": "메인 피사체(회색 정장 남성)가 화면 왼쪽 전경의 인물을 노려보고 있습니다.",
    "built_space": "집무실 내 소파와 벽면 간판, 캐비닛이 존재하며, 피사체는 소파에 위치해 있습니다.",
    "entities": "나주경찰서장 캐릭터 레퍼런스(제복)를 반영하지 않고 이전 컷의 회색 정장 인물을 등장시켰습니다. 전경의 전택수는 비니와 주황색 점퍼를 입고 있습니다. 벽면 간판에 '신뢰받는 경찰' 텍스트 위에 불필요한 깨진 문자가 포함되어 있습니다.",
    "hard_violations": [
     "지시된 인물이 아닌 이전 컷의 다른 인물(회색 정장)의 얼굴과 의상을 그대로 가져옴 (invented people/wrong identity).",
     "벽면 간판에 해독할 수 없는 문자가 누출됨 (leaked text)."
    ],
    "physics": "메인 피사체는 소파에 무게를 싣고 상체를 앞으로 굽히고 있으며 물리적으로 자연스럽습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 4,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 4,
    "verdict_ko": "두 후보 모두 캐릭터 레퍼런스(제복 입은 서장)를 무시하고 이전 컷의 회색 정장 인물을 메인 피사체로 잘못 렌더링하여 심각한 신원 오류가 있으나, B가 간판 텍스트를 더 깔끔하게 처리했습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "캐릭터 레퍼런스의 제복 입은 인물 대신 금지된 회색 정장 인물의 외모와 의상을 그대로 가져오는 치명적인 신원 오류를 범했으며, 벽면 간판에 깨진 텍스트가 추가되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 나주경찰서장 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S93sh1_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 나주경찰서장: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:832602>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 중앙의 주요 인물(나주경찰서장)이 지정된 캐릭터 레퍼런스의 제복 입은 인물이 아닌, 이전 장면 참고 사진 속 회색 정장을 입은 다른 인물로 완전히 잘못 묘사되었습니다.",
     "fix_en": "Replace the central man in the grey suit with the police chief from the character reference, wearing a dark blue short-sleeved police uniform shirt with a black tie. Maintain the man's angry expression, his forward-leaning posture with a strained neck, the camera framing, the lighting, the background with the framed sign, and the blurry foreground character.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "왼쪽 전택수 뒷모습이 좁은 가장자리가 아니라 프레임 왼쪽을 크게 차지한다.",
     "fix_en": "Reduce the mass of the out-of-focus foreground character on the left so they form only a narrow soft edge along the frame, revealing more of the background behind them. Do not change the central character, his expression, the lighting, or the camera angle.",
     "severity": "major",
     "observation_index": 2
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 중앙의 주요 인물(나주경찰서장)이 지정된 캐릭터 레퍼런스의 제복 입은 인물이 아닌, 이전 장면 참고 사진 속 회색 정장을 입은 다른 인물로 완전히 잘못 묘사되었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "화면 중앙 인물이 참조·이전 장면의 나주경찰서장(제복·짧은 머리)이 아니라 맞은편 회색 양복 남성이다.",
     "severity": "critical"
    },
    {
     "issue_ko": "왼쪽 전택수 뒷모습이 좁은 가장자리가 아니라 프레임 왼쪽을 크게 차지한다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 1,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "왼쪽 전택수 뒷모습이 좁은 가장자리가 아니라 프레임 왼쪽을 크게 차지한다.",
    "fix_en": "Reduce the mass of the out-of-focus foreground character on the left so they form only a narrow soft edge along the frame, revealing more of the background behind them. Do not change the central character, his expression, the lighting, or the camera angle.",
    "severity": "major",
    "observation_index": 2
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the central man in the grey suit with the police chief from the character reference, wearing a dark blue short-sleeved police uniform shirt with a black tie. Maintain the man's angry expression, his forward-leaning posture with a strained neck, the camera framing, the lighting, the background with the framed sign, and the blurry foreground character.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 클로즈업, 오버더숄더(OTS) 앵글, 그리고 상체를 기울인 채 노려보는 분노의 감정 표현은 완벽하게 연출했으나, 경찰서장의 신원을 레퍼런스(경찰 제복 인물)가 아닌 이전 컷의 회색 정장 인물로 잘못 식별한 치명적인 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "요구된 클로즈업 및 앵글을 완전히 무시하고 이전 컷의 와이드 샷을 그대로 반복했으며, 제외해야 할 추가 인물들을 모두 포함시켰고 의상마저 제멋대로 변경하여 지시사항을 모두 위반했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "회색 정장을 입은 남성이 화면 왼쪽 앞부분에 걸쳐진 비니 쓴 남성(전택수)을 향해 맹렬하게 시선을 고정하고 노려봄.",
      "built_space": "사무실 내부. 인물이 소파에 앉아 있으며, 배경 벽면에 '신뢰받는 경찰' 텍스트가 정확히 적힌 액자가 보임.",
      "entities": "지시된 카메라 프레이밍(전택수의 어깨를 걸친 서장의 클로즈업)은 훌륭히 구현되었으나, 서장의 신원이 캐릭터 레퍼런스(제복 입은 남성)와 전혀 다른 인물(이전 컷의 회색 정장 남성)로 렌더링됨.",
      "hard_violations": [],
      "physics": "상체를 앞으로 강하게 기울이고 목에 힘을 준 긴장된 자세가 시트 위에 물리적으로 자연스럽게 지탱됨."
     },
     {
      "label": "B",
      "direction": "5명의 인물이 소파에 앉아 대기하거나 서로를 향해 시선을 분산하고 있음. 특정인을 향한 강한 시선 집중이 없음.",
      "built_space": "사무실 소파 구역 전체가 보이며, 인물들이 제자리에 착석해 있음. 테이블, 신문, 책장 등이 보임.",
      "entities": "프롬프트에서 배제하도록 명시한 3명의 추가 인물이 모두 등장함. 또한 회색 정장을 입어야 할 인물이 경찰 제복으로 변경되어 제복 입은 사람이 2명이 되는 등 신원과 의상이 완전히 꼬임.",
      "hard_violations": [
       "프롬프트가 제외하라고 지시한 다수의 추가 인물 포함 (extra bodies)",
       "명시된 클로즈업 카메라 앵글 및 프레이밍 완전 무시"
      ],
      "physics": "인물들이 소파에 앉아 있는 자세와 체중의 분산은 물리적으로 정상임."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 4,
      "verdict_ko": "지시된 클로즈업, 오버더숄더(OTS) 앵글, 그리고 상체를 기울인 채 노려보는 분노의 감정 표현은 완벽하게 연출했으나, 경찰서장의 신원을 레퍼런스(경찰 제복 인물)가 아닌 이전 컷의 회색 정장 인물로 잘못 식별한 치명적인 오류가 있습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "요구된 클로즈업 및 앵글을 완전히 무시하고 이전 컷의 와이드 샷을 그대로 반복했으며, 제외해야 할 추가 인물들을 모두 포함시켰고 의상마저 제멋대로 변경하여 지시사항을 모두 위반했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "회색 정장을 입은 남성이 화면 왼쪽 앞부분에 걸쳐진 비니 쓴 남성(전택수)을 향해 맹렬하게 시선을 고정하고 노려봄.",
      "built_space": "사무실 내부. 인물이 소파에 앉아 있으며, 배경 벽면에 '신뢰받는 경찰' 텍스트가 정확히 적힌 액자가 보임.",
      "entities": "지시된 카메라 프레이밍(전택수의 어깨를 걸친 서장의 클로즈업)은 훌륭히 구현되었으나, 서장의 신원이 캐릭터 레퍼런스(제복 입은 남성)와 전혀 다른 인물(이전 컷의 회색 정장 남성)로 렌더링됨.",
      "hard_violations": [],
      "physics": "상체를 앞으로 강하게 기울이고 목에 힘을 준 긴장된 자세가 시트 위에 물리적으로 자연스럽게 지탱됨."
     },
     {
      "label": "B",
      "direction": "5명의 인물이 소파에 앉아 대기하거나 서로를 향해 시선을 분산하고 있음. 특정인을 향한 강한 시선 집중이 없음.",
      "built_space": "사무실 소파 구역 전체가 보이며, 인물들이 제자리에 착석해 있음. 테이블, 신문, 책장 등이 보임.",
      "entities": "프롬프트에서 배제하도록 명시한 3명의 추가 인물이 모두 등장함. 또한 회색 정장을 입어야 할 인물이 경찰 제복으로 변경되어 제복 입은 사람이 2명이 되는 등 신원과 의상이 완전히 꼬임.",
      "hard_violations": [
       "프롬프트가 제외하라고 지시한 다수의 추가 인물 포함 (extra bodies)",
       "명시된 클로즈업 카메라 앵글 및 프레이밍 완전 무시"
      ],
      "physics": "인물들이 소파에 앉아 있는 자세와 체중의 분산은 물리적으로 정상임."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "카메라 앵글, 클로즈업 프레이밍, 액션 지시(오버더숄더에서의 노려보는 연기)는 훌륭하게 구현했으나, 레퍼런스의 경찰서장(제복)이 아닌 회색 정장 인물을 피사체로 삼아 신원 일치에서 치명적으로 실패했습니다."
     },
     {
      "label": "A",
      "score": 1,
      "verdict_ko": "클로즈업 지시를 무시하고 5명을 모두 노출시켰으며, 엉뚱한 인물에게 경찰 제복을 입혀 서장을 두 명으로 만드는 등 신원과 프레이밍 모두 실패했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "회색 정장의 남성이 카메라 앞 전경에 걸친 어깨(전택수) 너머를 향해 날카로운 눈총을 쏘아봄.",
      "built_space": "경찰서장실 내부. 피사체 뒤쪽 벽면에 '신뢰받는 경찰' 액자가 위치함.",
      "entities": "주 피사체로 클로즈업된 인물이 캐릭터 레퍼런스의 나주경찰서장(제복 남성)이 아니라 이전 컷의 회색 정장 남성이며 얼굴과 복장 모두 불일치함. 전경의 어깨는 주황색 점퍼를 입은 전택수로 보임.",
      "hard_violations": [
       "주 피사체의 신원 불일치: 캐릭터 레퍼런스의 나주경찰서장(제복 착용)이 아닌 전혀 다른 인물(회색 정장 남성)을 클로즈업함.",
       "복장 위반: 서장의 복장(경찰 제복)이 유지되지 않음."
      ],
      "physics": "상체를 앞으로 강하게 기울인 자세이며, 보이지 않는 하체와 의자에 의해 지지됨."
     },
     {
      "label": "A",
      "direction": "중앙에 위치한 남성(제복으로 변경됨)이 왼쪽의 비니 쓴 남성을 향해 노려보며, 좌측의 실제 서장은 중앙을 응시함.",
      "built_space": "경찰서장실 소파 공간에 5명이 이전 컷과 같은 배치로 앉아 있음.",
      "entities": "5명이 모두 등장함. 프롬프트가 지정한 나주경찰서장(좌측 두 번째 제복 남성)이 주된 행동을 하지 않음. 대신 중앙의 정장 남성이 경찰 제복으로 복장이 바뀌어 서장 역할을 대신하고 있음.",
      "hard_violations": [
       "지시문에 없는 인물 포함: 서장의 1인 클로즈업을 지시하고 나머지를 제외하라고 했으나 5명 모두 등장함.",
       "프레임 스케일 위반: 클로즈업이 아닌 와이드샷.",
       "신원 및 복장 위반: 레퍼런스의 서장이 아닌 다른 인물에게 제복을 입히고 주된 행동을 묘사함."
      ],
      "physics": "인물들 모두 소파에 체중을 싣고 앉아 있음."
     }
    ],
    "all_candidates_fail": true,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "카메라 앵글, 클로즈업 프레이밍, 액션 지시(오버더숄더에서의 노려보는 연기)는 훌륭하게 구현했으나, 레퍼런스의 경찰서장(제복)이 아닌 회색 정장 인물을 피사체로 삼아 신원 일치에서 치명적으로 실패했습니다."
     },
     {
      "label": "B",
      "score": 1,
      "verdict_ko": "클로즈업 지시를 무시하고 5명을 모두 노출시켰으며, 엉뚱한 인물에게 경찰 제복을 입혀 서장을 두 명으로 만드는 등 신원과 프레이밍 모두 실패했습니다."
     }
    ],
    "all_candidates_fail": true,
    "readings": [
     {
      "label": "A",
      "direction": "회색 정장의 남성이 카메라 앞 전경에 걸친 어깨(전택수) 너머를 향해 날카로운 눈총을 쏘아봄.",
      "built_space": "경찰서장실 내부. 피사체 뒤쪽 벽면에 '신뢰받는 경찰' 액자가 위치함.",
      "entities": "주 피사체로 클로즈업된 인물이 캐릭터 레퍼런스의 나주경찰서장(제복 남성)이 아니라 이전 컷의 회색 정장 남성이며 얼굴과 복장 모두 불일치함. 전경의 어깨는 주황색 점퍼를 입은 전택수로 보임.",
      "hard_violations": [
       "주 피사체의 신원 불일치: 캐릭터 레퍼런스의 나주경찰서장(제복 착용)이 아닌 전혀 다른 인물(회색 정장 남성)을 클로즈업함.",
       "복장 위반: 서장의 복장(경찰 제복)이 유지되지 않음."
      ],
      "physics": "상체를 앞으로 강하게 기울인 자세이며, 보이지 않는 하체와 의자에 의해 지지됨."
     },
     {
      "label": "B",
      "direction": "중앙에 위치한 남성(제복으로 변경됨)이 왼쪽의 비니 쓴 남성을 향해 노려보며, 좌측의 실제 서장은 중앙을 응시함.",
      "built_space": "경찰서장실 소파 공간에 5명이 이전 컷과 같은 배치로 앉아 있음.",
      "entities": "5명이 모두 등장함. 프롬프트가 지정한 나주경찰서장(좌측 두 번째 제복 남성)이 주된 행동을 하지 않음. 대신 중앙의 정장 남성이 경찰 제복으로 복장이 바뀌어 서장 역할을 대신하고 있음.",
      "hard_violations": [
       "지시문에 없는 인물 포함: 서장의 1인 클로즈업을 지시하고 나머지를 제외하라고 했으나 5명 모두 등장함.",
       "프레임 스케일 위반: 클로즈업이 아닌 와이드샷.",
       "신원 및 복장 위반: 레퍼런스의 서장이 아닌 다른 인물에게 제복을 입히고 주된 행동을 묘사함."
      ],
      "physics": "인물들 모두 소파에 체중을 싣고 앉아 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 7,
     "B": 2
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S93sh1"
  }
 },
 "S93sh3::cine": {
  "applied": true,
  "fingerprint": "1bec8d633e3483d71c13af06775fbc5b2f9b409a60e3961cd50126b21694b364",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S93sh3_sel.png",
  "source_sha256": "c87e0b46f4dd4044213cae161044bb4edc41fbe29df25d7fc48706d9ca66915e",
  "file": "S93sh3_cine.png",
  "latency_ms": 13383
 },
 "S93sh5::signage": {
  "fp": "d4e613f919d7cf08",
  "inscriptions": [
   {
    "surface_native": "신문 1면 헤드라인",
    "text_native": "나주 드들강 여고생 피살 사건, 미궁에 빠진 수사",
    "reason_ko": "경찰서장실 탁자 위 신문 헤드라인을 클로즈업하는 장면이므로 사건의 배경이 되는 실제 뉴스를 시각적으로 전달하기 위해 필요합니다."
   }
  ]
 },
 "S93sh5": {
  "input_fingerprint": "d2dca26b88d7ad55",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 신문 1면에 실린 '나주 드들강 여고생 살인사건' 머리기사 텍스트 클로즈업\n\nLOCATION (lock): On the low table between the sofas in the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane downward just above the table's near edge at a steep oblique high angle, adopting 전택수's lowered visual path without showing him. Isolate the printed headline, “나주 드들강 여고생 살인사건, 16년 만에 단죄,” while retaining only a narrow margin of the surrounding front page and tabletop for physical context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: newspaper front page (lying on the table) — The upward-facing printed side is visible obliquely, with the headline “나주 드들강 여고생 살인사건, 16년 만에 단죄” held in focus; used as Primary isolated evidence text at the end of the descending camera move; table (supporting the newspaper) — The near edge crosses the lower margin beneath the front page; used as Narrow contextual border around the newspaper.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime office illumination keeps the printed text legible within a subdued investigative-drama palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper lies open to the front-page headline announcing that the Drdeul River schoolgirl murder was punished after sixteen years.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 신문 1면 헤드라인: \"나주 드들강 여고생 피살 사건, 미궁에 빠진 수사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 신문 1면에 실린 '나주 드들강 여고생 살인사건' 머리기사 텍스트 클로즈업\n\nLOCATION (lock): On the low table between the sofas in the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane downward just above the table's near edge at a steep oblique high angle, adopting 전택수's lowered visual path without showing him. Isolate the printed headline, “나주 드들강 여고생 살인사건, 16년 만에 단죄,” while retaining only a narrow margin of the surrounding front page and tabletop for physical context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: newspaper front page (lying on the table) — The upward-facing printed side is visible obliquely, with the headline “나주 드들강 여고생 살인사건, 16년 만에 단죄” held in focus; used as Primary isolated evidence text at the end of the descending camera move; table (supporting the newspaper) — The near edge crosses the lower margin beneath the front page; used as Narrow contextual border around the newspaper.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime office illumination keeps the printed text legible within a subdued investigative-drama palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper lies open to the front-page headline announcing that the Drdeul River schoolgirl murder was punished after sixteen years.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 신문 1면 헤드라인: \"나주 드들강 여고생 피살 사건, 미궁에 빠진 수사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 테이블 위 신문 1면에 실린 '나주 드들강 여고생 살인사건' 머리기사 텍스트 클로즈업\n\nLOCATION (lock): On the low table between the sofas in the police chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Complete the crane downward just above the table's near edge at a steep oblique high angle, adopting 전택수's lowered visual path without showing him. Isolate the printed headline, “나주 드들강 여고생 살인사건, 16년 만에 단죄,” while retaining only a narrow margin of the surrounding front page and tabletop for physical context.\n- FRAMING SCALE: insert close-up on a detail\n- KEY BACKGROUND ELEMENTS: newspaper front page (lying on the table) — The upward-facing printed side is visible obliquely, with the headline “나주 드들강 여고생 살인사건, 16년 만에 단죄” held in focus; used as Primary isolated evidence text at the end of the descending camera move; table (supporting the newspaper) — The near edge crosses the lower margin beneath the front page; used as Narrow contextual border around the newspaper.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime office illumination keeps the printed text legible within a subdued investigative-drama palette.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established — the place, its fixed features and wear, each person's clothing and state — persists. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The newspaper lies open to the front-page headline announcing that the Drdeul River schoolgirl murder was punished after sixteen years.\n\nPEOPLE: the SHOT TEXT alone decides who is visible in this shot. People known to appear somewhere in this scene: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리); 나주경찰서장 (Korean 남성, 50대 초반, 넓은 얼굴, 짧게 정돈된 검은색 머리); 나상혁 (Korean 남성, 30대 초반, 해사한 얼굴, 단정한 짧은 검은색 머리); 강주철 (Korean 남성, 50대 초반, 둥근 얼굴, 짧은 검은색 머리); 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리). That list is scene-level, not a cast list for this frame — it may name someone this shot does not show, and it may omit someone this shot does show. If the shot text names a person who is not on the list, draw that person exactly as the shot text describes them; the list does not override the shot text. Never add a person the shot text does not show. IF only part of a person is in frame (a hand, arm, foot, back, silhouette), that body part belongs to the specific person the shot text names — its sex, age, build, skin and grooming must unmistakably match that person.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 신문 1면 헤드라인: \"나주 드들강 여고생 피살 사건, 미궁에 빠진 수사\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "카메라가 테이블 위 평평하게 놓인 신문을 높은 각도에서 비스듬히 내려다봅니다.",
    "built_space": "경찰서장실의 목재 테이블과 뒤쪽의 어두운 가죽 소파가 보이며, 레퍼런스의 공간 구조와 일치합니다.",
    "entities": "신문 지면 위에 '나주 드들강 여고생 살인사건 16년 만에 단죄'와 '나주 드들강 여고생 피살 사건, 미궁에 빠진 수사' 텍스트가 모두 명확하게 인쇄되어 있습니다.",
    "hard_violations": [],
    "physics": "신문이 테이블 표면에 중력을 받아 자연스럽고 평평하게 놓여 있습니다."
   },
   {
    "label": "A",
    "direction": "카메라가 테이블 위 두껍게 접힌 신문을 높은 각도에서 내려다봅니다.",
    "built_space": "테이블, 뒤쪽 가죽 소파, 그리고 벽에 걸린 '신뢰받는 경찰' 액자가 보여 레퍼런스의 배경 요소를 잘 유지했습니다.",
    "entities": "신문에 '나주 드들강 여고생 살인사건, 16년 만에 단죄' 텍스트만 존재하며, 프롬프트의 필수 사인 텍스트가 누락되었습니다.",
    "hard_violations": [],
    "physics": "신문이 테이블 위에 놓여 있으나, 종이의 두께나 접힌 형태가 지나치게 단단한 블록처럼 부자연스럽습니다."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 4
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "요구된 두 가지 헤드라인 텍스트를 모두 포함하였으며, 신문의 질감과 형태를 사실적으로 잘 구현했습니다."
   },
   {
    "label": "A",
    "score": 4,
    "verdict_ko": "카메라 텍스트는 정확하나 필수 사인 텍스트(미궁에 빠진 수사)가 누락되었고, 신문이 빈 종이 블록처럼 부자연스럽습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features, lighting mood and each person's clothing are LOCKED to this photo; never copy its camera framing. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S93sh3_sel.png"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "1면 큰 제목은 「나주 드들강 여고생 살인사건」「16년 만에 단죄」이고, 지정 헤드라인 「나주 드들강 여고생 피살 사건, 미궁에 빠진 수사」는 아래 보조 제목으로 같이 있어 문구가 이중·모순이다.",
     "fix_en": "Change the large top headline to read exactly '나주 드들강 여고생 피살 사건, 미궁에 빠진 수사' and turn the sub-headline below it into illegible, blurry text lines. Preserve the newspaper, the table surface, and the lighting exactly as they are.",
     "severity": "major",
     "observation_index": 0
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "1면 큰 제목은 「나주 드들강 여고생 살인사건」「16년 만에 단죄」이고, 지정 헤드라인 「나주 드들강 여고생 피살 사건, 미궁에 빠진 수사」는 아래 보조 제목으로 같이 있어 문구가 이중·모순이다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 0,
    "openrouter:x-ai/grok-4.6": 1
   }
  },
  "fix_severity_skipped_count": 1,
  "fix_severity_skipped": [
   {
    "issue_ko": "1면 큰 제목은 「나주 드들강 여고생 살인사건」「16년 만에 단죄」이고, 지정 헤드라인 「나주 드들강 여고생 피살 사건, 미궁에 빠진 수사」는 아래 보조 제목으로 같이 있어 문구가 이중·모순이다.",
    "fix_en": "Change the large top headline to read exactly '나주 드들강 여고생 피살 사건, 미궁에 빠진 수사' and turn the sub-headline below it into illegible, blurry text lines. Preserve the newspaper, the table surface, and the lighting exactly as they are.",
    "severity": "major",
    "observation_index": 0
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev만 (배경 전용·공유 계획)",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S93sh3"
  },
  "lane_policy": "share_plan_prev_bgonly"
 },
 "S93sh5::cine": {
  "applied": true,
  "fingerprint": "2678396e2a673bdbec740be9472240c266fd28a006ac11f8c5b10701f5fbc0fd",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S93sh5_sel.png",
  "source_sha256": "33b8ea677ec591070b23ef2aac3de0abda725c14e91fcba0d7ec7cb61eba3c30",
  "file": "S93sh5_cine.png",
  "latency_ms": 13510
 },
 "S94sh6::signage": {
  "fp": "46286a6eb9f27e9f",
  "inscriptions": []
 },
 "S94sh6": {
  "input_fingerprint": "adf30bb89959713f",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 낡은 지갑 안에서 손가락으로 흑백사진을 절반쯤 빼낸 채 쥐고 있는 전택수의 손 클로즈업\n\nLOCATION (lock): In the police-station corridor immediately outside the chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height on 전택수's 서의용-facing side, tighten the downward tilt into an extreme close view of his fingers drawing the black-and-white photograph halfway from the open worn wallet. The hands and photograph carry the frame while a narrow strip of his clothing and the corridor beyond preserve human scale; the outward direction of the photograph anticipates its transfer to 서의용.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: black-and-white photograph in the middle-right of the frame, foreground, moves toward 서의용 outside the tight frame.\n- KEY BACKGROUND ELEMENTS: black-and-white photograph (being drawn halfway out) — Its image-bearing face is angled obliquely toward the camera and remains only halfway exposed from the wallet; used as Primary focal object emerging between 전택수's fingers; 전택수's wallet (worn and open) — The opened inner side faces obliquely toward the camera at waist height; used as Frames and partially conceals the photograph; police-station corridor (daytime interior); used as Minimal soft background preserving the location of the hand action.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient illumination appropriate to the corridor renders the hands, worn wallet, and monochrome photograph with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same police-station corridor, institutional walls, daylight, and the senior officer's clothing. Exclude the chief and reporter, focusing on the wallet and partially withdrawn photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's old wallet is open, and the same black-and-white photograph of the school-uniformed girl is halfway removed from its sleeve.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 낡은 지갑 안에서 손가락으로 흑백사진을 절반쯤 빼낸 채 쥐고 있는 전택수의 손 클로즈업\n\nLOCATION (lock): In the police-station corridor immediately outside the chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height on 전택수's 서의용-facing side, tighten the downward tilt into an extreme close view of his fingers drawing the black-and-white photograph halfway from the open worn wallet. The hands and photograph carry the frame while a narrow strip of his clothing and the corridor beyond preserve human scale; the outward direction of the photograph anticipates its transfer to 서의용.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: black-and-white photograph in the middle-right of the frame, foreground, moves toward 서의용 outside the tight frame.\n- KEY BACKGROUND ELEMENTS: black-and-white photograph (being drawn halfway out) — Its image-bearing face is angled obliquely toward the camera and remains only halfway exposed from the wallet; used as Primary focal object emerging between 전택수's fingers; 전택수's wallet (worn and open) — The opened inner side faces obliquely toward the camera at waist height; used as Frames and partially conceals the photograph; police-station corridor (daytime interior); used as Minimal soft background preserving the location of the hand action.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient illumination appropriate to the corridor renders the hands, worn wallet, and monochrome photograph with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same police-station corridor, institutional walls, daylight, and the senior officer's clothing. Exclude the chief and reporter, focusing on the wallet and partially withdrawn photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's old wallet is open, and the same black-and-white photograph of the school-uniformed girl is halfway removed from its sleeve.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 자신의 낡은 지갑 안에서 손가락으로 흑백사진을 절반쯤 빼낸 채 쥐고 있는 전택수의 손 클로즈업\n\nLOCATION (lock): In the police-station corridor immediately outside the chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: At waist height on 전택수's 서의용-facing side, tighten the downward tilt into an extreme close view of his fingers drawing the black-and-white photograph halfway from the open worn wallet. The hands and photograph carry the frame while a narrow strip of his clothing and the corridor beyond preserve human scale; the outward direction of the photograph anticipates its transfer to 서의용.\n- FRAMING SCALE: close-up\n- FRAME LAYOUT: black-and-white photograph in the middle-right of the frame, foreground, moves toward 서의용 outside the tight frame.\n- KEY BACKGROUND ELEMENTS: black-and-white photograph (being drawn halfway out) — Its image-bearing face is angled obliquely toward the camera and remains only halfway exposed from the wallet; used as Primary focal object emerging between 전택수's fingers; 전택수's wallet (worn and open) — The opened inner side faces obliquely toward the camera at waist height; used as Frames and partially conceals the photograph; police-station corridor (daytime interior); used as Minimal soft background preserving the location of the hand action.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Neutral daytime ambient illumination appropriate to the corridor renders the hands, worn wallet, and monochrome photograph with restrained contrast.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same police-station corridor, institutional walls, daylight, and the senior officer's clothing. Exclude the chief and reporter, focusing on the wallet and partially withdrawn photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Taeksu's old wallet is open, and the same black-and-white photograph of the school-uniformed girl is halfway removed from its sleeve.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 전택수 right now — 전택수's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 전택수 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "두 손이 지갑과 사진을 쥐고 있으며, 사진의 앞면은 카메라를 비스듬히 향함.",
    "built_space": "경찰서 복도가 부드러운 아웃포커싱 배경으로 처리됨.",
    "entities": "전택수의 상의(레퍼런스와 텍스트 배열 다름), 체크 셔츠, 낡은 가죽 지갑, 소녀의 흑백 사진이 명확히 식별됨.",
    "hard_violations": [],
    "physics": "양손이 지갑과 사진을 물리적으로 자연스럽게 쥐고 지탱함."
   },
   {
    "label": "B",
    "direction": "손들이 지갑과 사진을 조작 중이며, 사진이 카메라를 향해 정렬됨.",
    "built_space": "흐릿한 경찰서 복도 배경이 프레임 뒤쪽에 위치함.",
    "entities": "경찰 재킷, 체크 셔츠, 낡은 지갑, 흑백 사진이 레퍼런스와 일치하게 나타나나, 총 3개의 손이 등장함.",
    "hard_violations": [
     "화면에 세 개의 손(왼쪽, 오른쪽 위, 오른쪽 아래)이 등장하여 해부학적으로 불가능한 신체(extra bodies) 구조를 보임."
    ],
    "physics": "세 개의 손이 각각 지갑의 양쪽과 사진의 윗부분을 쥐고 지탱함."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "A",
    "openrouter:x-ai/grok-4.6": "A"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "A": 6,
   "B": 2
  },
  "selected": "A",
  "ranking": [
   "A",
   "B"
  ],
  "verdicts": [
   {
    "label": "A",
    "score": 6,
    "verdict_ko": "요청된 익스트림 클로즈업보다 프레이밍이 다소 넓고 재킷의 텍스트가 레퍼런스와 다르지만, 치명적인 오류 없이 주요 피사체들을 무난하게 구현했습니다."
   },
   {
    "label": "B",
    "score": 2,
    "verdict_ko": "사진과 지갑의 묘사는 우수하나, 화면에 세 개의 손이 등장하는 치명적인 해부학적 오류(하드 위반)가 발생하여 최하위로 평가되었습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The clothing, hair and overall look of 전택수 — who appear both in that photo and in this shot — are LOCKED to that photo. Anyone else visible in that photo is NOT in this shot: never carry their face, body or clothing onto anyone here. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S51sh7_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   },
   {
    "label": "PROP REFERENCE — 낡은 가죽 지갑: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:950222>"
   },
   {
    "label": "PROP REFERENCE — 교복 입은 소녀의 흑백사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1073154>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "의상 불일치: 재킷 오른쪽 가슴에 레퍼런스('수사관')와 전혀 다른 커다란 '경찰 POLICE' 텍스트가 적혀 있습니다.",
     "fix_en": "Replace the large '경찰 POLICE' text on the jacket's right chest with a small text patch. Preserve hands, wallet, photograph, jacket, lighting, and background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "소품 상태 불일치: 흑백사진이 지갑 포켓에서 절반쯤 빼낸 상태가 아니라, 포켓 밖으로 완전히 꺼내져 지갑 안쪽 면에 온전히 노출된 채 놓여 있습니다.",
     "fix_en": "Redraw the photograph so its lower half is inside the wallet's sleeve, halfway pulled out. Preserve hands, jacket, wallet, lighting, background, and the girl's face.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "손가락·사진 극근접이 아니라 상의 가슴부터 허리까지 넓게 담겨 의복이 좁은 띠로 보이지 않는다.",
     "fix_en": "Crop to an extreme close-up of hands and wallet.",
     "severity": "major",
     "observation_index": 3,
     "needs_regeneration": true
    },
    {
     "issue_ko": "배경이 이전 컷의 계단과 강력계 표지판이 있는 복도가 아니라 나무문이 있는 다른 복도이다.",
     "fix_en": "Replace background with blurred stairs and sign.",
     "severity": "major",
     "observation_index": 4,
     "needs_regeneration": true
    },
    {
     "issue_ko": "흑백사진이 프레임 중우측이 아니라 한가운데에 놓여 있다.",
     "fix_en": "Crop left to shift the photograph middle-right.",
     "severity": "major",
     "observation_index": 6,
     "needs_regeneration": true
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "의상 불일치: 재킷 오른쪽 가슴에 레퍼런스('수사관')와 전혀 다른 커다란 '경찰 POLICE' 텍스트가 적혀 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "소품 상태 불일치: 흑백사진이 지갑 포켓에서 절반쯤 빼낸 상태가 아니라, 포켓 밖으로 완전히 꺼내져 지갑 안쪽 면에 온전히 노출된 채 놓여 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "흑백사진이 지갑 슬리브에서 절반만 빠져 나온 채가 아니라 완전히 분리되어 지갑 앞 중앙에서 정면으로 들려 있다.",
     "severity": "critical"
    },
    {
     "issue_ko": "손가락·사진 극근접이 아니라 상의 가슴부터 허리까지 넓게 담겨 의복이 좁은 띠로 보이지 않는다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경이 이전 컷의 계단과 강력계 표지판이 있는 복도가 아니라 나무문이 있는 다른 복도이다.",
     "severity": "major"
    },
    {
     "issue_ko": "재킷 글자가 이전 컷의 수사/POLICE가 아닌 경찰 POLICE이고 바지가 어두운색이 아닌 카키색이다.",
     "severity": "major"
    },
    {
     "issue_ko": "흑백사진이 프레임 중우측이 아니라 한가운데에 놓여 있다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 5
   }
  },
  "fix_severity_skipped_count": 3,
  "fix_severity_skipped": [
   {
    "issue_ko": "손가락·사진 극근접이 아니라 상의 가슴부터 허리까지 넓게 담겨 의복이 좁은 띠로 보이지 않는다.",
    "fix_en": "Crop to an extreme close-up of hands and wallet.",
    "severity": "major",
    "observation_index": 3,
    "needs_regeneration": true
   },
   {
    "issue_ko": "배경이 이전 컷의 계단과 강력계 표지판이 있는 복도가 아니라 나무문이 있는 다른 복도이다.",
    "fix_en": "Replace background with blurred stairs and sign.",
    "severity": "major",
    "observation_index": 4,
    "needs_regeneration": true
   },
   {
    "issue_ko": "흑백사진이 프레임 중우측이 아니라 한가운데에 놓여 있다.",
    "fix_en": "Crop left to shift the photograph middle-right.",
    "severity": "major",
    "observation_index": 6,
    "needs_regeneration": true
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 5,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Replace the large '경찰 POLICE' text on the jacket's right chest with a small text patch. Preserve hands, wallet, photograph, jacket, lighting, and background.\n- Redraw the photograph so its lower half is inside the wallet's sleeve, halfway pulled out. Preserve hands, jacket, wallet, lighting, background, and the girl's face.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 클로즈업 앵글과 경찰서 복도 배경, 지갑 및 흑백 사진의 디테일을 충실히 구현했으나, 재킷에 지침에 없는 커다란 텍스트 패치가 추가된 점은 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "샷 텍스트가 요구한 손 클로즈업 프레이밍과 경찰서 복도 배경을 완전히 무시한 채 캐릭터 레퍼런스를 그대로 복사해 단순 스튜디오 배경에 배치했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "사진이 바깥쪽을 향해 비스듬히 당겨져 나오고 있으며, 손가락이 정확히 사진을 향해 작용하고 있습니다.",
      "built_space": "아웃포커싱된 경찰서 복도 내부로, 문과 벽면의 구조가 흐릿하게 식별됩니다.",
      "entities": "레퍼런스와 일치하는 낡은 가죽 지갑, 교복 입은 소녀의 흑백 사진, 그리고 전택수의 의상(체크 셔츠와 남색 재킷)이 확인됩니다.",
      "hard_violations": [],
      "physics": "양손이 지갑을 안정적으로 쥐고 있으며, 엄지와 검지가 사진을 물리적으로 꼬집어 당기고 있는 형태가 자연스럽게 지탱됩니다."
     },
     {
      "label": "B",
      "direction": "인물이 카메라 렌즈를 정면으로 똑바로 응시하고 있습니다.",
      "built_space": "건축물이나 배경 요소가 전혀 없는 평면적인 회색 스튜디오 배경입니다.",
      "entities": "전택수의 얼굴과 의상, 지갑, 사진이 등장하지만 재킷의 텍스트가 'SALICE'로 왜곡되어 있습니다.",
      "hard_violations": [],
      "physics": "양손이 지갑의 양옆을 쥐고 공중에 든 상태로, 손의 위치 자체는 지갑을 지탱하고 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "요구된 클로즈업 앵글과 경찰서 복도 배경, 지갑 및 흑백 사진의 디테일을 충실히 구현했으나, 재킷에 지침에 없는 커다란 텍스트 패치가 추가된 점은 아쉽습니다."
     },
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "샷 텍스트가 요구한 손 클로즈업 프레이밍과 경찰서 복도 배경을 완전히 무시한 채 캐릭터 레퍼런스를 그대로 복사해 단순 스튜디오 배경에 배치했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "A",
      "direction": "사진이 바깥쪽을 향해 비스듬히 당겨져 나오고 있으며, 손가락이 정확히 사진을 향해 작용하고 있습니다.",
      "built_space": "아웃포커싱된 경찰서 복도 내부로, 문과 벽면의 구조가 흐릿하게 식별됩니다.",
      "entities": "레퍼런스와 일치하는 낡은 가죽 지갑, 교복 입은 소녀의 흑백 사진, 그리고 전택수의 의상(체크 셔츠와 남색 재킷)이 확인됩니다.",
      "hard_violations": [],
      "physics": "양손이 지갑을 안정적으로 쥐고 있으며, 엄지와 검지가 사진을 물리적으로 꼬집어 당기고 있는 형태가 자연스럽게 지탱됩니다."
     },
     {
      "label": "B",
      "direction": "인물이 카메라 렌즈를 정면으로 똑바로 응시하고 있습니다.",
      "built_space": "건축물이나 배경 요소가 전혀 없는 평면적인 회색 스튜디오 배경입니다.",
      "entities": "전택수의 얼굴과 의상, 지갑, 사진이 등장하지만 재킷의 텍스트가 'SALICE'로 왜곡되어 있습니다.",
      "hard_violations": [],
      "physics": "양손이 지갑의 양옆을 쥐고 공중에 든 상태로, 손의 위치 자체는 지갑을 지탱하고 있습니다."
     }
    ]
   },
   "reverse_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "요구된 익스트림 클로즈업 샷 크기를 무시하고 인물의 상반신 전체를 보여주는 미디엄 샷으로 렌더링했으며, 지정된 경찰서 복도 배경이 아닌 스튜디오 배경을 사용하여 지시를 크게 위반했습니다."
     },
     {
      "label": "B",
      "score": 7,
      "verdict_ko": "손과 지갑, 흑백사진에 집중하는 허리 높이의 클로즈업 샷 크기를 완벽하게 구현했으며, 흐릿한 복도 배경과 의상, 소품의 디테일까지 프롬프트와 레퍼런스를 충실히 반영했습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "정면을 향해 서서 카메라를 응시하고 있음.",
      "built_space": "지정된 경찰서 복도가 아닌 텅 빈 회색 스튜디오 배경임.",
      "entities": "전택수의 상반신, 낡은 가죽 지갑, 소녀의 흑백사진. 의상은 레퍼런스와 비슷하나 'SALICE'라는 잘못된 텍스트가 적혀 있음.",
      "hard_violations": [
       "지정된 샷 크기(손 중심의 클로즈업) 위반 (미디엄 샷으로 렌더링됨)",
       "지정된 장소(경찰서 복도) 위반 (스튜디오 배경으로 렌더링됨)"
      ],
      "physics": "양손이 지갑의 양쪽을 쥐고 있음."
     },
     {
      "label": "B",
      "direction": "카메라가 허리 높이에서 지갑과 사진을 내려다보고 있음.",
      "built_space": "아웃포커싱된 경찰서 복도의 문과 벽면이 배경에 보임.",
      "entities": "전택수의 몸통과 양손(경찰 재킷과 체크 셔츠 일치), 낡은 가죽 지갑(레퍼런스 일치), 교복 입은 소녀의 흑백사진(레퍼런스 일치).",
      "hard_violations": [],
      "physics": "왼손은 지갑을 받치고 오른손 엄지와 검지로 사진을 쥐어 빼내고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 3,
      "verdict_ko": "요구된 익스트림 클로즈업 샷 크기를 무시하고 인물의 상반신 전체를 보여주는 미디엄 샷으로 렌더링했으며, 지정된 경찰서 복도 배경이 아닌 스튜디오 배경을 사용하여 지시를 크게 위반했습니다."
     },
     {
      "label": "A",
      "score": 7,
      "verdict_ko": "손과 지갑, 흑백사진에 집중하는 허리 높이의 클로즈업 샷 크기를 완벽하게 구현했으며, 흐릿한 복도 배경과 의상, 소품의 디테일까지 프롬프트와 레퍼런스를 충실히 반영했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "정면을 향해 서서 카메라를 응시하고 있음.",
      "built_space": "지정된 경찰서 복도가 아닌 텅 빈 회색 스튜디오 배경임.",
      "entities": "전택수의 상반신, 낡은 가죽 지갑, 소녀의 흑백사진. 의상은 레퍼런스와 비슷하나 'SALICE'라는 잘못된 텍스트가 적혀 있음.",
      "hard_violations": [
       "지정된 샷 크기(손 중심의 클로즈업) 위반 (미디엄 샷으로 렌더링됨)",
       "지정된 장소(경찰서 복도) 위반 (스튜디오 배경으로 렌더링됨)"
      ],
      "physics": "양손이 지갑의 양쪽을 쥐고 있음."
     },
     {
      "label": "A",
      "direction": "카메라가 허리 높이에서 지갑과 사진을 내려다보고 있음.",
      "built_space": "아웃포커싱된 경찰서 복도의 문과 벽면이 배경에 보임.",
      "entities": "전택수의 몸통과 양손(경찰 재킷과 체크 셔츠 일치), 낡은 가죽 지갑(레퍼런스 일치), 교복 입은 소녀의 흑백사진(레퍼런스 일치).",
      "hard_violations": [],
      "physics": "왼손은 지갑을 받치고 오른손 엄지와 검지로 사진을 쥐어 빼내고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 14,
     "B": 6
    },
    "ranking": [
     "A",
     "B"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "A",
   "fix_won": false,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S51sh7"
  }
 },
 "S94sh6::cine": {
  "applied": true,
  "fingerprint": "36be489afb80748eb4d101780a884d07e521cab702f7b91781138e36fbce45a6",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S94sh6_sel.png",
  "source_sha256": "9ea6a54d1d6aa94fdfbbbad86481562fb298273d494d6f6303f21fbbf87cd27b",
  "file": "S94sh6_cine.png",
  "latency_ms": 12660
 },
 "S94sh8::signage": {
  "fp": "a863875071159581",
  "inscriptions": [
   {
    "surface_native": "서장실 문 옆 표지판",
    "text_native": "서장실",
    "reason_ko": "경찰서장실 앞 복도라는 공간적 배경을 명확히 나타내기 위해 문 옆의 부서 표지판에 표기가 필요합니다."
   }
  ]
 },
 "S94sh8": {
  "input_fingerprint": "85fb4836a00e7a16",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수가 건넨 흑백사진을 두 손으로 쥔 채 멍하니 내려다보는 서의용의 상체\n\nLOCATION (lock): In the corridor outside the police chief's office, where the photograph has just been handed over. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close to 서의용's front three-quarter side, the camera finishes its tilt at slightly above chest height with a mild downward angle, framing his upper torso while his two hands and the black-and-white photograph remain at the lower center. 서의용 occupies most of the center frame, shoulders slack and attention fixed downward; the emphasized shift is the closer distance into his silent reaction, with 전택수 kept outside the frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 나주서 서장실 앞 복도 (Daytime police-station corridor) — The camera is oriented along a short section of the corridor behind 서의용 rather than opening its full depth; used as The visible corridor margins provide restrained spatial context around the isolated reaction; 흑백사진 (Held in both hands) — The image-bearing face is angled upward enough to be visible to the camera, while its specific pictured content remains unspecified; used as Held at the lower center as the visual anchor for 서의용's downward gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime interior ambient illumination is rendered with a subdued neutral-to-cool palette, moderate contrast, and naturalistic facial tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor finishes, daylight, and the detective's clothing. Exclude the chief, reporter, and wallet, showing the detective holding the transferred photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The black-and-white photograph has been removed from Taeksu's old wallet and transferred into Euiyong's hands.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서장실 문 옆 표지판: \"서장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수가 건넨 흑백사진을 두 손으로 쥔 채 멍하니 내려다보는 서의용의 상체\n\nLOCATION (lock): In the corridor outside the police chief's office, where the photograph has just been handed over. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close to 서의용's front three-quarter side, the camera finishes its tilt at slightly above chest height with a mild downward angle, framing his upper torso while his two hands and the black-and-white photograph remain at the lower center. 서의용 occupies most of the center frame, shoulders slack and attention fixed downward; the emphasized shift is the closer distance into his silent reaction, with 전택수 kept outside the frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 나주서 서장실 앞 복도 (Daytime police-station corridor) — The camera is oriented along a short section of the corridor behind 서의용 rather than opening its full depth; used as The visible corridor margins provide restrained spatial context around the isolated reaction; 흑백사진 (Held in both hands) — The image-bearing face is angled upward enough to be visible to the camera, while its specific pictured content remains unspecified; used as Held at the lower center as the visual anchor for 서의용's downward gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime interior ambient illumination is rendered with a subdued neutral-to-cool palette, moderate contrast, and naturalistic facial tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor finishes, daylight, and the detective's clothing. Exclude the chief, reporter, and wallet, showing the detective holding the transferred photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The black-and-white photograph has been removed from Taeksu's old wallet and transferred into Euiyong's hands.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서장실 문 옆 표지판: \"서장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 전택수가 건넨 흑백사진을 두 손으로 쥔 채 멍하니 내려다보는 서의용의 상체\n\nLOCATION (lock): In the corridor outside the police chief's office, where the photograph has just been handed over. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: From close to 서의용's front three-quarter side, the camera finishes its tilt at slightly above chest height with a mild downward angle, framing his upper torso while his two hands and the black-and-white photograph remain at the lower center. 서의용 occupies most of the center frame, shoulders slack and attention fixed downward; the emphasized shift is the closer distance into his silent reaction, with 전택수 kept outside the frame.\n- FRAMING SCALE: medium shot\n- FRAME LAYOUT: 서의용 in the middle-center of the frame, midground.\n- KEY BACKGROUND ELEMENTS: 나주서 서장실 앞 복도 (Daytime police-station corridor) — The camera is oriented along a short section of the corridor behind 서의용 rather than opening its full depth; used as The visible corridor margins provide restrained spatial context around the isolated reaction; 흑백사진 (Held in both hands) — The image-bearing face is angled upward enough to be visible to the camera, while its specific pictured content remains unspecified; used as Held at the lower center as the visual anchor for 서의용's downward gaze.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Daytime interior ambient illumination is rendered with a subdued neutral-to-cool palette, moderate contrast, and naturalistic facial tones.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same corridor finishes, daylight, and the detective's clothing. Exclude the chief, reporter, and wallet, showing the detective holding the transferred photograph.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): The black-and-white photograph has been removed from Taeksu's old wallet and transferred into Euiyong's hands.\n\nTHE HAND THAT IS DOING THIS: the central object is being used by 서의용 right now — 서의용's hand (with as much wrist and forearm as the framing reaches) grips or rests on it exactly as real use requires, and that hand's size, build, skin, age and sleeve belong to 서의용 and no one else. The framing stays on the object.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 서의용 (Korean 남성, 40대 중반, 넓은 얼굴, 짧은 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서장실 문 옆 표지판: \"서장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "B",
    "direction": "서의용의 시선이 양손에 든 흑백사진을 향해 아래로 고정되어 있음.",
    "built_space": "경찰서 복도로, 배경에 문과 이전 샷 레퍼런스와 일치하는 파란색 바탕의 표지판이 배치됨.",
    "entities": "서의용은 얼굴과 의상(비니 포함) 모두 레퍼런스와 일치함. 손에 든 흑백사진의 내용이 레퍼런스와 동일하며, 표지판에 '서장실' 텍스트가 정확히 표기됨.",
    "hard_violations": [],
    "physics": "양손이 사진의 양옆을 안정적으로 파지하고 있으며 물리적으로 자연스러움."
   },
   {
    "label": "A",
    "direction": "서의용의 시선이 양손에 든 흑백사진을 향해 아래로 고정되어 있음.",
    "built_space": "경찰서 복도로, 배경에 문과 흰색 바탕의 표지판이 배치됨(레퍼런스와 색상 다름).",
    "entities": "서의용의 얼굴과 가죽 재킷은 일치하나 레퍼런스에 있는 비니를 착용하지 않음. 흑백사진과 '서장실' 텍스트는 지시사항을 따름.",
    "hard_violations": [],
    "physics": "양손으로 사진을 쥐고 있는 자세가 물리적으로 큰 무리 없이 묘사됨."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 8,
   "A": 5
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 8,
    "verdict_ko": "캐릭터의 의상(비니 포함)과 소품을 정확히 구현했으며, 이전 샷 레퍼런스의 배경 디테일(파란색 표지판)까지 충실히 반영함."
   },
   {
    "label": "A",
    "score": 5,
    "verdict_ko": "지정된 샷의 구도와 행동은 따랐으나, 캐릭터 레퍼런스의 주요 의상(비니)을 누락했고 배경 표지판의 색상이 레퍼런스와 다름."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S94sh6_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 서의용: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:929743>"
   },
   {
    "label": "PROP REFERENCE — 교복 입은 소녀의 흑백사진: the exact object appearing in this shot; match its look, material and wear exactly.",
    "path": "<bytes:1073154>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "우측 배경의 파란색 표지판에 프롬프트가 지시한 '서장실' 텍스트 아래에 지시되지 않은 정체불명의 문자가 추가되어 있습니다.",
     "fix_en": "Erase the unreadable characters below the '서장실' text on the blue sign, leaving that lower section blank. Preserve Eui-yong, his pose, his clothing, the photograph, and the rest of the corridor.",
     "severity": "minor",
     "observation_index": 0
    },
    {
     "issue_ko": "배경 우측 문의 손잡이 위치가 이전 샷 레퍼런스에서는 문 왼쪽에 있었으나, 생성된 이미지에서는 오른쪽으로 바뀌어 공간의 연속성이 어긋납니다.",
     "fix_en": "Relocate the door handle on the background wooden door to its left side. Preserve Eui-yong, his clothing, the held photograph, and the surrounding walls.",
     "severity": "minor",
     "observation_index": 1
    },
    {
     "issue_ko": "서의용이 이전 스틸에서 가져오라는 남색 경찰 점퍼가 아니라 가죽 재킷과 비니를 입고 있다.",
     "fix_en": "Replace Eui-yong's leather jacket and beanie with the navy blue police jacket from the previous shot reference. Preserve his face, his hand position, the black-and-white photograph, and the background corridor.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "배경 복도의 투톤 벽과 유리창 달린 문 마감이 이전 스틸에 고정된 장소와 다르다.",
     "fix_en": "Redraw the background door as a solid wooden door without a glass panel and adjust the wall's paint tones to match the previous shot reference exactly. Preserve Eui-yong, his clothing, his hands, the photograph, and the camera framing.",
     "severity": "major",
     "observation_index": 3
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "우측 배경의 파란색 표지판에 프롬프트가 지시한 '서장실' 텍스트 아래에 지시되지 않은 정체불명의 문자가 추가되어 있습니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "배경 우측 문의 손잡이 위치가 이전 샷 레퍼런스에서는 문 왼쪽에 있었으나, 생성된 이미지에서는 오른쪽으로 바뀌어 공간의 연속성이 어긋납니다.",
     "severity": "minor"
    },
    {
     "issue_ko": "서의용이 이전 스틸에서 가져오라는 남색 경찰 점퍼가 아니라 가죽 재킷과 비니를 입고 있다.",
     "severity": "major"
    },
    {
     "issue_ko": "배경 복도의 투톤 벽과 유리창 달린 문 마감이 이전 스틸에 고정된 장소와 다르다.",
     "severity": "major"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 2,
    "openrouter:x-ai/grok-4.6": 2
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "우측 배경의 파란색 표지판에 프롬프트가 지시한 '서장실' 텍스트 아래에 지시되지 않은 정체불명의 문자가 추가되어 있습니다.",
    "fix_en": "Erase the unreadable characters below the '서장실' text on the blue sign, leaving that lower section blank. Preserve Eui-yong, his pose, his clothing, the photograph, and the rest of the corridor.",
    "severity": "minor",
    "observation_index": 0
   },
   {
    "issue_ko": "배경 우측 문의 손잡이 위치가 이전 샷 레퍼런스에서는 문 왼쪽에 있었으나, 생성된 이미지에서는 오른쪽으로 바뀌어 공간의 연속성이 어긋납니다.",
    "fix_en": "Relocate the door handle on the background wooden door to its left side. Preserve Eui-yong, his clothing, the held photograph, and the surrounding walls.",
    "severity": "minor",
    "observation_index": 1
   },
   {
    "issue_ko": "서의용이 이전 스틸에서 가져오라는 남색 경찰 점퍼가 아니라 가죽 재킷과 비니를 입고 있다.",
    "fix_en": "Replace Eui-yong's leather jacket and beanie with the navy blue police jacket from the previous shot reference. Preserve his face, his hand position, the black-and-white photograph, and the background corridor.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "배경 복도의 투톤 벽과 유리창 달린 문 마감이 이전 스틸에 고정된 장소와 다르다.",
    "fix_en": "Redraw the background door as a solid wooden door without a glass panel and adjust the wall's paint tones to match the previous shot reference exactly. Preserve Eui-yong, his clothing, his hands, the photograph, and the camera framing.",
    "severity": "major",
    "observation_index": 3
   }
  ],
  "fix_skipped": true,
  "fix_skip_reason": "no_critical_issue",
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S94sh6"
  }
 },
 "S94sh8::cine": {
  "applied": true,
  "fingerprint": "e326b5c1ad10b4f3759a4240adea4c4176ab7097dc12181cf7638c11d4f40d4a",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S94sh8_sel.png",
  "source_sha256": "1846d9cddb00bbc52c95456ab82a0aafe0b9e5ec0c5bf2c30ada78b41dae38bd",
  "file": "S94sh8_cine.png",
  "latency_ms": 12755
 },
 "S94sh11::signage": {
  "fp": "ba2149ce07b5b211",
  "inscriptions": [
   {
    "surface_native": "서장실 문 표지판",
    "text_native": "서장실",
    "reason_ko": "경찰서 서장실 밖 복도라는 설정을 시각적으로 명확하게 전달하기 위해 서장실 표지판이 필요합니다."
   }
  ]
 },
 "S94sh11": {
  "input_fingerprint": "5341963ac9557433",
  "prompt": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 한가운데 뒤돌아선 채 서의용과 나상혁을 향해 따뜻한 미소를 짓는 전택수의 전경\n\nLOCATION (lock): Farther along the police-station corridor outside the chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside and slightly behind 서의용 and 나상혁 at below-shoulder height, the dolly-out settles into a stable wide frame looking down the corridor, with their partial backs anchoring the lower left and lower right while 전택수 appears several meters ahead near the center. The sole emphasized change from the preceding reaction shot is camera distance: the broadened view holds as 전택수 completes his turn, shifts his weight back, and smiles toward 서의용, while both foreground men lift their attention to him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, background, looks toward 서의용 and 나상혁; 서의용 in the lower-left of the frame, foreground, looks toward 전택수; 나상혁 in the lower-right of the frame, foreground, looks toward 전택수.\n- KEY BACKGROUND ELEMENTS: 나주서 서장실 앞 복도 (Daytime police-station corridor) — The camera looks along the corridor's length, with 전택수 centered on its receding axis; used as Its receding layout separates the two foreground observers from 전택수 and gives the final turn visual closure; 흑백사진 (Still held by 서의용) — The image-bearing face remains angled upward but is only partly visible from behind 서의용; used as A small continuity detail held low by 서의용, secondary to 전택수's smile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime interior ambience maintains subdued neutral-to-cool color, moderate contrast, and a humane naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same long station corridor, institutional daylight, and the three investigators' clothing. Exclude the chief and reporter, showing the senior officer turned back with a warm smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong continues holding the black-and-white photograph Taeksu entrusted to him as Taeksu turns back and smiles.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서장실 문 표지판: \"서장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
  "roll_prompts": {
   "A": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 한가운데 뒤돌아선 채 서의용과 나상혁을 향해 따뜻한 미소를 짓는 전택수의 전경\n\nLOCATION (lock): Farther along the police-station corridor outside the chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside and slightly behind 서의용 and 나상혁 at below-shoulder height, the dolly-out settles into a stable wide frame looking down the corridor, with their partial backs anchoring the lower left and lower right while 전택수 appears several meters ahead near the center. The sole emphasized change from the preceding reaction shot is camera distance: the broadened view holds as 전택수 completes his turn, shifts his weight back, and smiles toward 서의용, while both foreground men lift their attention to him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, background, looks toward 서의용 and 나상혁; 서의용 in the lower-left of the frame, foreground, looks toward 전택수; 나상혁 in the lower-right of the frame, foreground, looks toward 전택수.\n- KEY BACKGROUND ELEMENTS: 나주서 서장실 앞 복도 (Daytime police-station corridor) — The camera looks along the corridor's length, with 전택수 centered on its receding axis; used as Its receding layout separates the two foreground observers from 전택수 and gives the final turn visual closure; 흑백사진 (Still held by 서의용) — The image-bearing face remains angled upward but is only partly visible from behind 서의용; used as A small continuity detail held low by 서의용, secondary to 전택수's smile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime interior ambience maintains subdued neutral-to-cool color, moderate contrast, and a humane naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same long station corridor, institutional daylight, and the three investigators' clothing. Exclude the chief and reporter, showing the senior officer turned back with a warm smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong continues holding the black-and-white photograph Taeksu entrusted to him as Taeksu turns back and smiles.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서장실 문 표지판: \"서장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.",
   "B": "Create ONE FINAL photorealistic live-action film still of the moment below — South Korea, 2001–2017; all people are Korean unless stated. TIME OF DAY (lock): day.\n\nSHOT TEXT (authoritative, Korean): 복도 한가운데 뒤돌아선 채 서의용과 나상혁을 향해 따뜻한 미소를 짓는 전택수의 전경\n\nLOCATION (lock): Farther along the police-station corridor outside the chief's office. The shot takes place here — the attached PREVIOUS SHOT STILL shows this exact place.\n\nCAMERA & FRAME (framing authority; references give identity and place, not framing):\n- CAMERA: Beside and slightly behind 서의용 and 나상혁 at below-shoulder height, the dolly-out settles into a stable wide frame looking down the corridor, with their partial backs anchoring the lower left and lower right while 전택수 appears several meters ahead near the center. The sole emphasized change from the preceding reaction shot is camera distance: the broadened view holds as 전택수 completes his turn, shifts his weight back, and smiles toward 서의용, while both foreground men lift their attention to him.\n- FRAMING SCALE: wide shot\n- FRAME LAYOUT: 전택수 in the middle-center of the frame, background, looks toward 서의용 and 나상혁; 서의용 in the lower-left of the frame, foreground, looks toward 전택수; 나상혁 in the lower-right of the frame, foreground, looks toward 전택수.\n- KEY BACKGROUND ELEMENTS: 나주서 서장실 앞 복도 (Daytime police-station corridor) — The camera looks along the corridor's length, with 전택수 centered on its receding axis; used as Its receding layout separates the two foreground observers from 전택수 and gives the final turn visual closure; 흑백사진 (Still held by 서의용) — The image-bearing face remains angled upward but is only partly visible from behind 서의용; used as A small continuity detail held low by 서의용, secondary to 전택수's smile.\nFollow exactly: angle, subject scale, placement. Keep true people/background scale: fixtures, far objects take only their size/distance share, never foreground-sized.\n\nLIGHTING & MOOD (these govern light, color and surface rendering):\n- LIGHTING & MOOD: Consistent daytime interior ambience maintains subdued neutral-to-cool color, moderate contrast, and a humane naturalistic tonal range.\n- MATERIAL REALISM: every object and surface must read as a real,\n  physical material — correct texture, weight, wear and light response\n  (metal reflects, fabric drapes and creases, liquid is glossy and\n  pools, painted or drawn marks sit ON a surface and follow its\n  curvature and lighting). Nothing may look like a flat sticker, a\n  doodle or a graphic overlay pasted onto the frame.\n\nTHIS SHOT CONTINUES THE PREVIOUS SHOT: everything the attached still established about the place — its fixed features, wear and lighting — persists. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. Any person in it who cannot move stays PRECISELY as photographed (body, pose, contact points, held objects); only the camera changes.\n\nPREVIOUS STILL USAGE (follow exactly — what to take from the attached still and what to exclude): Take the same long station corridor, institutional daylight, and the three investigators' clothing. Exclude the chief and reporter, showing the senior officer turned back with a warm smile.\n\nREALIZE FIGURATIVE LANGUAGE: figurative or exaggerated shot text becomes what a real movie camera records on a physical set — staging, light, distance, angle — never literal fantasy imagery. Every character is a real, fully-formed human (build, face, hands, clothing); never a shape, solid silhouette or abstract mass, unless the story explicitly declares a non-human being.\n\nEXPRESSIONS ARE ACTED, NEVER ANATOMICAL: emotional or figurative descriptions of eyes, face or presence are an actor's performance — realize them only through gaze direction, focus, eyelids, facial muscles and posture. Eyes stay anatomically normal human eyes; never whiten, blank, cloud, glow or enlarge eyeballs, skin or anatomy unless the story declares the being non-human.\n\nPROPS FACE THE RIGHT WAY: a handheld or used object is oriented as its real use requires — its functional side (screen, front, page) faces its user's eyes; the camera sees whatever side the staging geometry implies. Never flip, mirror or reverse an object's front and back.\n\nNATURAL PERFORMANCE (default only — every explicit direction above wins): absent a pose, immobility or shot-text direction, people are alive in mid-moment — weight shifted, hands engaged with something real, gaze on the target the SHOT TEXT implies; no stiff attention stance, no blank stare into the lens. If the text stages stillness, sleep, death, restraint, drill or a direct-to-camera look, follow that exactly.\n\nDRAWN MARKS KEEP THEIR SHAPE: a mark described as drawn, painted or traced is a stroke on the surface — an open outline of brush width with the surface showing inside; never filled into a solid disc unless the text says filled.\n\nBODY & SUPPORT (default; direction above wins): seated = hips on seat, back to backrest, legs to floor, one per adjacent seat side by side; hands, hips, feet contact their support; frozen, stunned or held still stops mid-action as caught, weight on one side, hands where movement left them, never symmetric at-attention; a staged arrangement above or in SHOT TEXT wins.\n\nCARRIED STATE (persist exactly — must match the neighbouring shots of this scene): Euiyong continues holding the black-and-white photograph Taeksu entrusted to him as Taeksu turns back and smiles.\n\nPEOPLE: the SHOT TEXT alone decides who is visible. Any person shown must be one of: 전택수 (Korean 남성, 50대 중반, 각진 얼굴, 짧게 정돈된 검은색 머리) — never anyone else; never add a person the shot text does not show. Any partial body part in frame belongs to the person the text names — sex, age, build, skin and grooming must match.\n\nCHARACTER REFERENCE ROLE (follow exactly): refs = identity only: face, hair, build, clothing. Pose, gaze, angle, framing follow this prompt's CAMERA/action, never the refs. Keep a face it covers (mask, hood, helmet, costume head) or turns away hidden as staged, no ref face under it; refs then guide only what shows.\n\nSIGNAGE TEXT (authoritative — the scene itself calls for these inscriptions; render each on its surface as period-real physical lettering in the native script, exactly as written; add no other readable text anywhere):\n- 서장실 문 표지판: \"서장실\"\n\nScene-implied signs, documents, markings may appear in this place and era's language, not another's; invent none, print no shot text; no caption, subtitle, watermark, logo, overlay; stage decisive numerals out of legibility — a hand across, an oblique angle, shallow focus.\n\nCINEMATIC FINISH: this must read as a key frame from a film, not a posed photograph — motivated practical light with soft falloff, depth layering of foreground, midground and background, subtle air and atmosphere, restrained film grain. Within the CAMERA contract above, favor the more cinematic realization of every unspecified choice.\n\nCOMPOSITION VARIATION (this candidate only): same place, people, moment; reframe: non-obvious angle, height, distance, foreground layer. CAMERA & FRAME is a start, not a lock; refs, shot text fix what is seen, not camera position. Keep stated SHOT TEXT framing."
  },
  "readings": [
   {
    "label": "A",
    "direction": "중앙의 인물이 전경에 위치한 두 사람을 향해 미소를 지으며 시선을 맞추고 있음.",
    "built_space": "경찰서 복도 내부. 화면 우측 문 옆에 '서장실'이라고 적힌 표지판이 보임.",
    "entities": "중앙 인물은 제공된 전택수의 캐릭터 레퍼런스와 일치하지 않으며, 사용이 금지된 이전 스틸컷 인물의 얼굴과 비니, 가죽 재킷을 그대로 가져옴. 중앙 인물이 흑백 사진을 들고 있음.",
    "hard_violations": [
     "이전 스틸컷의 얼굴을 복사하여 캐릭터 레퍼런스 지시를 위반함",
     "서의용(좌측 하단)이 들고 있어야 할 사진을 전택수(중앙) 위치의 인물이 들고 있어 샷 연출(Staging) 및 소품 상태(Carried state)를 완전히 위반함"
    ],
    "physics": "중앙 인물이 양손으로 사진의 하단 모서리를 쥐고 지탱하고 있으며 땅에 서 있음."
   },
   {
    "label": "B",
    "direction": "중앙의 전택수가 몸을 돌려 전경의 두 사람(서의용, 나상혁)을 향해 따뜻한 미소를 띠며 바라보고 있음.",
    "built_space": "경찰서 복도 내부. 우측 벽에 '서장실' 텍스트가 적힌 표지판이 올바르게 부착되어 있음.",
    "entities": "중앙의 전택수는 캐릭터 레퍼런스에 제시된 얼굴, 머리 스타일, 남색 재킷을 정확히 반영함. 좌측 전경의 인물(서의용)이 흑백 사진을 들고 있음.",
    "hard_violations": [],
    "physics": "좌측 전경 인물이 한 손으로 흑백 사진을 안정적으로 들고 있으며, 중앙 인물은 두 발로 바닥을 디디고 서 있음."
   }
  ],
  "gq": {
   "route": "agree",
   "gap": 0.0,
   "per_model_winner": {
    "gemini-pro": "B",
    "openrouter:x-ai/grok-4.6": "B"
   },
   "models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ]
  },
  "totals": {
   "B": 7,
   "A": 3
  },
  "selected": "B",
  "ranking": [
   "B",
   "A"
  ],
  "verdicts": [
   {
    "label": "B",
    "score": 7,
    "verdict_ko": "전택수(중앙)가 뒤를 돌아보며 미소 짓는 구도와 캐릭터 레퍼런스(얼굴, 머리)를 잘 구현했으며, 서의용(좌측 전경)이 흑백 사진을 들고 있는 프레임 배치와 표지판 텍스트 요구사항을 정확히 따랐습니다."
   },
   {
    "label": "A",
    "score": 3,
    "verdict_ko": "좌측 전경의 인물이 들고 있어야 할 사진을 중앙 인물이 들고 있어 프레임 배치 지시를 어겼으며, 이전 스틸컷의 얼굴을 그대로 가져와 캐릭터 레퍼런스 규정을 심각하게 위반했습니다."
   }
  ],
  "refs": [
   {
    "label": "PREVIOUS SHOT STILL — a visually related earlier shot of this same place: the location's look, materials, fixed features and lighting mood are LOCKED to this photo; never copy its camera framing. The people visible in that photo are NOT in this shot: never carry their faces, bodies or clothing onto anyone here. Each person in THIS shot is defined solely by the PEOPLE list, their own CHARACTER REFERENCE images and their wardrobe notes. If this photo shows a character who cannot move (dead or unconscious), that character's exact pose, position and orientation are ALSO LOCKED — treat their body as a fixed prop of the set that only the camera moves around.",
    "path": "/Users/manta/Documents/Projects/TheRoad-I1/projects/5bddbdfc-2681-42a6-9837-43f35f60049d/images/f5372927-bbec-405d-ad2c-100d587f5373/scene/recipe/S94sh8_sel.png"
   },
   {
    "label": "CHARACTER REFERENCE — 전택수: the exact person appearing in this shot; match face, hair and build exactly.",
    "path": "<bytes:917702>"
   }
  ],
  "critique": {
   "issues": [
    {
     "issue_ko": "화면 중앙의 전택수가 캐릭터 레퍼런스의 의상('수사' 모자와 네이비 재킷)을 착용하지 않았으며, 대신 좌측 전경의 인물이 해당 의상을 입고 있어 캐릭터 간 의상이 뒤바뀌었습니다.",
     "fix_en": "Dress the center man in a navy jacket and cap, and the left man in a black beanie, grey hoodie, and leather jacket. Maintain all faces, poses, lighting, and the background.",
     "severity": "critical",
     "observation_index": 0
    },
    {
     "issue_ko": "좌측 전경 인물이 흑백 사진을 쥐고 있는 손가락 형태가 뭉개지고 사진 표면과 부자연스럽게 융합되어 있습니다.",
     "fix_en": "Redraw the left hand to have normal fingers gripping the photo edge without melting into the paper. Maintain the photo image, character poses, and background.",
     "severity": "critical",
     "observation_index": 1
    },
    {
     "issue_ko": "우측 벽면의 '서장실' 표지판 아래에 지시되지 않은 의미 불명의 뭉개진 글씨가 추가되었습니다.",
     "fix_en": "Erase the blurred text below '서장실' on the wall sign, leaving that space blank. Maintain the top text, characters, and corridor.",
     "severity": "major",
     "observation_index": 2
    },
    {
     "issue_ko": "화면 중앙의 전택수가 레퍼런스와 다른 얼굴이며 짧고 정돈된 검은 머리가 아니다",
     "fix_en": "Redraw the center man's face to precisely match the Jeon Taek-su reference. Maintain his expression, clothing, foreground characters, and the set.",
     "severity": "critical",
     "observation_index": 3
    },
    {
     "issue_ko": "전택수가 복도 몇 미터 앞 배경이 아니라 두 사람과 가까운 중경에 크게 서 있다",
     "fix_en": "Reduce the center man's scale to push him farther down the hall. Maintain all poses, the foreground figures, and lighting.",
     "severity": "major",
     "observation_index": 5,
     "needs_regeneration": true
    },
    {
     "issue_ko": "서의용 손의 흑백사진이 뒤에서 일부만 위를 향해 보여야 하는데 전면이 완전히 드러나 있다",
     "fix_en": "Angle the held photo upward and away so its face is mostly obscured. Maintain the holding hand, character positions, and background.",
     "severity": "major",
     "observation_index": 7
    },
    {
     "issue_ko": "왼쪽 인물 소매 패치 글자가 깨져 있다",
     "fix_en": "Remove the garbled letters from the left jacket's sleeve patch, making it a blank dark rectangle. Maintain the jacket, characters, and location.",
     "severity": "minor",
     "observation_index": 8
    }
   ],
   "observer_observations": [
    {
     "issue_ko": "화면 중앙의 전택수가 캐릭터 레퍼런스의 의상('수사' 모자와 네이비 재킷)을 착용하지 않았으며, 대신 좌측 전경의 인물이 해당 의상을 입고 있어 캐릭터 간 의상이 뒤바뀌었습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "좌측 전경 인물이 흑백 사진을 쥐고 있는 손가락 형태가 뭉개지고 사진 표면과 부자연스럽게 융합되어 있습니다.",
     "severity": "critical"
    },
    {
     "issue_ko": "우측 벽면의 '서장실' 표지판 아래에 지시되지 않은 의미 불명의 뭉개진 글씨가 추가되었습니다.",
     "severity": "major"
    },
    {
     "issue_ko": "화면 중앙의 전택수가 레퍼런스와 다른 얼굴이며 짧고 정돈된 검은 머리가 아니다",
     "severity": "critical"
    },
    {
     "issue_ko": "전택수가 레퍼런스 경찰 점퍼·수사 모자가 아닌 민무늬 남색 재킷을 입고, 수사 모자는 왼쪽 서의용이 쓰고 있다",
     "severity": "major"
    },
    {
     "issue_ko": "전택수가 복도 몇 미터 앞 배경이 아니라 두 사람과 가까운 중경에 크게 서 있다",
     "severity": "major"
    },
    {
     "issue_ko": "복도 축을 내려다보는 와이드가 아니라 세 사람이  bunched된 미디엄에 가깝다",
     "severity": "major"
    },
    {
     "issue_ko": "서의용 손의 흑백사진이 뒤에서 일부만 위를 향해 보여야 하는데 전면이 완전히 드러나 있다",
     "severity": "major"
    },
    {
     "issue_ko": "왼쪽 인물 소매 패치 글자가 깨져 있다",
     "severity": "minor"
    }
   ],
   "observer_models": [
    "gemini-pro",
    "openrouter:x-ai/grok-4.6"
   ],
   "observer_counts": {
    "gemini-pro": 3,
    "openrouter:x-ai/grok-4.6": 6
   }
  },
  "fix_severity_skipped_count": 4,
  "fix_severity_skipped": [
   {
    "issue_ko": "우측 벽면의 '서장실' 표지판 아래에 지시되지 않은 의미 불명의 뭉개진 글씨가 추가되었습니다.",
    "fix_en": "Erase the blurred text below '서장실' on the wall sign, leaving that space blank. Maintain the top text, characters, and corridor.",
    "severity": "major",
    "observation_index": 2
   },
   {
    "issue_ko": "전택수가 복도 몇 미터 앞 배경이 아니라 두 사람과 가까운 중경에 크게 서 있다",
    "fix_en": "Reduce the center man's scale to push him farther down the hall. Maintain all poses, the foreground figures, and lighting.",
    "severity": "major",
    "observation_index": 5,
    "needs_regeneration": true
   },
   {
    "issue_ko": "서의용 손의 흑백사진이 뒤에서 일부만 위를 향해 보여야 하는데 전면이 완전히 드러나 있다",
    "fix_en": "Angle the held photo upward and away so its face is mostly obscured. Maintain the holding hand, character positions, and background.",
    "severity": "major",
    "observation_index": 7
   },
   {
    "issue_ko": "왼쪽 인물 소매 패치 글자가 깨져 있다",
    "fix_en": "Remove the garbled letters from the left jacket's sleeve patch, making it a blank dark rectangle. Maintain the jacket, characters, and location.",
    "severity": "minor",
    "observation_index": 8
   }
  ],
  "repair_mode": "edit",
  "fix_ref_count": 3,
  "fix_prompt": "Edit this photograph. Apply ONLY the corrections listed below,\neach in place, changing nothing else:\n\nCORRECTIONS:\n- Dress the center man in a navy jacket and cap, and the left man in a black beanie, grey hoodie, and leather jacket. Maintain all faces, poses, lighting, and the background.\n- Redraw the left hand to have normal fingers gripping the photo edge without melting into the paper. Maintain the photo image, character poses, and background.\n- Redraw the center man's face to precisely match the Jeon Taek-su reference. Maintain his expression, clothing, foreground characters, and the set.\n\nPRESERVE EVERYTHING ELSE EXACTLY as in the original photograph —\nevery person, pose, face, garment, object, set feature, lighting,\ncamera and framing stay identical. Real-world writing that belongs\nto the scene — signs, banners, documents — stays exactly as the\noriginal shows it; do not add captions, subtitles, watermarks or\nany overlay text.",
  "fix_rejudge": {
   "forward_raw": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 컷에 등장한 서의용의 복장(비니, 가죽 재킷)과 레퍼런스의 전택수 복장(모자, 패치, 체크 셔츠)을 완벽하게 분리하여 유지했으며, 지시된 구도와 상황을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "서의용이 전택수의 모자를 쓰고 이전 컷의 복장을 잃어버렸으며, 전택수의 의상 디테일이 모두 누락되어 인물 설정과 연속성을 크게 훼손했습니다."
     }
    ],
    "readings": [
     {
      "label": "B",
      "direction": "중앙의 전택수가 전경의 두 사람을 향해 시선을 맞추며 미소 짓고 있으며, 두 사람 역시 전택수를 바라봄.",
      "built_space": "경찰서 복도 배경으로 우측 문에 '서장실' 표지판이 올바르게 부착되어 있음.",
      "entities": "전택수(중앙)는 레퍼런스와 일치하는 수사 모자, 남색 재킷, 체크 셔츠를 착용함. 서의용(좌측)은 이전 컷과 동일한 비니와 가죽 재킷을 입고 흑백 사진을 들고 있음. 표지판 텍스트 '서장실'이 정확함.",
      "hard_violations": [],
      "physics": "서의용이 두 손으로 사진을 안정적으로 쥐고 있으며, 서 있는 인물들의 자세와 무게 중심이 자연스러움."
     },
     {
      "label": "A",
      "direction": "전택수가 전경의 두 사람을 향해 미소 지으며 시선을 향하고, 전경의 두 사람도 그를 바라봄.",
      "built_space": "복도 배경과 우측 '서장실' 표지판이 존재함.",
      "entities": "전택수(중앙)는 레퍼런스의 모자와 셔츠가 누락됨. 서의용(좌측)이 써야 할 비니 대신 전택수의 '수사' 모자를 쓰고 있어 복장이 뒤바뀜. 흑백 사진은 서의용이 들고 있음.",
      "hard_violations": [],
      "physics": "서의용이 사진의 모서리를 잡고 있으며, 물리적으로 불가능한 부분은 없음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "B",
      "openrouter:x-ai/grok-4.6": "B"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "forward_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "이전 컷에 등장한 서의용의 복장(비니, 가죽 재킷)과 레퍼런스의 전택수 복장(모자, 패치, 체크 셔츠)을 완벽하게 분리하여 유지했으며, 지시된 구도와 상황을 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 3,
      "verdict_ko": "서의용이 전택수의 모자를 쓰고 이전 컷의 복장을 잃어버렸으며, 전택수의 의상 디테일이 모두 누락되어 인물 설정과 연속성을 크게 훼손했습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "중앙의 전택수가 전경의 두 사람을 향해 시선을 맞추며 미소 짓고 있으며, 두 사람 역시 전택수를 바라봄.",
      "built_space": "경찰서 복도 배경으로 우측 문에 '서장실' 표지판이 올바르게 부착되어 있음.",
      "entities": "전택수(중앙)는 레퍼런스와 일치하는 수사 모자, 남색 재킷, 체크 셔츠를 착용함. 서의용(좌측)은 이전 컷과 동일한 비니와 가죽 재킷을 입고 흑백 사진을 들고 있음. 표지판 텍스트 '서장실'이 정확함.",
      "hard_violations": [],
      "physics": "서의용이 두 손으로 사진을 안정적으로 쥐고 있으며, 서 있는 인물들의 자세와 무게 중심이 자연스러움."
     },
     {
      "label": "A",
      "direction": "전택수가 전경의 두 사람을 향해 미소 지으며 시선을 향하고, 전경의 두 사람도 그를 바라봄.",
      "built_space": "복도 배경과 우측 '서장실' 표지판이 존재함.",
      "entities": "전택수(중앙)는 레퍼런스의 모자와 셔츠가 누락됨. 서의용(좌측)이 써야 할 비니 대신 전택수의 '수사' 모자를 쓰고 있어 복장이 뒤바뀜. 흑백 사진은 서의용이 들고 있음.",
      "hard_violations": [],
      "physics": "서의용이 사진의 모서리를 잡고 있으며, 물리적으로 불가능한 부분은 없음."
     }
    ]
   },
   "reverse_raw": {
    "winner": "A",
    "ranking": [
     "A",
     "B"
    ],
    "verdicts": [
     {
      "label": "A",
      "score": 8,
      "verdict_ko": "캐릭터 레퍼런스의 의상(전택수의 모자와 재킷)과 이전 샷의 연속성(좌측 인물의 비니와 가죽 재킷, 사진)을 정확히 유지하며 프롬프트의 구도를 훌륭하게 구현했습니다."
     },
     {
      "label": "B",
      "score": 2,
      "verdict_ko": "전택수의 레퍼런스 의상(수사 모자, 체크 셔츠)을 전경 좌측 인물이 입고 있으며, 전택수 본인은 모자가 없는 등 캐릭터 식별 및 복장 연속성에 치명적인 오류가 있습니다."
     }
    ],
    "readings": [
     {
      "label": "A",
      "direction": "중앙의 택수가 전경의 두 사람을 향해 미소 지으며 시선을 맞추고 있고, 두 사람의 시선도 택수를 향함.",
      "built_space": "이전 샷과 일치하는 경찰서 복도 구조이며, 우측 벽에 '서장실' 표지판이 올바르게 위치함.",
      "entities": "택수는 레퍼런스에 주어진 '수사' 모자와 남색 재킷, 체크 셔츠를 입고 있음. 전경 좌측의 의용은 이전 샷과 동일하게 비니와 가죽 재킷을 입고 흑백 사진을 들고 있음.",
      "hard_violations": [],
      "physics": "세 인물 모두 지면에 안정적으로 서 있으며, 사진을 들고 있는 손의 파지법과 자세가 물리적으로 자연스러움."
     },
     {
      "label": "B",
      "direction": "중앙의 택수가 뒤돌아 미소 짓고 전경의 두 사람이 그를 바라보는 방향성은 일치함.",
      "built_space": "경찰서 복도 형태를 유지하고 우측 벽에 '서장실' 표지판이 보임.",
      "entities": "복장 할당에 심각한 오류가 있음. 전경 좌측 인물이 이전 샷의 비니 대신 택수의 레퍼런스 의상인 '수사' 모자와 체크 셔츠를 입고 있으며, 정작 택수는 모자를 벗고 다른 겉옷을 입고 있음.",
      "hard_violations": [
       "이전 샷의 복장 연속성 파괴 (좌측 인물의 비니와 가죽 재킷 누락)",
       "캐릭터 간 의상 뒤바뀜 (좌측 인물이 전택수의 레퍼런스 모자와 셔츠를 착용)"
      ],
      "physics": "사진을 든 손이나 인물들의 자세 자체는 지면을 잘 지지하고 있음."
     }
    ],
    "all_candidates_fail": false,
    "gq": {
     "route": "agree",
     "gap": 0.0,
     "per_model_winner": {
      "gemini-pro": "A",
      "openrouter:x-ai/grok-4.6": "A"
     },
     "models": [
      "gemini-pro",
      "openrouter:x-ai/grok-4.6"
     ]
    }
   },
   "reverse_normalized": {
    "winner": "B",
    "ranking": [
     "B",
     "A"
    ],
    "verdicts": [
     {
      "label": "B",
      "score": 8,
      "verdict_ko": "캐릭터 레퍼런스의 의상(전택수의 모자와 재킷)과 이전 샷의 연속성(좌측 인물의 비니와 가죽 재킷, 사진)을 정확히 유지하며 프롬프트의 구도를 훌륭하게 구현했습니다."
     },
     {
      "label": "A",
      "score": 2,
      "verdict_ko": "전택수의 레퍼런스 의상(수사 모자, 체크 셔츠)을 전경 좌측 인물이 입고 있으며, 전택수 본인은 모자가 없는 등 캐릭터 식별 및 복장 연속성에 치명적인 오류가 있습니다."
     }
    ],
    "all_candidates_fail": false,
    "readings": [
     {
      "label": "B",
      "direction": "중앙의 택수가 전경의 두 사람을 향해 미소 지으며 시선을 맞추고 있고, 두 사람의 시선도 택수를 향함.",
      "built_space": "이전 샷과 일치하는 경찰서 복도 구조이며, 우측 벽에 '서장실' 표지판이 올바르게 위치함.",
      "entities": "택수는 레퍼런스에 주어진 '수사' 모자와 남색 재킷, 체크 셔츠를 입고 있음. 전경 좌측의 의용은 이전 샷과 동일하게 비니와 가죽 재킷을 입고 흑백 사진을 들고 있음.",
      "hard_violations": [],
      "physics": "세 인물 모두 지면에 안정적으로 서 있으며, 사진을 들고 있는 손의 파지법과 자세가 물리적으로 자연스러움."
     },
     {
      "label": "A",
      "direction": "중앙의 택수가 뒤돌아 미소 짓고 전경의 두 사람이 그를 바라보는 방향성은 일치함.",
      "built_space": "경찰서 복도 형태를 유지하고 우측 벽에 '서장실' 표지판이 보임.",
      "entities": "복장 할당에 심각한 오류가 있음. 전경 좌측 인물이 이전 샷의 비니 대신 택수의 레퍼런스 의상인 '수사' 모자와 체크 셔츠를 입고 있으며, 정작 택수는 모자를 벗고 다른 겉옷을 입고 있음.",
      "hard_violations": [
       "이전 샷의 복장 연속성 파괴 (좌측 인물의 비니와 가죽 재킷 누락)",
       "캐릭터 간 의상 뒤바뀜 (좌측 인물이 전택수의 레퍼런스 모자와 셔츠를 착용)"
      ],
      "physics": "사진을 든 손이나 인물들의 자세 자체는 지면을 잘 지지하고 있음."
     }
    ]
   },
   "combined": {
    "totals": {
     "A": 5,
     "B": 16
    },
    "ranking": [
     "B",
     "A"
    ],
    "agreement": true,
    "policy": 1
   },
   "winner": "B",
   "fix_won": true,
   "all_candidates_fail": false,
   "policy": "fix_rejudge_v2_orig_on_tie"
  },
  "ref_mode": "prev+엔티티",
  "share_plan": {
   "ref_plan": "prev",
   "prev_anchor_tag": "S94sh8"
  }
 },
 "S94sh11::cine": {
  "applied": true,
  "fingerprint": "41084bd84cf4f900c579885420efe12046a7716b7d8c8a10ad3feb8e68913074",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S94sh11_sel.png",
  "source_sha256": "ab1a571f7f0184e7a1e8ebb9b8b45a7534d3199452fd0feac3ea140326267376",
  "file": "S94sh11_cine.png",
  "latency_ms": 13586
 },
 "era_ref::e170e9acef772e5b": {
  "subject": "2000년대 대한민국 교도소 조사실",
  "terms": [
   "교도소 조사실",
   "교정시설 조사실",
   "교도소 접견실 내부",
   "교도소 심문실"
  ],
  "queries": [
   [
    "2000년대 대한민국 교도소 조사실 교정시설 조사실 내부",
    "2000년대 대한민국 교도소 접견실 심문실 내부"
   ],
   [
    "대한민국 교도소 접견실 내부 사진",
    "대한민국 교정시설 조사실 내부 사진"
   ]
  ],
  "candidates": 4,
  "picked_index": 3,
  "picked_url": "https://img3.daumcdn.net/thumb/R658x0.q70/?fname=https%3A%2F%2Ft1.daumcdn.net%2Fnews%2F201910%2F31%2Fmoneytoday%2F20191031060122221xane.jpg",
  "picked_reason_ko": "후보 중 유일하게 2000년대 한국 교정시설의 소규모 조사실로 읽히며, 보안문·모니터·벽체와 바닥 마감이 명확하다.",
  "sha256": "9f57044a0f067bfddd4d4b7ee4217f7f7c74efec6f5dc3fb097a5f01aa23a659",
  "file": "eraref_e170e9acef772e5b.png"
 },
 "era_ref::c32c2848d77f5e21": {
  "subject": "2000년대 한국 지방 다방의 배달용 오토바이와 보온통",
  "terms": [
   "다방 배달 보온통",
   "시티백 다방 배달",
   "다방 커피 배달 가방"
  ],
  "queries": [
   [
    "다방 배달 보온통 시티백 다방 배달",
    "다방 커피 배달 가방"
   ],
   [
    "다방 배달 보온통",
    "시티백 다방 배달",
    "다방 커피 배달 가방"
   ]
  ],
  "candidates": 4,
  "picked_index": 2,
  "picked_url": "https://shop1.daumcdn.net/shophow/p/E5263829000.jpg?opt=R500x500.q90&type=thumb&ut=20260226172048",
  "picked_reason_ko": "2번은 당시 다방 배달에서 쓰던 형태에 가까운 검은 보온 가방과 컵 고정 구조를 제품 전체가 잘 보이도록 명확하게 담았다.",
  "sha256": "a277f963b7df89a199637f692890c7bbdf3f805332ede79bdb17646555ac34cf",
  "file": "eraref_c32c2848d77f5e21.png"
 },
 "S88sh2::cine": {
  "applied": true,
  "fingerprint": "f2ca5ec4bff3d5791039858d451bda28f7e34fc4fda9678d9cb4ac76863af765",
  "model": "x-ai/grok-imagine-image-2.0",
  "pack": "20.202608141300",
  "source_file": "S88sh2_sel.png",
  "source_sha256": "050f66ee19946fb67160e6b4ca24848a841325479be4d6ecacc87238ea5ddef4",
  "file": "S88sh2_cine.png",
  "latency_ms": 13953
 },
 "era_ref::2cc51f61abf22711": {
  "subject": "대한민국 교도소 수의 (2001년~2017년)",
  "terms": [
   "교도소 수의",
   "재소자 수의",
   "교도소 죄수복"
  ],
  "queries": [
   [
    "대한민국 교도소 수 2001년 2017년 통계",
    "대한민국 재소자 수 2001년 2017년 통계",
    "대한민국 교도소 죄수복 2001년 2017년 수의 변화"
   ],
   [
    "법무부 교정통계연보 2018 교정시설 수 연도별 2001",
    "교정통계연보 2018 수용인원 2001 2017",
    "대한민국 교정시설 현황 2001년 교도소 구치소 수",
    "대한민국 수용자 복장 규칙 2001년 2017년 색상"
   ],
   [
    "\"2001\" \"1일 평균 수용인원\" 교정시설",
    "\"2001년\" \"수용인원\" \"교정시설\"",
    "\"2001\" \"수형자\" \"미결수용자\" 교정",
    "\"교정기관\" \"2001\" \"2017\" 교도소"
   ],
   [
    "\"교정기관 수\" \"2001\" 법무부",
    "\"교정시설 수\" \"2001\" \"2017\" 법무부",
    "\"교도소\" \"구치소\" \"지소\" \"2017\" \"개\" 교정시설",
    "\"전국 교정기관\" \"2001\" \"개\""
   ],
   [
    "대한민국 교도소 수용자복 2001년 2006년",
    "대한민국 교도소 수용자복 2007년 2017년"
   ],
   [
    "2017년 수용자 의류 색상 수형자 미결수용자 법무부령",
    "2017 교도소 수용자복 기결수 미결수 색상",
    "형의 집행 및 수용자의 처우에 관한 법률 시행규칙 2017 수용자 의류 색상"
   ]
  ],
  "candidates": 4,
  "picked_index": 4,
  "picked_url": "https://yaejongshop.com/web/product/big/202302/da7209f63e3d2641076a4b9fa9a3921f.jpg",
  "picked_reason_ko": "후보 4가 실제 촬영된 수의의 전체 구성과 세부 형태를 가장 명료하게 보여 주는 참고 사진이다.",
  "sha256": "0fbc28c48b489e9721dffcedbc166992dfa88f1181effbc3579defa1d498848a",
  "file": "eraref_2cc51f61abf22711.png"
 }
}